Unveiling Groundwater Contamination Sources: A Comprehensive Guide to Principal Component Analysis (PCA)

Abstract

This article provides a thorough exploration of Principal Component Analysis (PCA) for identifying and apportioning groundwater contamination sources. Tailored for environmental researchers and scientists, it covers foundational principles, step-by-step methodologies, and advanced applications, including integration with receptor models like APCS-MLR. It also addresses common limitations, optimization strategies, and comparative analyses with other techniques, offering a complete resource for conducting robust groundwater geochemistry studies and informing effective remediation strategies.

Understanding PCA: The Geochemical Detective for Groundwater

Principal Component Analysis (PCA) is a powerful multivariate statistical technique used to simplify complex datasets by reducing their dimensionality. In essence, it transforms a large number of interrelated variables into a smaller set of uncorrelated variables called Principal Components (PCs), which capture the most significant patterns of variation within the data [1]. This transformation allows researchers to visualize high-dimensional data in two or three dimensions, identify underlying structures, and pinpoint the key factors driving observed patterns [2].

In the context of hydrochemical research, groundwater quality is characterized by numerous physical and chemical parameters (e.g., ions, nutrients, metals). Interpreting this multivariate data to identify pollution sources and natural geochemical processes is challenging. PCA serves as an indispensable tool for this task, distilling the essential information from extensive water quality datasets and providing insights into the factors influencing groundwater composition [3] [4].

The Core Mechanics of PCA

The Fundamental Principle: Variance as Information

PCA operates on a fundamental premise: variance equals information. Features or directions in the data with greater variance are assumed to contain more information. Thus, PCA seeks to find the new axes (principal components) that successively maximize the captured variance [1].

• First Principal Component (PC1): The axis through the data that captures the greatest possible variance.
• Second Principal Component (PC2): The axis orthogonal to PC1 that captures the next greatest amount of residual variance.
• Subsequent Components: This process continues, with each new component being orthogonal to all previous ones and capturing the next highest level of variance. The total number of components equals the number of original variables, but typically, only the first few are meaningful and retained for analysis [2].

Mathematically, these principal components are the eigenvectors of the data's covariance matrix, and the amount of variance each component captures is given by its corresponding eigenvalue [2]. A larger eigenvalue signifies a component that captures more of the total variance in the dataset.

A Visual Analogy

Imagine a dataset containing the heights and weights of 50 individuals, plotted on a two-dimensional scatter plot. The PCA algorithm would first find the line of best fit through this data—the direction where the spread of the points is greatest. This is the first principal component. The next line would be drawn perpendicular to the first, capturing the remaining largest spread. Previously, each individual was represented by two numbers (height and weight); after PCA, they can be represented by their position along these new principal axes, effectively reducing the data's dimensionality [2].

Application of PCA to Hydrochemical Data

PCA is particularly valuable in groundwater studies for differentiating between natural geochemical processes and anthropogenic pollution sources. The following workflow outlines its standard application.

Key Analytical Steps

1. Data Collection and Standardization Hydrochemical datasets are typically standardized before performing PCA. This involves transforming each variable to have a mean of zero and a standard deviation of one (Z-score normalization). This step is critical because it prevents variables with larger inherent scales (e.g., total dissolved solids) from dominating the analysis over those with smaller scales (e.g., pH) simply due to their numerical range [5].

2. Performing PCA and Interpreting Outputs After standardization, PCA is performed to generate key outputs:

• Explained Variance: The proportion of the total dataset variance captured by each principal component.
• Loadings: Indicate the correlation between the original variables and the principal components. High absolute loadings show which variables are most influential for a given PC.
• Scores: The coordinates of each individual water sample in the new principal component space [6].

Critical Visualizations for Interpretation

Scree Plots help determine how many principal components to retain. They display the eigenvalues (variance explained) by each component in descending order. The "elbow" of the plot—where the curve bends and flattens—indicates the optimal number of components to keep, as those before the elbow capture the majority of the meaningful variance [5] [6].

PCA Biplots are the most informative visualization, overlaying two types of information:

• Sample Scores (dots): Show the location of each groundwater sample in the PC space. Samples clustering together have similar hydrochemical compositions.
• Variable Loadings (vectors/arrows): Show the influence and direction of original water quality parameters on the PCs. The direction and length of an arrow indicate how strongly that variable influences the distribution of samples on the plot [6].

The interpretation of angles between variable vectors in a biplot is crucial:

• Small Angle: The two variables are positively correlated.
• 90° Angle: The variables are not correlated.
• Angle close to 180°: The variables are negatively correlated [6].

Deciphering Groundwater Chemistry: Common PCA Findings

When applied to groundwater datasets, PCA consistently identifies major categories of influencing factors, which can be quantified as shown in the table below.

Table 1: Common Pollution Sources Identified by PCA in Groundwater Studies

Source Category	Typical Hydrochemical Signature	Contribution Example	Study Location
Agricultural Activities	High loadings on NO₃⁻, NH₄⁺, NO₂⁻, K⁺, Ca²⁺, Mg²⁺ (from fertilizers) [3] [4]	38.5% (mixed agricultural and domestic) [7]	Qujiang River Basin, China [7]
Industrial Wastewater	High loadings on specific heavy metals (e.g., Fe, Mn), SO₄²⁻, Cl⁻, COD (Chemical Oxygen Demand) [3] [8]	35.2% [7]	Limin Groundwater Resource Area, China [3]
Domestic Sewage	High loadings on NH₄⁺, NO₂⁻, vital organisms (fecal bacteria), COD [4] [9]	Part of mixed domestic/agricultural source (38.5%) [7]	Foggia Province, Italy [9]
Natural Geochemical Processes	High loadings on Ca²⁺, Mg²⁺, HCO₃⁻, Na⁺ (from water-rock interaction) [3] [7]	26.3% [7]	Qujiang River Basin, China [7]
Seawater Intrusion	High loadings on Cl⁻, Na⁺, Electrical Conductivity (EC), Total Dissolved Solids (TDS) [10] [9]	Identified as a major source in coastal wells [9]	Kızılırmak Delta, Turkey [10]

The following diagram illustrates how these different sources and processes manifest on a hypothetical PCA biplot.

Advanced Protocols: Quantitative Source Apportionment

Beyond qualitative identification, PCA can be coupled with other statistical models to quantify the contribution of each pollution source. The most common method is the Absolute Principal Component Score-Multiple Linear Regression (APCS-MLR) model [3].

The APCS-MLR Protocol

This protocol provides a step-by-step methodology for implementing the APCS-MLR model, based on established research practices [3] [7].

Table 2: Essential Research Reagents and Computational Tools

Item Category	Specific Examples & Functions
Field Sampling Equipment	Peristaltic pump or bailer (for representative groundwater sampling); Multi-parameter probe (for in-situ measurement of pH, EC, T, DO, Eh); HDPE sample bottles (for trace metal and organic analysis).
Laboratory Analytical Reagents	ICP-MS/OES standards (for cation and trace metal quantification); IC eluents and standards (for anion quantification); Titrants and buffers (for HCO₃⁻ and COD analysis); Microbial growth media (for analysis of fecal indicator bacteria).
Statistical Computing Software	R (with FactoMineR, psych packages); Python (with scikit-learn, pandas libraries); Commercial software (e.g., SPSS, SAS) for PCA and regression analysis.
Data Quality Control	Certified Reference Materials (CRMs) for water quality; Field blanks and duplicate samples; Standardization solutions for all analytical instruments.

Step 1: Perform PCA and Calculate Absolute Principal Component Scores (APCS)

• Conduct standard PCA on the standardized hydrochemical dataset.
• Introduce an "artificial" sample with all values at zero (representing a theoretical zero-concentration background) and project it into the PC space.
• Calculate the APCS, which represent the coordinates of all real samples relative to this artificial zero point, providing an absolute measure of pollution for each component [3].

Step 2: Conduct Multiple Linear Regression (MLR)

• The concentration of each water quality parameter is regressed against all the calculated APCS using multiple linear regression.
• The regression equation for a parameter (e.g., nitrate) takes the form: ( C = b₀ + (b₁ * APCS₁) + (b₂ * APCS₂) + ... + (bₙ * APCSₙ) ) where ( C ) is the concentration, ( b₀ ) is the constant, and ( b₁, b₂, ... bₙ ) are the regression coefficients for each APCS [3].

Step 3: Apportion Contributions

• The factor contribution from the k-th source to the concentration ( C ) is calculated as ( bk * APCSk ).
• The percentage contribution from each source is then derived by normalizing these values. This quantifies the proportion of each pollutant, and the overall groundwater quality degradation, attributable to the identified sources (e.g., agriculture, industry) [3].

Case Study: APCS-MLR in Practice

A study of the Limin Groundwater Resource Area (China) from 2006 to 2016 used the PCA-APCS-MLR model to identify and quantify three major pollution source categories: water-rock interaction, agricultural fertilizer, and domestic/industrial wastewater. The model successfully calculated the average contribution of each source to specific pollutant categories like heavy metals and nutrients, demonstrating a clear temporal evolution of pollution sources linked to changes in land use [3].

Principal Component Analysis is a versatile and powerful tool for deciphering the complex narratives embedded in hydrochemical data. By reducing dimensionality, it illuminates the primary factors—be it natural water-rock interactions, agricultural runoff, or industrial discharge—governing groundwater composition. Its application, especially when coupled with quantitative receptor models like APCS-MLR, moves beyond mere identification to provide a scientifically defensible apportionment of pollution sources. This information is critical for developing targeted and effective groundwater management and remediation strategies, ensuring the protection of this vital natural resource.

In the realm of principal component analysis (PCA) for groundwater chemistry studies, understanding the relationship between eigenvectors, eigenvalues, and explained variance is fundamental. These mathematical concepts form the backbone of dimensionality reduction, allowing researchers to distill complex hydrochemical datasets into interpretable components that reveal underlying environmental processes [11].

Eigenvectors and eigenvalues are intrinsic properties of a square matrix that capture its fundamental behavior. In linear algebra, an eigenvector is a nonzero vector that changes only by a scalar factor when a linear transformation is applied to it. This scalar factor is the eigenvalue corresponding to that eigenvector [12]. Mathematically, for a matrix A, this relationship is expressed as:

Av = λv

Where v is the eigenvector and λ is the eigenvalue [12]. Geometrically, eigenvectors represent the axes along which a linear transformation acts by stretching or compressing, while eigenvalues indicate the magnitude of this stretching or compressing [11]. In the context of PCA, the covariance matrix of the data becomes the matrix A, and its eigenvectors define the principal components—the new directions in which the data varies the most [13] [14].

Variance, a statistical measure of data spread, becomes intrinsically linked to eigenvalues in PCA. The total variance in a standardized dataset is equal to the sum of all eigenvalues derived from the covariance matrix [15]. Each eigenvalue represents the amount of variance captured by its corresponding principal component, with larger eigenvalues indicating components that explain more variance [15]. This relationship allows researchers to quantify how much information each principal component retains from the original dataset.

The Mathematics of Variance Explanation in PCA

Theoretical Framework

In PCA, the connection between eigenvalues and variance is both mathematical and intuitive. When data is standardized (mean-centered and scaled to unit variance), the total variance equals the number of variables [15]. PCA transforms this variance into a new coordinate system defined by the eigenvectors of the covariance matrix.

The eigenvalue of each principal component equals the variance of the data when projected onto that component's axis [16]. Mathematically, if we have a covariance matrix S and an eigenvector μ, the variance along the direction of μ is given by:

μ^TSμ = λ

Where λ is the eigenvalue corresponding to μ [16]. This relationship means that the eigenvalues directly measure how "spread out" the data points are along each principal direction [13].

The proportion of total variance explained by each principal component is calculated as:

Variance Explained_i = λ_i / Σλ

Where λ_i is the eigenvalue of the i-th principal component and Σλ is the sum of all eigenvalues [15]. This quantitative measure allows researchers to make informed decisions about how many components to retain for analysis.

Geometric Interpretation

Geometrically, principal components align with the natural axes of the data cloud. The first principal component corresponds to the direction of maximum variance, the second principal component captures the next greatest variance direction while being orthogonal (uncorrelated) to the first, and so on [11] [14].

This relationship can be visualized by considering a scatter plot of data points. The line that minimizes the perpendicular distances from points to the line (reconstruction error) simultaneously maximizes the variance of the projections onto that line [13]. This duality principle means that the same eigenvector solves both optimization problems.

Table 1: Interpreting Eigenvalues and Variance in PCA

Mathematical Concept	Statistical Interpretation	Role in Groundwater Chemistry
Eigenvector	Direction of a principal component in original variable space	Represents a linear combination of hydrochemical parameters (e.g., mineral dissolution factor)
Eigenvalue	Amount of variance explained by the principal component	Quantifies importance of a pollution source or natural process
Sum of All Eigenvalues	Total variance in the standardized dataset	Represents total hydrochemical variability across sampling sites
Eigenvalue Ratio	Proportion of total variance explained by a component	Helps determine significance of identified pollution sources

Quantitative Analysis of Variance Explanation

Mathematical Proofs and Derivations

The fundamental relationship between eigenvalues and variance in PCA can be derived from the properties of the covariance matrix. For a dataset with variables centered to have zero mean, the sample covariance matrix S is defined as:

S = (1/(n-1)) X^TX

PCA involves finding the eigenvectors and eigenvalues of this covariance matrix. The eigenvectors v_i satisfy:

Sv_i = λ_iv_i

The variance of the projections of the data onto each eigenvector (the principal component scores) is given by:

Var(Xv_i) = v_i^TSv_i = v_i^T(λ_iv_i) = λ_i

This derivation confirms that the eigenvalue λ_i equals the variance of the i-th principal component [16].

Another important property is that the total variance in the data equals the trace of the covariance matrix (sum of its diagonal elements), which for standardized data equals the number of variables. This total variance is preserved in the principal component space:

Trace(S) = Σλ_i

This relationship ensures that the sum of all eigenvalues equals the total variance in the original dataset [15].

Practical Computation

In practice, eigenvalues and eigenvectors are computed using numerical algorithms such as the QR algorithm or singular value decomposition (SVD) [12]. For a groundwater chemistry dataset with p measured parameters, the covariance matrix is a p × p symmetric matrix. The eigenvectors of this matrix form an orthogonal basis, and the eigenvalues are always real numbers due to the symmetry of the covariance matrix.

Table 2: Worked Example of Variance Calculation from a PCA Model

Component	Eigenvalue	Individual Variance Explained	Cumulative Variance Explained
PC1	4.12	37.39%	37.39%
PC2	1.71	15.52%	52.91%
PC3	1.22	11.07%	63.98%
PC4	1.13	10.24%	74.22%
PC5	0.85	7.71%	81.93%
Remaining Components	1.99	18.07%	100.00%
Total	11.02	100.00%	100.00%

Data derived from a groundwater study analyzing 11 parameters across 215 sampling sites [17]

Experimental Protocols for PCA in Groundwater Research

Standardized PCA Workflow

The following protocol outlines the standardized procedure for implementing PCA in groundwater chemistry studies, with particular attention to the calculation and interpretation of eigenvectors and eigenvalues for variance explanation.

Phase 1: Data Preparation and Standardization

• Data Collection: Assemble hydrochemical data from monitoring wells, ensuring representative spatial coverage. A typical study might include 106-215 sampling sites with 12-22 parameters per site [18] [19] [17].
• Data Cleaning: Address missing values through appropriate imputation methods or exclusion, documenting all decisions.
• Data Standardization: Apply Z-score standardization to each variable:
  • Calculate mean (μ) and standard deviation (σ) for each parameter
  • Transform each value: Z = (X - μ)/σ [14]
  • This ensures all variables contribute equally to the covariance matrix

Phase 2: Covariance Matrix and Eigen Analysis

• Covariance Matrix Computation: Calculate the p × p covariance matrix from the standardized data, where p is the number of hydrochemical parameters [14].
• Eigen Decomposition: Perform eigen decomposition of the covariance matrix to extract:
  • Eigenvectors (principal component directions)
  • Eigenvalues (variances along each principal component) [14]
• Component Selection: Determine the number of components to retain based on:
  • Kaiser criterion (eigenvalues > 1)
  • Scree plot analysis
  • Cumulative variance explained (typically 70-80% for groundwater studies)

Phase 3: Interpretation and Validation

• Variance Explanation Calculation: For each component i:
  • Individual variance explained = (λ_i/Σλ) × 100%
  • Cumulative variance = Σ(λ₁ to λ_i)/Σλ × 100% [15]
• Component Rotation: Apply varimax rotation to improve interpretability while maintaining orthogonality [18].
• Source Identification: Interpret rotated components as potential pollution sources or natural processes based on factor loadings.
• Model Validation: Implement cross-validation procedures to assess stability of the PCA model, particularly checking for sensitivity to sample size and composition [19].

Visualizing the Eigenvalue-Variance Relationship

The following diagram illustrates the computational workflow for extracting and interpreting eigenvectors and eigenvalues in groundwater PCA studies:

Application in Groundwater Chemistry Source Identification

Case Study Implementation

In a comprehensive study of the Huaihe River Basin, researchers applied PCA to 215 groundwater samples analyzing 11 hydrochemical parameters [17]. The eigen analysis revealed four significant principal components with eigenvalues exceeding 1, collectively explaining 74.22% of the total variance in the dataset.

The relationship between eigenvalues and variance explanation was crucial for interpreting the results:

• PC1 (eigenvalue = 4.12) explained 37.39% of variance and was interpreted as "dissolved filtration, migration enrichment"
• PC2 (eigenvalue = 1.71) explained 15.52% of variance and represented "agricultural surface pollution"
• PC3 (eigenvalue = 1.22) explained 11.07% of variance, identified as "leaching and agricultural surface pollution"
• PC4 (eigenvalue = 1.13) explained 10.24% of variance, corresponding to "industrial pollution factor" [17]

This eigenvalue-based variance allocation provided a quantitative foundation for prioritizing pollution mitigation efforts, with the dissolution processes (PC1) identified as the most significant contributor to water quality variation.

Methodological Considerations and Limitations

While the eigenvalue-variance relationship provides a mathematically sound framework for dimensionality reduction, groundwater researchers must acknowledge several important limitations:

• Sample Size Sensitivity: PCA-based models can be unstable with small sample sizes. Studies with 106-215 samples have shown that WQI values can vary significantly when the model is applied to different subsets of the data [19].

• Parameter Selection Impact: The exclusion of a single water quality parameter from the PCA model can cause up to 60% deviation in WQI scores for some samples, highlighting the sensitivity of eigenvalues and eigenvectors to variable selection [19].

• Interpretation Challenges: While eigenvalues quantitatively measure variance explanation, connecting these statistical constructs to specific hydrogeochemical processes requires domain expertise and supporting evidence from other analytical techniques.

The following diagram illustrates how eigenvectors and eigenvalues function within the broader context of a groundwater PCA study, from data collection to source apportionment:

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Analytical Tools for PCA-Based Groundwater Studies

Tool/Reagent	Specification	Function in PCA Workflow
Multiparameter Water Quality Probe	pH, EC, TDS, DO sensors	Field measurement of fundamental parameters for initial dataset
Ion Chromatography System	Anions (F-, Cl-, NO3-, SO42-), Cations (Ca2+, Mg2+)	Quantification of major ions for hydrochemical characterization
ICP-MS Apparatus	Trace element detection (ppb level)	Analysis of heavy metals and trace elements as potential PCA variables
Statistical Software Package	R (FactoMineR), Python (scikit-learn), SPSS	Implementation of PCA algorithm and eigen decomposition
Data Standardization Module	Z-score normalization routine	Preprocessing to ensure equal variable contribution to covariance matrix
Varimax Rotation Algorithm	Orthogonal rotation method	Enhancement of component interpretability while maintaining mathematical properties
Cross-Validation Framework	Repeated random sub-sampling	Assessment of PCA model stability and eigenvalue consistency [19]

The precise relationship between eigenvectors, eigenvalues, and variance explanation forms the mathematical foundation of PCA's application in groundwater research. By quantifying how much each component contributes to total dataset variance through its eigenvalue, researchers can objectively identify the most significant pollution sources and natural processes affecting water quality. This mathematical rigor, when properly applied with domain expertise and validation procedures, enables evidence-based decision-making in environmental management and remediation planning.

Principal Component Analysis (PCA) is a powerful multivariate statistical method widely used in geochemistry and groundwater research to simplify complex datasets. It reduces the dimensionality of large sets of interrelated variables while retaining the trends and patterns present in the data [8]. For geochemists studying groundwater chemistry, PCA helps identify the dominant processes and sources influencing water quality, such as natural rock weathering, agricultural runoff, industrial discharge, and wastewater infiltration [20] [21] [7]. The method transforms original measured variables (e.g., ion concentrations, physicochemical parameters) into a new set of uncorrelated variables called principal components (PCs). Understanding the key terminology of loadings, scores, and principal components is essential for proper interpretation of PCA results in hydrogeochemical studies.

Key Terminology Explained

Principal Components (PCs)

Principal Components are new variables constructed as linear combinations of the original measured variables. They are orthogonal (uncorrelated) and are extracted in order of decreasing variance explained. The first principal component (PC1) captures the largest possible variance in the data, the second component (PC2) captures the next largest variance while being uncorrelated to the first, and so on [22]. In groundwater studies, these components often represent underlying geochemical processes or contamination sources. For example, a study in the Qujiang River Basin identified three principal components representing natural rock weathering, agricultural/domestic activities, and industrial wastewater discharge [7].

Loadings

Loadings are coefficients that indicate the contribution and direction of each original variable to a principal component. Mathematically, they represent the cosine of the angle of rotation between the original variable axis and the principal component axis [22]. Loadings range from -1 to +1, where:

• Higher absolute values (closer to ±1) indicate stronger influence of that variable on the component.
• Positive loadings suggest the variable contributes positively to the component.
• Negative loadings indicate an inverse relationship with the component.

In the Tunisia groundwater study, high positive loadings of radium and nitrates on the first principal component helped identify contamination from phosphate mining and agricultural activities [20].

Scores

PCA scores are the coordinates of the samples (e.g., groundwater samples) in the new coordinate system defined by the principal components. They represent the projection of each sample onto the principal components and indicate how similar or different samples are from each other with respect to the dominant patterns in the data [22] [23]. In practice, scoring allows researchers to:

• Classify groundwater samples into groups with similar characteristics.
• Identify spatial patterns of contamination.
• Recognize outliers or unusual samples.

In the Lion Creek watershed study, PCA scores helped distinguish water inflows with different chemical signatures and identify their likely sources, including mine water connections [24].

Table 1: Summary of Key PCA Terminology in Geochemical Context

Term	Mathematical Meaning	Interpretation in Geochemistry	Range/Properties
Principal Components	New orthogonal variables maximizing variance	Underlying geochemical processes or contamination sources	Uncorrelated, ordered by variance explained
Loadings	Correlation between original variables and PCs	Influence strength and direction of chemical parameters on processes	-1 to +1; higher absolute value = stronger influence
Scores	Projection of samples onto new PC axes	Position of each water sample along the geochemical processes	Continuous values; used for sample classification

Workflow and Relationships

The diagram below illustrates the logical relationship between original data, loadings, scores, and principal components in a groundwater chemistry study:

Application Protocols for Groundwater Studies

Standard Experimental Protocol for PCA in Groundwater Source Research

1. Study Design and Sampling

• Define study objectives and identify potential contamination sources.
• Establish sampling network considering hydrogeological setting and potential pollution point sources.
• Collect representative groundwater samples from monitoring wells, springs, or production wells.
• For temporal studies, establish appropriate sampling frequency to capture seasonal variations.

2. Field Sampling and Measurement

• Measure in-situ parameters (pH, temperature, electrical conductivity, dissolved oxygen, redox potential) using calibrated multiparameter meters.
• Collect samples in appropriate containers with necessary preservation (filtration, acidification, refrigeration).
• Follow standardized protocols such as EPA guidelines or HJ/T 164-2004 (Technical Specifications for Groundwater Environmental Monitoring) [7].
• Implement quality control measures including field blanks, duplicates, and trip spikes.

3. Laboratory Analysis

• Analyze major cations (Ca²⁺, Mg²⁺, Na⁺, K⁺) and anions (Cl⁻, SO₄²⁻, HCO₃⁻, NO₃⁻) using appropriate methods (IC, ICP-OES/MS, titration).
• Measure relevant trace elements, isotopes, or organic contaminants based on study objectives.
• Include quality assurance/quality control procedures with certified reference materials, method blanks, and duplicate analyses.

4. Data Preprocessing

• Compile data into a matrix with samples as rows and variables as columns.
• Address missing data through appropriate methods (deletion, imputation).
• Assess data distribution and apply transformations (log, centered log-ratio) if necessary.
• Standardize variables if they have different units or scales (common in geochemistry).

5. PCA Implementation and Interpretation

• Perform PCA using statistical software (R, SPSS, SAS, Python).
• Determine the number of significant components to retain (eigenvalue >1, scree plot, cumulative variance).
• Interpret loadings to identify which variables contribute most to each component.
• Analyze scores to identify sample groupings, spatial patterns, and outliers.
• Validate interpretations with hydrochemical knowledge and supplementary methods (Piper diagrams, ion ratios).

Table 2: Essential Research Reagents and Materials for Groundwater Geochemistry Studies

Category/Item	Specification/Function	Application Context
Field Equipment	Multiparameter water quality meter (pH, EC, DO, T)	On-site measurement of physical-chemical parameters
Sample Containers	HDPE bottles (various sizes), preservatives	Sample collection and storage for different analytes
Filtration Setup	0.45 μm membrane filters, syringes	Removal of suspended particles prior to cation analysis
Cation Analysis	Nitric acid (HNO₃) for preservation, ICP standards	Measurement of major and trace cations by ICP-OES/MS
Anion Analysis	IC eluents, standards, cartridge filters	Determination of major anions by ion chromatography
Data Analysis Tools	Statistical software (R, Python with scikit-learn)	PCA implementation and visualization

Advanced Applications and Integration

Integrated Methodological Frameworks

Contemporary groundwater studies increasingly combine PCA with complementary statistical methods to enhance source identification and apportionment. A study in the Qujiang River Basin demonstrated an integrated PCA-PMF-Mantel framework that enabled full-process assessment from qualitative identification to quantitative apportionment and spatial validation of pollution drivers [7]. This approach identified that anthropogenic sources accounted for 73.7% of total contribution, with mixed agricultural and domestic inputs dominating (38.5%), followed by industrial effluents (35.2%), while natural weathering contributed 26.3% [7].

Kernel PCA for Non-Linear Relationships

Traditional PCA assumes linear relationships between variables, which may not always hold in complex groundwater systems. Kernel PCA addresses this limitation by mapping data into a higher-dimensional feature space using non-linear kernel functions [25]. A recent study in Saudi Arabia's coastal aquifers employed Kernel PCA with polynomial kernels to develop a robust Water Quality Index that effectively handled non-linear hydrochemical relationships, particularly for salinity parameters affected by seawater intrusion [25].

Synoptic Sampling with PCA

In mineralized watersheds affected by mining, combining synoptic sampling with PCA can effectively discretize chemistry of inflows and source areas. This approach was successfully applied in the Lion Creek watershed, Colorado, where it identified primary contamination sources under low flow conditions and revealed hydraulic connections between bank inflows and mine water [24]. The method enabled development of a conceptual model of contaminant dynamics to inform remediation strategies.

Principal Component Analysis (PCA) is a powerful multivariate statistical technique extensively used in environmental research, such as identifying groundwater chemistry sources [26] [24] [8]. Its proper application hinges on verifying several key prerequisites concerning the dataset's structure and properties. Failure to meet these prerequisites can lead to misleading components that poorly represent the underlying data, ultimately compromising the scientific conclusions.

This protocol details the core assumptions of data normality, linearity, and sampling adequacy—assessed via the Kaiser-Meyer-Olkin (KMO) test—providing researchers with a structured framework for preparing and validating data for PCA within the context of groundwater geochemistry studies.

The Statistical Assumptions of PCA

PCA is based on linear algebra and operates on a correlation or covariance matrix. While the mathematical computation of PCA does not have strict distributional assumptions, the interpretation and reliability of the results are heavily influenced by the data's properties [27] [28].

The table below summarizes the core data prerequisites for a robust and interpretable PCA:

Table 1: Core Prerequisites for Principal Component Analysis

Prerequisite	Formal Requirement	Practical Implication in Groundwater Studies
Variable Type	Continuous (Interval/Ratio) [29] [30]	Constituent concentrations (e.g., As, Fe, pH) are ideal. Ordinal data can be used but may relax linearity.
Linearity	Linear relationships between variables [31] [29]	PCA models linear associations. Non-linearities can be addressed via data transformations.
Sampling Adequacy	Sufficient cases for stable correlations [29] [32]	A minimum of 150 observations or 5-10 cases per variable is often recommended.
Outliers	No significant outliers [31] [29]	Outliers disproportionately influence the correlation matrix and component orientation.
Data Reduction Suitability	Adequate correlations among variables [31] [29]	Tested via Bartlett's Test of Sphericity; variables must be sufficiently correlated to be reduced.

For groundwater studies, data should undergo specific checks before PCA. Normality is not a strict formal requirement for performing PCA [27] [28]. However, the Pearson correlation coefficient, which forms the basis of the PCA, is most informative and powerful when variables have a bivariate normal distribution [28] [30]. Furthermore, some methods for determining the number of significant components or for statistical inference may assume normality [28]. Skewed distributions, common for trace metal concentrations (e.g., Arsenic), can distort correlations. Applying transformations (e.g., log, square root) is often necessary to approximate normality and linearize relationships [28].

Experimental Protocol for Assumption Testing

This section provides a step-by-step workflow and detailed methodologies for testing the key prerequisites for PCA.

Workflow for PCA Prerequisite Testing

The following diagram outlines the sequential protocol for data preparation and assumption checking before proceeding with the main PCA.

Detailed Methodology

Step 1: Data Preparation and Screening

• Objective: Ensure data is structured and coded correctly for analysis.
• Protocol:
  • Structure data in a matrix format with rows representing individual groundwater samples (e.g., from 90 different wells) and columns representing the measured physicochemical parameters (e.g., As, Fe, NH₄-N, pH, SO₄) [26].
  • Address missing data through appropriate methods (e.g., deletion, imputation).
  • Confirm that all variables are continuous.

Step 2: Testing the Linearity Assumption

• Objective: Verify that relationships between pairs of variables are sufficiently linear.
• Protocol:
  • Visual Inspection: Generate a matrix of scatterplots for all variable pairs. Visually inspect the plots for clear non-linear patterns (e.g., U-shaped curves) [29].
  • Action: If non-linearity is detected, apply transformations (e.g., logarithmic, square root) to the variables and re-check the scatterplots.

Step 3: Checking for Normality

• Objective: Assess if variables are reasonably normally distributed to ensure correlations are representative.
• Protocol:
  • Graphical Methods: Create histograms with normal distribution curves or Quantile-Quantile (Q-Q) plots for each variable. In a Q-Q plot, data points that roughly follow the diagonal line indicate normality.
  • Statistical Tests: Optionally, use normality tests like Shapiro-Wilk or Kolmogorov-Smirnov. However, these can be overly sensitive with large sample sizes, so graphical inspection is often more practical.
  • Action for Skewed Data: For moderately to highly skewed parameters (common with trace metal concentrations), apply a log10 transformation. Re-check the distribution post-transformation [28].

Step 4: Testing Sampling Adequacy with KMO and Bartlett's Test

• Objective: Statistically confirm that the dataset is suitable for data reduction.
• Protocol:
  • Kaiser-Meyer-Olkin (KMO) Measure:
    • Compute the overall KMO statistic and the KMO for each variable (Measure of Sampling Adequacy, MSA) [29] [32].
    • Interpretation: An overall KMO value ≥ 0.6 is considered the minimum threshold, with values above 0.8 being good [30]. Variables with an individual MSA < 0.5 should be considered for removal from the analysis [32].
  • Bartlett's Test of Sphericity:
    • This test evaluates the null hypothesis that the correlation matrix is an identity matrix (variables are uncorrelated).
    • Interpretation: A statistically significant test result (p-value < .05) indicates that there are sufficient correlations in the data to proceed with PCA [29] [30].

Table 2: Essential Reagents and Statistical Resources for PCA

Item / Resource	Function / Description	Application Note
Statistical Software (R, SPSS)	Provides the computational environment to perform PCA and associated diagnostic tests.	R offers packages like FactorAssumptions for automated KMO and communality checks [32]. SPSS has built-in PCA procedures in the "Dimension Reduction" menu [29].
KMO & Bartlett's Test	Diagnostic tools to quantitatively assess data suitability for factor analysis/PCA.	KMO measures sampling adequacy; Bartlett's test checks if variables are sufficiently correlated [29] [32].
Tracer Compounds (e.g., LiCl, NaBr)	Used in synoptic sampling of watersheds to estimate streamflow and quantify constituent loading [24].	Enables the calculation of contaminant mass flux, providing a more accurate spatial pattern of contamination for the PCA dataset.
Data Transformation Library	A set of functions (e.g., log, sqrt, Box-Cox) to handle skewed data and improve linearity and normality.	Critical for pre-processing environmental concentration data, which often follows a log-normal distribution.

Applications in Groundwater Chemistry Research

In groundwater studies, PCA is primarily used for contaminant source attribution [8]. For example, research in the Hetao basin used PCA to demonstrate that high Arsenic groundwater was controlled by geological, reducing, and oxic factors, with Arsenic species highly correlated with Fe, NH₄-N, and pH [26]. Similarly, PCA has been applied to differentiate PFAS (per- and polyfluoroalkyl substances) signatures from different airports, helping to identify distinct anthropogenic sources [8].

The prerequisites outlined in this document are fundamental to the success of such applications. A valid PCA model relies on a well-structured dataset that has passed checks for linearity, sampling adequacy, and sufficient inter-correlations, ensuring the resulting components accurately reflect the true geochemical processes in the aquifer system.

From Theory to Practice: A Step-by-Step PCA Protocol for Groundwater Studies

Principal Component Analysis (PCA) has emerged as a powerful multivariate statistical technique for interpreting complex hydrochemical datasets in groundwater studies. Within the broader context of groundwater chemistry source research, PCA serves as a dimensionality reduction tool that identifies dominant patterns and sources of variation in water quality parameters, effectively distinguishing between natural geochemical processes and anthropogenic influences [33] [34]. The reliability of PCA outcomes is fundamentally dependent on the quality of initial data collection and the rigor of pre-processing methodologies applied before analysis. This protocol establishes comprehensive guidelines for the critical first phase of hydrochemical investigation: systematic data collection and pre-processing through standardization and centering techniques.

The application of PCA to groundwater chemistry enables researchers to interpret complex hydrochemical patterns from public supply well fields and other monitoring networks, providing valuable insights for natural background groundwater quality determination [35]. Proper pre-processing ensures that the resulting principal components accurately reflect true geochemical relationships rather than artifacts of measurement scale or data structure. Studies have demonstrated that appropriate data treatment significantly enhances the interpretability of PCA outputs for identifying pollution sources, including natural rock weathering, agricultural activities, and industrial contamination [33] [7].

Hydrochemical Data Collection Protocol

Field Sampling Methodology

Groundwater sample collection must follow standardized protocols to ensure data quality and comparability. The sampling procedure should begin with well purging for approximately 15 minutes or until in-situ parameters (pH, temperature, electrical conductivity, dissolved oxygen, and redox potential) stabilize, as measured by a multiparameter water quality analyzer [7]. Following stabilization, samples should be collected in pre-cleaned containers appropriate for the target analytes, preserved according to standard methods, and transported under controlled conditions to the analytical laboratory.

Comprehensive hydrochemical characterization should include major cations (Ca²⁺, Mg²⁺, Na⁺, K⁺), major anions (Cl⁻, SO₄²⁻, HCO₃⁻, NO₃⁻), and ancillary parameters including pH, electrical conductivity (EC), total dissolved solids (TDS), and temperature [33] [7]. For studies investigating specific contamination issues, additional parameters such as heavy metals, nutrients, or organic contaminants may be included based on hypothesized pollution sources. Documentation should include precise well locations, sampling depths, dates, times, and relevant field conditions.

Data Quality Assurance

Historical hydrochemical data from public supply well fields can provide valuable long-term perspectives but require careful validation regarding changes in analytical methods and reporting units over time [35]. As explicitly noted in hydrochemical research guidelines, "historical data must be checked for these inconsistencies and it is not uncommon that unit conversions have been applied twice, which is extremely difficult to identify" [35]. Methodological documentation should be maintained for all parameters, including analytical techniques, detection limits, and precision estimates.

Quality control measures should include field blanks, duplicate samples, and standard reference materials analyzed at predetermined frequencies. Any data points below detection limits should be handled consistently, either through substitution with a fraction of the detection limit or using statistical methods designed for censored data. For multivariate analysis, the completeness of data records across all parameters for each sampling location is crucial, as missing values can complicate subsequent statistical treatment [36].

Table 1: Essential Hydrochemical Parameters for Groundwater PCA Studies

Parameter Category	Specific Parameters	Measurement Units	Significance in PCA
Physical Parameters	pH, Temperature, EC, TDS	Standard units (pH, μS/cm, mg/L)	Indicators of general water chemistry and mineralization
Major Cations	Ca²⁺, Mg²⁺, Na⁺, K⁺	mg/L or meq/L	Water-rock interactions, salinity sources
Major Anions	Cl⁻, SO₄²⁻, HCO₃⁻, NO₃⁻	mg/L or meq/L	Anthropogenic influences, natural processes
Nutrients	NO₃⁻, NO₂⁻, NH₄⁺, PO₄³⁻	mg/L as N or P	Agricultural pollution indicators
Trace Elements	Fe, Mn, As, F⁻	μg/L or mg/L	Natural geochemistry and specific contamination sources

Data Pre-processing Methodology

Data Integration and Cleaning

Before statistical analysis, hydrochemical data must undergo systematic cleaning and integration. This process involves several critical steps to ensure data quality and compatibility. First, unit standardization must be applied across all parameters to establish consistency; for example, converting all concentrations to mg/L or meq/L as appropriate [35]. Temporal consistency should be verified, particularly when working with historical records where analytical methods may have changed over time.

Data should be structured in a matrix format with rows representing individual samples and columns representing measured parameters. This matrix serves as the fundamental input for subsequent multivariate analysis. A critical assessment for potential sampling bias must be conducted, considering factors such as well construction characteristics, screen intervals, and capture zones that might influence chemical compositions [35]. Documentation of all data transformations and the rationale for inclusion/exclusion of specific samples or parameters is essential for methodological transparency.

Handling Missing Values

Missing data presents a significant challenge in hydrochemical datasets and must be addressed prior to PCA implementation. The optimal approach depends on the extent and nature of missingness. For minimal missing values (<5% of dataset), mean imputation using the variable average may be acceptable, though this approach reduces variance in the dataset [36]. For more substantial missing data, sophisticated imputation techniques such as k-nearest neighbors (kNN) regression or multiple imputation by chained equations (MICE) provide more statistically robust solutions.

Specialized statistical packages offer specific functionality for handling missing values in multivariate analysis. As noted in computational guidelines, "the pca() and spca() from mixOmics can natively handle NA values in the input data through the implementation of the Non-linear Iterative Partial Least Squares (NIPALS) algorithm" [36]. The selected method for handling missing data should be clearly documented, as different approaches can influence subsequent PCA results.

Data Transformation

Hydrochemical parameters often exhibit right-skewed distributions and varying measurement scales that can disproportionately influence PCA results. Application of appropriate transformations helps normalize distributions and stabilize variances. The most common transformation approaches include:

• Logarithmic transformation: Particularly effective for parameters with positive skewness and large value ranges (e.g., ion concentrations). The log10 transformation is most commonly applied [36].
• Square root transformation: A moderate transformation suitable for count data or moderately skewed distributions.
• Box-Cox transformation: A power transformation that identifies the optimal transformation parameter to achieve normality.

Statistical assessment of normality (using Shapiro-Wilk test, Q-Q plots, or skewness/kurtosis measures) before and after transformation guides selection of the most appropriate method. As noted in omics data analysis protocols that are equally applicable to hydrochemistry, "we log10-transform our data frame to minimize the influence of outliers" before PCA implementation [36].

Standardization and Centering

Centering and standardization are critical pre-processing steps that directly impact PCA interpretation. These procedures address the fundamental issue of parameters with different measurement units and variances disproportionately influencing principal components.

• Centering: Subtraction of the variable mean from each value, transforming data to a zero-centered scale. This ensures that the first principal component describes the direction of maximum variance rather than being influenced by parameter means. Mathematically, for a variable x with mean μ, centered values are calculated as (x - μ).

• Standardization (Auto-scaling): Division of centered values by the variable standard deviation, converting all parameters to unit variance. This approach gives equal weight to all parameters regardless of their original measurement scale. Standardization is particularly important when parameters have substantially different variances or measurement units.

The decision to center versus standardize depends on research objectives and data characteristics. When parameters share comparable units and scales, centering alone may be sufficient. For heterogeneous parameters with different units (e.g., pH, mg/L, μS/cm), standardization is generally recommended [33] [36]. As explicitly stated in multivariate analysis guidelines, "we change the scale argument to TRUE to prevent dominating the PCA by high-abundance" parameters [36].

Table 2: Data Pre-processing Techniques and Their Applications

Pre-processing Technique	Mathematical Operation	Application Context	Effect on PCA
Centering	x́ = (x - μ)	Parameters with similar scales and units	PC1 describes variance direction rather than mean influence
Standardization (Auto-scaling)	x́ = (x - μ)/σ	Parameters with different units and variances	Equal weight to all parameters regardless of original scale
Log Transformation	x́ = log₁₀(x)	Right-skewed distributions (e.g., concentration data)	Normalizes distributions, reduces outlier influence
Range Scaling	x́ = (x - min)/(max - min)	Parameters with bounded ranges	Compresses all values to [0,1] interval
Pareto Scaling	x́ = (x - μ)/√σ	Compromise between auto and no scaling	Reduces but does not eliminate variance influence

Workflow Visualization

The following diagram illustrates the complete hydrochemical data collection and pre-processing workflow:

The Researcher's Toolkit: Essential Reagents and Materials

Table 3: Essential Research Reagents and Materials for Hydrochemical Analysis

Category	Item/Reagent	Specification/Purity	Application in Hydrochemical Studies
Field Equipment	Multiparameter water quality analyzer	pH, EC, TDS, DO, ORP sensors	In-situ measurement of physical-chemical parameters [7]
Sample Containers	HDPE bottles	Acid-washed, pre-cleaned	Cation and trace element sample collection
	Glass bottles	Pre-sterilized	Nutrient and organic parameter sampling
Preservation Reagents	Nitric acid (HNO₃)	Trace metal grade, ultrapure	Cation preservation at pH < 2
	Sodium hydroxide (NaOH)	Analytical grade	Alkalinity titration and pH adjustment
Laboratory Standards	Certified reference materials	NIST-traceable	Analytical quality control and method validation
	Anion standard solutions	Multi-element, certified concentrations	Ion chromatography calibration
Cation Analysis	ICP-MS/ICP-OES calibration standards	Custom mixed, certified	Major and trace cation quantification
Anion Analysis	Ion chromatography eluents	Carbonate/bicarbonate-based	Separation of major anions [7]

Implementation Considerations

Software and Computational Tools

Implementation of the pre-processing workflow requires appropriate statistical software. R programming environment offers several specialized packages for multivariate analysis. The FactoMineR package provides comprehensive PCA functionality through its PCA() function, while the mixOmics package offers enhanced capabilities for handling missing values through its pca() function [36]. As noted in computational protocols, "the mixOmics package also offers the Sparse PCA function spca(), which is an alternative to regular PCA and is suitable for large omics datasets" [36] - an approach equally beneficial for extensive hydrochemical datasets.

Python-based implementations through scikit-learn's decomposition.PCA module provide alternative computational frameworks. Regardless of platform selection, documentation of software versions, function parameters, and random seed settings ensures computational reproducibility.

Methodological Documentation

Comprehensive documentation of all pre-processing decisions is essential for research transparency and reproducibility. This should include specific descriptions of: (1) criteria for handling missing data, (2) transformation methods applied to each parameter with statistical justification, (3) standardization approach selected with rationale, and (4) any data exclusion criteria applied. Such documentation enables critical evaluation of methodological choices and facilitates comparative analysis across studies.

Properly pre-processed hydrochemical data establishes the foundation for meaningful PCA implementation, enabling researchers to accurately identify natural background conditions, discriminate between geogenic and anthropogenic influences, and apportion contamination sources in groundwater systems [35] [33] [7]. The rigorous application of these standardized protocols enhances the reliability and interpretability of multivariate statistical outcomes in groundwater chemistry research.

In the analysis of groundwater chemistry data, Principal Component Analysis (PCA) serves as a powerful dimensionality reduction technique that transforms complex, multidimensional hydrochemical datasets into a simpler structure by identifying dominant patterns of variance [37]. This step is crucial for distinguishing between natural geogenic processes and anthropogenic contamination sources in aquifer systems [7] [38]. The core mathematical foundation of PCA lies in constructing the covariance/correlation matrix and performing eigen-decomposition, which collectively identify the orthogonal directions of maximum variance in the original data [39] [37]. This protocol provides researchers with a standardized methodology for executing this critical phase of PCA within the context of groundwater hydrogeochemistry, enabling consistent identification of pollution sources and natural water-rock interactions across diverse study regions.

Theoretical Foundation

Covariance vs. Correlation Matrix

The covariance matrix represents a fundamental construct in PCA that captures how variables in the dataset co-vary with one another. For a data matrix X with dimensions n×p (where n is the number of groundwater samples and p is the number of hydrochemical parameters), the covariance matrix is a p×p symmetric matrix where diagonal elements represent the variances of individual variables, and off-diagonal elements represent the covariances between variable pairs [39] [37]. Mathematically, the sample covariance matrix is computed as Q = XᵀX/(n-1) for mean-centered data [37].

In hydrochemical applications, the correlation matrix is often preferred when variables exhibit different measurement units or scales (e.g., pH, mg/L for ions, μS/cm for conductivity) [39]. The correlation matrix is essentially a normalized covariance matrix where each element is scaled by the product of the standard deviations of the corresponding variables, resulting in values bounded between -1 and 1 [40]. This standardization prevents variables with inherently larger numerical ranges from dominating the PCA results merely due to their measurement scale [39].

Eigen-Decomposition Fundamentals

Eigen-decomposition, also known as spectral decomposition, is the mathematical procedure that identifies the principal components of the data [41]. For a square symmetric matrix like the covariance or correlation matrix C, eigen-decomposition solves the equation Cv = λv, where λ represents eigenvalues (scalars) and v represents eigenvectors (vectors) [41] [37]. The eigenvalues quantify the amount of variance captured by each principal component, while the corresponding eigenvectors define the direction of these components in the original variable space [39]. The eigenvectors are always orthogonal (perpendicular) to each other, forming an optimal basis for representing the variance structure in the data [37].

Table 1: Key Mathematical Components in Eigen-Decomposition

Component	Mathematical Symbol	Interpretation in Groundwater PCA
Covariance Matrix	C	Captures how hydrochemical parameters co-vary across samples
Correlation Matrix	R	Standardized version of C for unequal variable units
Eigenvalues	λ₁, λ₂, ..., λₚ	Amount of variance captured by each principal component
Eigenvectors	v₁, v₂, ..., vₚ	Directions of principal components in original variable space
Explained Variance	λᵢ/Σλ	Percentage of total variance explained by the i-th component

Experimental Protocol

Step-by-Step Procedure

Constructing the Covariance/Correlation Matrix

• Input Prepared Data: Begin with the standardized data matrix Xₛₜₚ of dimensions n×p from Step 1 (Data Preprocessing), where n is the number of groundwater samples and p is the number of hydrochemical parameters [39] [40].

• Matrix Selection Decision: Based on your research question and data characteristics, determine whether to use the covariance or correlation matrix:

  • For datasets with consistent measurement units where preserving original variance structure is prioritised, select the covariance matrix.
  • For hydrochemical datasets with parameters measured in different units (e.g., meq/L, mg/L, μS/cm, pH units), select the correlation matrix to prevent scale-based dominance [39].

• Compute Covariance Matrix: Calculate the sample covariance matrix using the formula: C = (Xₛₜₚᵀ × Xₛₜₚ)/(n-1) where Xₛₜₚ is the standardized data matrix and ᵀ denotes matrix transpose [40].

• Compute Correlation Matrix (Alternative): If using correlation instead of covariance, compute: R = (1/(n-1)) × Zᵀ × Z where Z is the z-score normalized matrix of the original data [39].

Performing Eigen-Decomposition

• Execute Decomposition: Perform eigen-decomposition of the covariance/correlation matrix to solve the characteristic equation: |C - λI| = 0 where I is the identity matrix of dimension p×p [37]. This yields p eigenvalues and corresponding eigenvectors.

• Sort Components: Sort eigenvalues in descending order (λ₁ ≥ λ₂ ≥ ... ≥ λₚ) and arrange eigenvectors accordingly [39] [40]. This ordering represents principal components from most to least significant in terms of variance explanation.

• Validate Results: Ensure that all eigenvalues are non-negative (a requirement for covariance/correlation matrices) and that eigenvectors are normalized to unit length [37].

• Compute Variance Explained: Calculate the proportion of total variance explained by each principal component as λᵢ/Σλ and cumulative variance as the running sum of these proportions [40].

Table 2: Workflow Output Specifications for Groundwater Applications

Process Stage	Expected Output	Quality Control Check
Matrix Construction	p×p symmetric matrix	Matrix should be positive semi-definite with no negative eigenvalues
Eigen-Decomposition	p eigenvalues and p eigenvectors	Sum of eigenvalues should equal total variance in the original data
Component Sorting	Descending-ordered eigenvalues	λ₁/Σλ ratio indicates compression efficiency
Variance Calculation	Explained variance proportions	Cumulative variance should approach 100% as components are added

Computational Implementation

The following Python code demonstrates the computational implementation of this protocol:

The Scientist's Toolkit

Research Reagent Solutions

Table 3: Essential Computational Tools for Matrix Construction and Eigen-Decomposition

Tool/Software	Specific Function	Application in Groundwater PCA
Python NumPy	np.cov(), np.corrcoef(), np.linalg.eig()	Core matrix operations and eigen-decomposition
R Statistical Environment	cov(), cor(), eigen()	Alternative open-source implementation
Scikit-learn PCA Module	sklearn.decomposition.PCA	High-level PCA implementation with validation
MATLAB	cov(), corr(), eig()	Commercial numerical computing platform
Covariance Matrix Algorithms	Bessel's correction (n-1 denominator)	Ensures unbiased sample covariance estimate
Eigen-Decomposition Algorithms	QR algorithm, Singular Value Decomposition	Numerical methods for stable decomposition

Applications in Groundwater Chemistry

Case Study Implementations

In the Qujiang River Basin study, researchers applied PCA to groundwater quality data, where the eigen-decomposition of the correlation matrix revealed three principal components accounting for a substantial portion of the total variance [7]. These components were interpretable as: (1) natural rock weathering processes, (2) agricultural and domestic activities, and (3) industrial wastewater discharges [7]. The eigenvalues provided quantitative measures of each source's contribution, with the first component typically capturing the largest variance proportion.

A similar approach in Nagpur, India demonstrated how eigen-decomposition of the correlation matrix identified two major components explaining approximately 61-62% of total variance across seasonal sampling campaigns [38]. The first component showed high loadings (eigenvector elements) for EC, TDS, TH, Cl⁻, NO₃⁻, Ca²⁺, and Mg²⁺, interpreted as pollution-controlled processes from anthropogenic sources [38]. The second component exhibited high loadings for Na⁺ and HCO₃⁻, representing alkalinity and pollution-controlled processes with mixed geogenic and anthropogenic influences [38].

Technical Validation

The numerical stability of eigen-decomposition can be verified through multiple approaches:

• Reconstruction Check: The original covariance matrix should be recoverable through C = VΛVᵀ, where V is the matrix of eigenvectors and Λ is the diagonal matrix of eigenvalues [37].
• Orthogonality Confirmation: Eigenvectors should be mutually orthogonal, confirmed by VᵀV = I (identity matrix) [41].
• Variance Preservation: The sum of eigenvalues should equal the total variance in the original dataset (sum of diagonal elements of covariance matrix) [39] [37].

For groundwater applications, the scree plot (eigenvalues ordered by magnitude) provides visual validation, typically showing a steep decline followed by an "elbow" point where subsequent components contribute minimally to explained variance [39] [37].

Workflow Visualization

Diagram 1: Matrix construction and eigen-decomposition workflow for groundwater PCA.

Principal Component Analysis (PCA) is a powerful multivariate statistical technique widely used in environmental chemistry to simplify complex datasets and identify the underlying sources of contamination. In groundwater studies, distinguishing between natural (geogenic) and human-made (anthropogenic) sources is crucial for effective resource management and remediation planning. PCA achieves this by transforming original, often correlated, water quality variables into a new set of uncorrelated variables called principal components (PCs). Each PC is a linear combination of the original variables, with the first component capturing the maximum possible variance in the data, and each subsequent component capturing the remaining variance in descending order [33]. The core outputs of PCA—loadings and scores—provide the key to interpreting these sources. Loadings indicate the contribution and direction of each original variable to a principal component, while scores position each water sample within the new component space, allowing for the grouping of samples with similar chemical characteristics [22].

The application of PCA to groundwater source identification is particularly valuable in areas impacted by multiple potential contamination pathways. For instance, studies have successfully utilized PCA to distinguish contaminants originating from phosphate mining (anthropogenic) from those arising from deep geothermal waters (natural) in the Gafsa basin of Southern Tunisia [20]. Similarly, an integrated approach combining PCA with other methods has been used to identify sources of water and metals in an acid mine drainage stream, revealing a hydraulic connection between mine water and contaminated seepages [42]. By analyzing the loadings, researchers can determine which chemical parameters are most strongly associated with each distinct source, providing a factual basis for targeted mitigation strategies.

Theoretical Framework: Loadings and Their Interpretation

What Are Loadings?

In PCA, loadings are the coefficients of the original variables in the linear equations that define the principal components. They represent the cosine of the angle between the original variable axis and the principal component axis, effectively describing how much each original variable contributes to the variance accounted for by each PC [22]. Geometrically, PCA is a process of rotating the original set of axes (the measured variables) to align with the directions of maximum variance in the data cloud. The loadings define the orientation of these new principal component axes relative to the original axes [22].

Loadings can range from -1 to +1. A loading with a large absolute value—whether positive or negative—indicates that the variable is highly influential on that component. The sign of the loading indicates the nature of the relationship. A positive loading suggests that the variable contributes positively to the component; a negative loading indicates that its absence (or low value) contributes to the component. When interpreting a component, one must examine all variables with large-magnitude loadings (both positive and negative) to understand the underlying pattern or source it represents [43] [44].

A Practical Example of Loadings Interpretation

Consider a simplified example from a general PCA context, which illustrates the interpretive process. Suppose a PCA of student test scores produces two principal components with the following loadings structure:

• PC1 Loadings: (0.5, 0.5, 0.5, 0.5) for Math, Physics, Reading, and Vocabulary tests.
• PC2 Loadings: (0.5, 0.5, -0.5, -0.5) for the same tests.

Interpreting this, PC1 has approximately equal, positive loadings for all four tests. This component can be interpreted as representing "overall academic ability" because it increases with high scores in all subjects. PC2, in contrast, has high positive loadings for Math and Physics and high negative loadings for Reading and Vocabulary. This component represents a "contrast between quantitative ability (Math/Physics) and verbal ability (Reading/Vocabulary)" [44]. This same logical framework is applied to water quality variables to identify contamination sources.

Step-by-Step Protocol for Interpreting Loadings

Protocol Workflow

The following diagram outlines the logical workflow for conducting a PCA analysis aimed at distinguishing natural and anthropogenic sources in groundwater.

Detailed Procedural Steps

Step 1: Data Preparation and Preprocessing Collect and prepare hydrochemical data from groundwater samples. Essential parameters often include major ions (Ca²⁺, Mg²⁺, Na⁺, K⁺, Cl⁻, SO₄²⁻, HCO₃⁻), nutrients (NO₃⁻, NO₂⁻, NH₄⁺), trace metals, and physical parameters like pH and Electrical Conductivity (EC) [33] [10]. Data should be checked for completeness and often standardized (e.g., z-score normalization) to avoid variables with larger numerical ranges artificially dominating the PCA [33].

Step 2: Perform PCA and Extract the Loadings Matrix Execute the PCA using standard statistical software. The key output for interpretation is the loadings matrix, which shows the loading of each original variable on each extracted principal component. The number of components to retain for interpretation should be based on objective criteria, such as the scree plot (retaining components before the plot levels off) or Kaiser’s rule (retaining components with eigenvalues >1) [43].

Step 3: Identify Significant Loadings To interpret a principal component, identify the variables that have a strong influence on it. This involves selecting loadings with large absolute magnitudes. The threshold for a "large" loading is subjective but should be determined a priori. A common rule of thumb is to consider loadings with an absolute value greater than 0.5 or 0.6 as significant for interpretation [43]. These high-loading variables are used to label and assign meaning to the component.

Step 4: Interpret the Principal Components and Assign Sources Analyze the pattern of high loadings for each component. A component with high positive loadings for Na⁺, Cl⁻, and EC might indicate a salinity source, such as seawater intrusion or dissolution of evaporite minerals. The specific context determines whether this is natural or anthropogenic. In the Kızılırmak Delta, Turkey, such a pattern was linked to natural geogenic processes and seawater intrusion [10]. Conversely, a component with a high loading for NO₃⁻, potentially accompanied by K⁺, could represent agricultural contamination from chemical fertilizers or manure [45]. A component with high loadings for radium, Cr(VI), or specific trace metals might be tied to an industrial or mining source, such as the phosphate mining in Southern Tunisia or the hexavalent chromium plume in Hinkley, California [20] [46].

Step 5: Correlate with Sample Scores and Spatial Distribution Plot the PCA scores to visualize how the individual groundwater samples are distributed along the new component axes. Samples with high positive or negative scores on a specific component are strongly influenced by the source that component represents. Mapping these scores using GIS can reveal spatial patterns, helping to confirm source locations. For example, a study might find that samples with high scores on the "agricultural" component cluster in regions of intense farming, while samples with high scores on the "geogenic" component are associated with a specific geological formation [33] [10].

Step 6: Validate the Interpretation Source identification should not rely on PCA alone. Validate the interpretations by:

• Cross-referencing with hydrogeology: Compare the PCA results with known geological layers and flow paths.
• Using supplementary tracers: Incorporate stable isotope data (e.g., δ¹⁵N-NO₃⁻ and δ¹⁸O-NO₃⁻ for nitrate sources) to corroborate findings [45].
• Comparing with known standards: Compare the chemical profiles of the identified sources to known source compositions (e.g., chemical composition of local fertilizers or mine tailings).

Case Study: Application in Southern Tunisia Groundwater

A study in the Gafsa basin of Southern Tunisia effectively demonstrates the application of PCA loadings for source apportionment [20]. The region faces contamination from phosphate mining and agricultural activities.

• Objective: To distinguish between radioactive and nitrate contamination sources in 33 groundwater samples.
• Method: PCA was applied to hydrochemical data, including radioactive elements (e.g., radium) and nitrate concentrations.
• Key Findings from Loadings Interpretation: The PCA results allowed the researchers to classify the groundwater samples into distinct groups based on their contamination profiles. The analysis revealed that samples most impacted by anthropogenic activities showed high levels of radium and nitrate. The loadings pattern helped delineate the contributions from different sources:
  • Phosphate Mining: A primary anthropogenic source of radioactivity.
  • Agricultural Runoff: The major source of nitrate contamination.
  • Fossil Geothermal Waters: A natural, geogenic source of salinity and radioactivity.
  • Low-Agricultural Areas: Relatively uncontaminated groundwater.
• Outcome: The study provided a clear conceptual model of contamination sources, underscoring the need for targeted strategies to address pollution from mining and agriculture separately [20].

The Scientist's Toolkit: Essential Reagents and Materials

Table 1: Essential Research Reagents and Materials for Groundwater PCA Studies.

Category	Item/Reagent	Function in Analysis
Field Sampling	High-Density Polyethylene (HDPE) Bottles	Inert container for sample collection, prevents leaching and contamination.
	Nitric Acid (HNO₃), Trace Metal Grade	Used for sample preservation, especially for metal analysis, to keep metals in solution.
Major Ion Analysis	Ion Chromatography (IC) System	Quantifies concentrations of major anions (Cl⁻, NO₃⁻, SO₄²⁻) and cations (Na⁺, K⁺, Ca²⁺, Mg²⁺).
	Inductively Coupled Plasma Mass Spectrometer (ICP-MS)	Provides highly sensitive measurement of trace metal and elemental concentrations.
Nutrient & Inorganic Carbon Analysis	Spectrophotometer / AutoAnalyzer	Measures concentrations of nutrients like nitrate (NO₃⁻), nitrite (NO₂⁻), and ammonium (NH₄⁺).
	Titrator	Measures alkalinity, reported as bicarbonate (HCO₃⁻) and carbonate (CO₃²⁻) concentration.
Field & Statistical Tools	Multiparameter Probe	In-situ measurement of pH, Electrical Conductivity (EC), Temperature, and Dissolved Oxygen.
	Statistical Software (R, Python, SPSS, etc.)	Platform for performing Principal Component Analysis and other multivariate statistics.

Table 2: Common groundwater quality parameters and their potential interpretation in PCA for source identification.

Parameter	Potential Link to Natural (Geogenic) Sources	Potential Link to Anthropogenic Sources	Example PCA Loading Context
Nitrate (NO₃⁻)	Typically very low background levels.	High loadings often link to agricultural fertilizer or manure/sewage [45].	High positive loading on an "Agricultural" PC.
Radium (Ra)	Can be released from aquifer minerals [20].	High loadings can indicate contamination from phosphate mining or other industrial waste [20].	High positive loading on a "Mining" PC.
Sodium (Na⁺) & Chloride (Cl⁻)	Saline intrusion, dissolution of halite deposits [10].	Road de-icing salts, industrial wastewater, sewage.	High positive loadings on a "Salinity" PC.
Chromium (Cr(VI))	Weathering of mafic and ultramafic rocks (typically low) [46].	Industrial plating, cooling water, historical releases (e.g., PG&E Hinkley) [46].	High positive loading on an "Industrial" PC.
Sulfate (SO₄²⁻)	Oxidation of sulfide minerals (e.g., pyrite), gypsum dissolution.	Acid mine drainage, industrial discharges [42].	High positive loading on a "Mining/Industrial" PC.
Bicarbonate (HCO₃⁻)	Carbonate mineral dissolution (calcite, dolomite), a primary natural buffer.	---	Often a dominant natural component with high positive loadings.
Calcium (Ca²⁺) & Magnesium (Mg²⁺)	Weathering of carbonate rocks (limestone, dolomite) and silicate minerals.	---	Typically indicate natural water-rock interaction.
Potassium (K⁺)	Weathering of K-feldspar.	Agricultural fertilizer (potash), manure.	Can appear with NO₃⁻ in an "Agricultural" PC.

Interpreting PCA loadings is a systematic process that moves from statistical output to environmental insight. By identifying variables with high loadings on each component and contextualizing them within the study area's known hydrology and land use, researchers can effectively fingerprint and distinguish between natural and anthropogenic contamination sources. This methodology, as demonstrated in diverse field studies, provides a robust factual foundation for developing targeted and effective groundwater protection and remediation strategies.

Principal Component Analysis (PCA) has emerged as a powerful statistical tool for environmental forensics, particularly in deciphering the complex origins of groundwater contaminants. This multivariate technique effectively reduces large, complex datasets into key components that explain the majority of variance in the data, revealing hidden patterns and relationships among parameters. In groundwater studies, PCA helps researchers distinguish between agricultural, industrial, and geogenic influences by identifying characteristic element associations and their spatial distributions. The integration of PCA with other receptor models and spatial analysis techniques has significantly advanced source apportionment capabilities, providing crucial insights for targeted pollution prevention and remediation strategies in diverse hydrogeological settings.

Theoretical Framework of PCA in Environmental Source Apportionment

Principal Component Analysis operates on the fundamental principle of dimensionality reduction, transforming correlated variables into a smaller set of uncorrelated principal components that capture the maximum variance in the data. The mathematical foundation of PCA involves eigenanalysis of the covariance or correlation matrix of the original variables, generating new hypothetical variables (principal components) that are linear combinations of the original measurements [47]. Each successive component accounts for as much of the remaining variance as possible, with the first few components typically explaining the majority of systematic variation in environmental datasets.

In groundwater chemistry, the application of PCA relies on the premise that different contamination sources produce distinct elemental or chemical signatures. Geogenic processes typically release elements through natural rock weathering and mineral dissolution, often characterized by associations between elements like fluoride, arsenic, and iron with specific geological formations [48] [49]. Agricultural activities introduce nitrates, phosphates, and potassium from fertilizers and manure, while industrial discharges often contribute heavy metals and complex organic compounds [45] [4]. PCA helps identify these signature patterns by grouping variables that co-vary across sampling locations, enabling researchers to trace contaminants back to their probable sources.

The strength of PCA lies in its ability to handle the complex, multi-dimensional nature of groundwater quality data where numerous parameters are measured simultaneously. By reducing data dimensionality while preserving essential information, PCA facilitates the identification of dominant contamination patterns and their spatial distributions, providing a scientific basis for prioritizing management interventions.

Comparative Case Studies

High Fluoride Groundwater in Sargodha, Pakistan

In the Sargodha region of Pakistan, groundwater contamination by fluoride exemplifies the complex interplay between geogenic and anthropogenic factors. A comprehensive study analyzing 48 groundwater samples revealed that 43.76% exceeded the WHO guideline value of 1.5 mg/L for fluoride, with concentrations ranging from 0.1 to 5.8 mg/L [48]. The application of PCA-MLR (Multiple Linear Regression) model identified five potential sources of groundwater pollution, with fluoride primarily originating from F-bearing minerals, ion exchange processes, rock-water interaction, and industrial and agricultural practices.

The hydrogeochemical facies in this region showed a transition from CaHCO₃ to NaHCO₃ water type, with alkaline pH, high Na⁺, HCO₃⁻ concentrations, and Ca-poor aquifers promoting fluoride dissolution. Positive correlations between Na⁺ and F⁻ suggested cation exchange processes where elevated Na⁺ occurs in Ca-poor aquifers, reducing Ca²⁺ availability and leading to higher F⁻ concentrations. The correlation between HCO₃⁻ and F⁻ indicated that carbonate mineral dissolution increases pH and HCO₃⁻, subsequently triggering F⁻ mobilization in aquifers [48]. Cluster analysis further categorized samples into three clusters: less polluted (10.4%), moderately polluted (39.5%), and severely polluted (50%), revealing the spatial variability of fluoride toxicity and aquifer vulnerability.

Health risk assessment demonstrated that children face higher risks from fluoride toxicity compared to adults, highlighting the public health implications of these findings. The study concluded that groundwater in the area is unsuitable for drinking, domestic, and agricultural needs without appropriate treatment [48].

Nitrate Contamination in Varying Aquifer Conditions

The impact of aquifer burial conditions on nitrate source apportionment was investigated using an integrated approach combining PCA-APCS-MLR (Absolute Principal Component Scores-Multiple Linear Regression) and MixSIAR (Mix Stable Isotope Analysis in R) models [45]. This research demonstrated that neglecting aquifer confinement could introduce absolute errors of 22-24% in source apportionment results, emphasizing the importance of considering hydrogeological settings in contamination studies.

For unconfined aquifers, the PCA-APCS-MLR analysis identified chemical fertilizers as the dominant source of NO₃⁻-N (52.5%), while the MixSIAR model further refined this assessment, identifying soil nitrogen (58%) as the primary contributor. In contrast, confined groundwater showed manure and sewage as the main nitrate source (53.9% via PCA-APCS-MLR and 37.9% via MixSIAR) [45]. These findings suggest that unconfined groundwater in regions with high soil nitrogen reserves faces persistent risk of NO₃⁻-N contamination, while confined aquifers are more vulnerable to sewage and manure inputs.

The study revealed that 75% of the groundwater samples exceeded the WHO drinking water standard for nitrate, underscoring the widespread nature of nitrate contamination and its threat to drinking water safety and ecosystem health. The differential source contributions between aquifer types highlight the necessity of tailored pollution control strategies based on specific hydrogeological conditions.

Radioactive and Nitrate Contamination in Southern Tunisia

In the Gafsa basin of southern Tunisia, PCA was successfully employed to distinguish between multiple contamination sources in a region affected by both phosphate mining and agricultural activities [20]. The study analyzed 33 groundwater samples and classified them into distinct groups based on contamination sources: phosphate mining, combined agricultural and mining activities, fossil geothermal waters, and low-agricultural areas.

Samples most affected by anthropogenic activities exhibited high levels of radium and nitrate, with contamination patterns correlating with specific environmental and chemical factors. The radioactivity in groundwater was primarily attributed to phosphate mining activities and deep groundwater sources from the North Western Sahara Aquifer System (NWSAS), while nitrate contamination was largely due to agricultural runoff, with secondary sources related to phosphate mining [20].

This case study underscores the complexity of groundwater contamination in regions with multiple overlapping pollution sources and demonstrates how PCA can effectively disentangle these complex influences. The findings provided critical insights for managing water quality in areas with similar environmental challenges, particularly where industrial and agricultural activities coexist.

Mixed Land-Use Area in Southwestern China

A comprehensive study in a multiple land-use area in southwestern China applied PCA combined with Geographic Information System (GIS) to explore spatial and temporal variations in groundwater quality and identify pollution sources [4]. The research analyzed groundwater samples from 26 wells in 2012 and 38 wells in 2018 for 13 water quality parameters, revealing evolving contamination patterns over time.

The PCA results identified four main factors governing groundwater quality: the hydro-geochemical process as the predominant factor, followed by agricultural activities, domestic sewage discharges, and industrial sewage discharges. The study found that agriculture expansion from 2012 to 2018 resulted in increased apportionment of agricultural pollution, while economic restructure and infrastructure improvement reduced the contributions of domestic sewage and industrial pollution [4].

Anthropogenic activities were identified as the major causes of elevated nitrogen concentrations (NO₃⁻, NO₂⁻, NH₄⁺) in groundwater, highlighting the necessity of controlling nitrogen sources through effective fertilizer management in agricultural areas and reducing sewage discharges in urban areas. The integration of GIS with PCA successfully identified pollutant sources and major factors driving groundwater quality variations, demonstrating the value of spatial analysis in contamination source tracking.

Comparative Analysis of Receptor Models in Chengdu Plain

A study in the Chengdu Plain of Southwestern China compared the performance of PMF (Positive Matrix Factorization) and PCA-APCS-MLR receptor models for groundwater pollution source identification and apportionment [50]. Both models identified five contamination sources with similar main load species for each potential source, including agricultural activities, domestic sewage, industrial wastewater, and geogenic processes.

The comparison revealed that PMF generally had higher R² values (0.603-0.931) compared to PCA-APCS-MLR (0.497-0.859) and smaller unexplained variability, suggesting that PMF provided a more physically plausible source apportionment in the study area [50]. However, both models showed reliable source estimation for species like NO₂⁻ and NO₃⁻, while contributions to species Fe, Mn, Cl⁻, SO₄²⁻ and NH₄⁺ were significantly different between models due to large data variability, differences in uncertainty analysis, and algorithm approaches.

This comparative study highlights the advantages of applying multiple receptor models to achieve reliable source identification and apportionment, particularly for understanding the applicability and limitations of different modeling approaches in groundwater pollution assessment.

Data Presentation

Table 1: Contamination Source Apportionment Across Case Studies

Case Study Location	Primary Contaminants	Agricultural Contribution	Industrial Contribution	Geogenic Contribution	Other Sources
Sargodha, Pakistan [48]	Fluoride	Significant (part of mixed sources)	Significant (part of mixed sources)	Dominant (F-bearing minerals)	Rock-water interaction
Unconfined Aquifers [45]	NO₃⁻-N	52.5% (PCA-APCS-MLR); 58% soil N (MixSIAR)	Not dominant	Not specified	Not applicable
Confined Aquifers [45]	NO₃⁻-N	Not dominant	Not specified	Not specified	Manure & sewage: 53.9% (PCA-APCS-MLR); 37.9% (MixSIAR)
Southern Tunisia [20]	Radium, Nitrate	Significant (nitrate)	Significant (mining-related radioactivity)	Significant (fossil geothermal waters)	Mixed sources
Southwestern China [4]	Nitrogen compounds	Increased from 2012 to 2018	Decreased from 2012 to 2018	Dominant factor	Domestic sewage

Table 2: Statistical Performance of Receptor Models in Source Apportionment

Model Type	R² Value Range	Unexplained Variability	Best Application Context	Limitations
PCA-APCS-MLR [50]	0.497-0.859	Moderate to high	Initial source identification, datasets with clear separation of sources	Higher unexplained variability for parameters with large variability
PMF [50]	0.603-0.931	Lower	Quantitative apportionment, complex mixed sources	Requires more sophisticated implementation
MixSIAR [45]	Not specified	Not specified	Isotope-assisted apportionment, agricultural vs. sewage differentiation	Requires isotopic data
PCA-GIS Integration [4]	Not specified	Not specified	Spatial pattern analysis, temporal trend assessment	Qualitative to semi-quantitative

Experimental Protocols

Standardized PCA Workflow for Groundwater Source Identification

The successful application of PCA for distinguishing contamination sources follows a systematic workflow encompassing study design, data collection, statistical analysis, and interpretation. The following protocol synthesizes best practices from the reviewed case studies:

Phase 1: Study Design and Sampling Strategy

• Define study boundaries based on hydrogeological units and land use patterns
• Establish sampling network using systematic grid or stratified random design
• Include reference sites in presumed uncontaminated areas for baseline comparison
• Consider aquifer type (confined vs. unconfined) as a stratification factor [45]
• Determine sample size based on statistical power requirements and practical constraints

Phase 2: Sample Collection and Analysis

• Collect groundwater samples following standardized protocols [51] [7]
• Purge wells for 15-20 minutes or until stabilization of pH, EC, and TDS before sampling
• Measure in-situ parameters (pH, Eh, DO, EC, T) using calibrated multiparameter instruments
• Collect samples in appropriate containers with necessary preservatives
• Analyze major cations (Ca²⁺, Mg²⁺, Na⁺, K⁺) and anions (Cl⁻, SO₄²⁻, HCO₃⁻, NO₃⁻, F⁻)
• Include trace elements relevant to suspected sources (As, Pb, Cd, Cr, Cu, Zn, etc.)
• Ensure quality control through blanks, duplicates, and standard reference materials

Phase 3: Data Preprocessing and Statistical Analysis

• Compile data into a matrix with samples as rows and parameters as columns
• Replace values below detection limits with DL/√2
• Test data for normality and apply appropriate transformations if necessary
• Standardize data using z-scores to avoid scale dependence among variables
• Perform PCA on the correlation matrix with Varimax rotation for better interpretation
• Retain components with eigenvalues >1 (Kaiser criterion) or based on scree plot
• Interpret factor loadings considering |r| > 0.5 as moderate and |r| > 0.75 as strong

Phase 4: Source Apportionment and Validation

• Apply APCS-MLR or PMF for quantitative source contribution estimates [50]
• Validate results using spatial analysis with GIS [4]
• Compare with known land use patterns and potential pollution sources
• Conduct uncertainty analysis through bootstrap methods or model comparison
• Integrate with hydrochemical diagrams and ionic ratios for geochemical consistency

Advanced Integrated Methodologies

For complex scenarios with multiple overlapping sources, integrated approaches yield more robust results:

Coupled PCA-PMF-Mantel Test Framework [7]

• Use PCA for initial qualitative source identification
• Apply PMF for quantitative source apportionment
• Validate spatial correlations using Mantel test between hydrochemistry and environmental factors (land use, geology, etc.)
• This integrated approach provides a full-process assessment from identification to spatial validation

Multi-Model Comparison Approach [50]

• Apply both PCA-APCS-MLR and PMF models to the same dataset
• Compare results for consistency and identify discrepancies
• Assess model performance based on R² values and unexplained variability
• Use consensus estimates where models agree and investigate differences where they diverge

Visualization

PCA Workflow for Groundwater Source Identification

Contaminant Source Signatures in PCA

The Scientist's Toolkit

Essential Research Reagents and Materials

Table 3: Essential Analytical Tools for PCA-Based Groundwater Studies

Category	Specific Items	Function in Analysis	Application Examples
Field Equipment	Multiparameter water quality analyzer (pH, EC, Eh, DO, T)	In-situ parameter measurement for quality assessment	HI9828 (HANNA) used in Qujiang River Basin [7]
	Portable XRF analyzer	Rapid elemental analysis in field conditions	Hitachi XMET 8000 for soil metal analysis [52]
	GPS device	Precise location mapping for spatial analysis	ArcGIS Field Maps with <1m accuracy [52]
Laboratory Analysis	Ion Chromatography (IC)	Anion analysis (NO₃⁻, SO₄²⁻, Cl⁻, F⁻)	Shimadzu LC-10ADvp [51], Dionex DX-120 [48]
	ICP-MS	Trace metal analysis with high sensitivity	Agilent 7500ce for heavy metals [51]
	Titration equipment	HCO₃⁻ determination	Standard acid-base titration [51]
Statistical Software	R Statistical Software	PCA, PMF, spatial analysis	Implementation of PCA-APCS-MLR and MixSIAR [45]
	SPSS, SAS	Multivariate statistical analysis	Factor analysis, cluster analysis [47]
	GIS Software (ArcGIS, QGIS)	Spatial interpolation and mapping	Kriging, spatial pattern analysis [4]
Reference Materials	Certified reference materials	Quality assurance and method validation	NIST 2711 for soil analysis [52]
	Standard solutions	Instrument calibration	Multi-element standards for ICP-MS [51]

The case studies presented demonstrate the robust application of Principal Component Analysis for distinguishing agricultural, industrial, and geogenic influences on groundwater quality across diverse hydrogeological settings. When properly implemented with appropriate sampling design, analytical protocols, and statistical validation, PCA emerges as an powerful tool for environmental forensics that can disentangle complex contamination patterns and provide scientifically-defensible basis for pollution management strategies. The integration of PCA with complementary methods like PMF, GIS, and isotopic analysis further enhances source apportionment capabilities, offering a comprehensive framework for addressing groundwater quality challenges in an increasingly human-impacted world. As groundwater resources face growing pressures from agricultural intensification, industrial expansion, and geogenic contamination, the continued refinement and application of multivariate statistical approaches will remain essential for developing targeted, effective protection and remediation strategies.

The Absolute Principal Component Score-Multiple Linear Regression (APCS-MLR) model is a powerful receptor modeling technique used for quantitative source apportionment in environmental studies. When combined with Principal Component Analysis (PCA), it provides a robust framework for identifying and quantifying the contributions of various pollution sources to groundwater chemistry [53]. This method was originally developed by Thurston and Spengler for air pollution studies but has since been successfully adapted for aquatic systems, including groundwater contamination assessment [54] [55].

The key advantage of the PCA-APCS-MLR approach lies in its ability to work with conventional hydrochemical data, making it more accessible and cost-effective compared to isotope-based methods that require sophisticated instrumentation and complex analyses [54] [56]. This model has demonstrated remarkable consistency with advanced isotope mixing models like SIAR (Stable Isotope Analysis in R), with comparative studies showing less than 4% difference in source contribution estimates for key pollutants like nitrate [54].

Theoretical Framework and Core Principles

Methodological Foundation

The PCA-APCS-MLR model operates on the fundamental principle that the chemical composition of groundwater represents a mixture of contributions from various sources. The model quantifies these contributions through a structured statistical approach that combines dimensionality reduction with regression analysis [53] [55]. The methodology is particularly valuable in scenarios where traditional forward modeling approaches become challenging due to complex hydrologic conditions or limited source emission data [57].

Comparative Advantages

Recent methodological comparisons have highlighted several distinct advantages of the PCA-APCS-MLR approach:

• Accessibility: Requires only conventional hydrochemical indicators that can be easily measured, avoiding the need for sophisticated isotope analysis [54] [56]
• Efficiency: Provides a faster alternative to isotope methods while maintaining comparable accuracy for many applications [54]
• Comprehensive Apportionment: Capable of simultaneously quantifying multiple pollution sources across various chemical parameters [53]
• Data Requirements: Effectively handles datasets with moderate variability, though performance may vary compared to PMF (Positive Matrix Factorization) for parameters with high variability [55]

Experimental Protocol and Workflow

Sample Collection and Data Preparation

Table 1: Essential Hydrochemical Parameters for PCA-APCS-MLR Analysis

Parameter Category	Specific Indicators	Measurement Methods	Significance in Source Apportionment
Basic Physicochemical	pH, TDS, DO, EC	In situ testing with calibrated meters (e.g., SX-620 pH Tester, Hanna DiST)	Determines hydrochemical environment and redox conditions [53]
Major Anions	Cl⁻, SO₄²⁻, NO₃⁻-N, NO₂⁻-N	Ion chromatography, spectrophotometry	Indicators of agricultural, industrial, and sewage inputs [53] [58]
Major Cations	K⁺, Na⁺, Ca²⁺, Mg²⁺	ICP-OES, atomic absorption spectroscopy	Traces natural water-rock interactions and industrial pollution [53] [17]
Nutrient Parameters	NH₄⁺-N, TP, COD	Spectrophotometric methods	Identifies agricultural runoff and organic pollution [53] [57]
Trace Elements	Mn, Fe, I, Sb	ICP-MS, specialized probes	Fingerprints specific industrial activities and natural geology [53] [58]

Sample Collection Protocol:

• Site Selection: Deploy monitoring wells considering pollution source distribution and groundwater flow direction using "point-face combination" distribution method [17]
• Spatial Distribution: Aim for approximately 1 km² per sampling point in regional studies [53]
• In Situ Measurements: Immediately measure temperature, pH, DO, and TDS using calibrated portable instruments [53]
• Sample Preservation: Collect samples in 1.5L polyethylene containers, preserve according to parameter requirements, and transport to laboratory under controlled conditions [53]
• Quality Assurance: Implement field blanks, duplicates, and standard reference materials to ensure data quality

Data Preprocessing and Validation

Figure 1: Computational Workflow for PCA-APCS-MLR Modeling

Statistical Validation Steps:

• Dataset Adequacy Testing:
  • Perform Kaiser-Meyer-Olkin (KMO) test (requires value >0.5 for PCA reliability)
  • Conduct Bartlett's test of sphericity (requires significance level <0.05) [53]
• Data Standardization: Normalize dataset using Z-score standardization to eliminate unit influence
• PCA Implementation:
  • Extract principal components with eigenvalues >1.0 based on Kaiser criterion [53]
  • Apply varimax rotation to redistribute and polarize loadings, creating varifactors (VFs)
• Component Loading Interpretation:
  • Strong loadings: |loading| > 0.75
  • Moderate loadings: 0.50 ≤ |loading| ≤ 0.75
  • Weak loadings: 0.30 ≤ |loading| ≤ 0.50 [53] [57]

APCS-MLR Calculation Procedure

The mathematical framework proceeds through these computational stages:

• Initial PCA Factor Scores Calculation: (Az)ij = ai1C1j + ai2C2j + … + aimCmj where Az represents component score, a stands for component loading, C is measured concentration, i = 1,2,…,p (components), j = 1,2,…,n (samples), and m is number of parameters [53]

• Absolute Principal Component Scores (APCS) Conversion: APCSjk = (Az)jk - (A0)jk where (Az)jk and (A0)jk are actual and zero score values of principal component k at sampling site j [57]

• Multiple Linear Regression Model: C_i = b_0 + Σ(APCS_k × b_ik) + ε where Ci is concentration of chemical parameter i, b0 is constant term, b_ik is regression coefficient for source k on parameter i, and ε is residual error [53] [55]

Application Case Studies in Groundwater Research

Comparative Performance Assessment

Table 2: Case Study Applications of PCA-APCS-MLR in Groundwater Source Apportionment

Study Area & Context	Identified Pollution Sources (Contribution %)	Key Parameters	Model Performance Metrics	Comparative Method Validation
Dagu River GW Source Area, China [54]	Chemical fertilizers (58.1%), Natural sources (22.7%), Manure & sewage (19.2%)	NO₃⁻-N, SO₄²⁻, Cl⁻, TDS	Close agreement with SIAR model (difference <3.8% for fertilizers)	SIAR model consistency: R² >0.85 for major sources
Yangtze River Delta, China [53]	Natural hydro-chemical evolution (18.9%), Textile industry (18.2%), Agricultural activities (17.1%), Other industry (15.1%), Domestic sewage (4.2%)	Multiple ions including NH₄⁺-N, NO₃⁻-N, SO₄²⁻, Mn, Sb	Comprehensive source profiling across 17 parameters	Method applicability confirmed for complex anthropogenic areas
Mixed Land-use Area, SW China [55]	Agricultural (24-27%), Geological (18-24%), Industrial (15-25%), Unexplained variability (balance)	NO₂⁻, NO₃⁻, Fe, Mn, Cl⁻, SO₄²⁻, NH₄⁺	R² = 0.497-0.859 for parameter predictions	Compared with PMF model; PMF showed higher R² (0.603-0.931)
Zhuji, East China [59]	Shallow GW: Hydrogeological conditions, Agricultural activities, Domestic sewage/manure (total 77.6%)	δ¹⁵N–NO₃, δ¹⁸O–NO₃, TN, NO₃⁻, NH₄⁺	Differentiated shallow vs. deep groundwater sources	Combined with SIAR using isotopic fractionation factors
Poyang Lake Basin, China [57]	Urban wastewater (34%), Agricultural non-point sources (16%), Other natural and anthropogenic sources	TP, NH₃–N, COD, organic contaminants	Improved accuracy with land-use parameters	GIS correlation strengthened source identification

Advanced Implementation Considerations

Aquifer Burial Condition Effects: Recent research highlights the critical importance of considering aquifer confinement conditions in source apportionment. Studies demonstrate that neglecting burial conditions can introduce absolute errors of 22-24% in source contribution estimates [45]. The dominant NO₃⁻ sources differ significantly between unconfined aquifers (primarily chemical fertilizers: 52.5%) and confined aquifers (dominated by manure & sewage: 53.9%) [45].

Integration with Complementary Methods:

• Land-use Parameter Integration: Incorporating GIS-based land-use parameters (cultivated land, urban areas, industrial zones) as auxiliary variables improves source identification accuracy and reduces subjectivity [57]
• Isotope Coupling Approach: Combining PCA-APCS-MLR with stable isotope analysis (δ¹⁵N–NO₃, δ¹⁸O–NO₃) provides enhanced validation and refines source apportionment, particularly for nitrogen pollution [59]
• Multi-Model Verification: Applying both PMF and PCA-APCS-MLR models achieves more reliable source identification and apportionment, with PMF potentially offering better performance for parameters with high variability [55]

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Research Reagents and Analytical Solutions for PCA-APCS-MLR Studies

Category	Specific Items	Technical Specifications	Application Purpose
Field Measurement Instruments	SX-620 pH Tester, SX-630 ORP Tester, Hanna DiST for TDS	Calibrated portable meters with appropriate electrodes	In situ measurement of fundamental parameters (pH, DO, TDS, EC) [53]
Sample Containers & Preservation	1.5L polyethylene containers, chemical preservatives (H₂SO₄ for nutrients, NaOH for metals)	Acid-washed, pre-cleaned containers; analytical grade preservatives	Maintain sample integrity during transport and storage [53]
Laboratory Analysis Systems	Ion Chromatography system, ICP-OES/MS, Spectrophotometer	Appropriate detection limits for expected concentration ranges	Quantification of major ions, trace elements, and nutrient parameters [53] [17]
Statistical Software Packages	SPSS, RStudio with appropriate packages	Latest versions with multivariate statistical capabilities	PCA execution, APCS calculation, and MLR analysis [53] [57]
GIS & Spatial Analysis Tools	ArcGIS, QGIS with spatial analysis extensions	Capability for land-use classification and spatial correlation	Integration of land-use parameters for enhanced source identification [57]
Isotope Analysis Materials	Stable isotope ratio mass spectrometer, specialized sample preparation equipment	High precision for δ¹⁵N and δ¹⁸O measurements (±0.2‰)	Optional validation and refinement of nitrate source apportionment [54] [59]

Methodological Limitations and Best Practices

Limitations and Considerations

While PCA-APCS-MLR presents a powerful approach for source apportionment, researchers should consider these limitations:

• Source Profile Similarity: Results may show large deviations when multiple source profiles exhibit similar chemical signatures (collinearity) [57]
• Variable Performance: Contribution estimates for parameters with high variability (Fe, Mn, Cl⁻, SO₄²⁻ and NH₄⁺) may show significant differences compared to PMF models [55]
• Unexplained Variability: The model may attribute substantial contributions to unexplained variability, particularly in complex hydrogeological settings [55]
• Data Requirements: Requires sufficient sample size (typically >50 samples) with comprehensive hydrochemical characterization for reliable results [53]

Recommendations for Optimal Implementation

• Study Design Phase:

  • Conduct preliminary hydrogeological assessment to understand aquifer characteristics and flow regimes
  • Implement stratified sampling based on land-use types and aquifer confinement conditions [45]
  • Ensure adequate spatial coverage with approximately 1 sample per km² in regional studies [53]

• Data Quality Assurance:

  • Validate dataset suitability through KMO (>0.5) and Bartlett's test (p<0.05) before PCA implementation [53]
  • Incorporate quality control samples (blanks, duplicates, standards) throughout analytical procedures

• Model Application and Validation:

  • Integrate land-use parameters through GIS to strengthen source identification [57]
  • Apply multiple receptor models (PMF, APCS-MLR) for comparative assessment where feasible [55]
  • Validate findings through isotope analysis when source ambiguity remains [54] [59]

The PCA-APCS-MLR methodology represents a robust, accessible approach for quantitative source apportionment in groundwater systems. When properly implemented with attention to its limitations and integration with complementary techniques, it provides valuable insights for developing targeted groundwater protection strategies and pollution management measures.

Beyond the Basics: Overcoming PCA Limitations and Optimizing Your Analysis

Principal Component Analysis (PCA) is a powerful linear dimensionality reduction technique widely applied in groundwater chemistry to identify contamination sources and understand hydrogeochemical processes [37]. In practice, however, researchers frequently encounter three significant pitfalls that can compromise their interpretations: sensitivity to scale variance, the potentially invalid orthogonality assumption of principal components, and the inability to capture non-linear relationships in data [7]. These challenges are particularly pronounced in groundwater datasets where parameter measurements span different units and scales, and where complex, non-linear geochemical processes often govern solute behavior. This protocol outlines targeted strategies to address these limitations, enhancing the reliability of PCA in groundwater source identification.

Pitfall 1: Scale Variance and Sensitivity

Scale variance refers to PCA's sensitivity to the measurement units and magnitudes of different variables. Groundwater datasets typically include parameters with vastly different numerical ranges (e.g., pH ~0-14, electrical conductivity ~100-10,000 µS/cm, major ions ~1-1000 mg/L). Without proper preprocessing, variables with larger variances will dominate the first principal components regardless of their true chemical significance [10].

Recommended Preprocessing Protocol

Step 1: Data Cleaning and Transformation

• Handle missing data using appropriate imputation methods (e.g., k-means clustering for pattern recognition or regression imputation)
• Apply log-transformation to right-skewed parameters (e.g., nitrate, chloride) to normalize distributions: X_transformed = log(X_original)

Step 2: Standardization Techniques Apply one of the following standardization methods before PCA:

• Z-score Standardization: Center to mean=0 and scale to standard deviation=1 X_std = (X - μ) / σ
• Min-Max Scaling: Scale to a fixed range [0, 1] X_scaled = (X - X_min) / (X_max - X_min)
• Robust Scaling: Use median and interquartile range for outlier-resistant scaling

Step 3: Validation

• Compare correlation matrices before and after standardization
• Conduct sensitivity analysis using different scaling methods

Table 1: Impact of Scaling Methods on Groundwater PCA

Scaling Method	Best Use Case	Advantages	Limitations
Z-score	Most groundwater applications	Preserves outlier information, interpretable	Sensitive to extreme outliers
Min-Max	Parameters with known valid ranges	Preserves original distribution shape	Compresses variance with outliers
Robust Scaling	Datasets with significant outliers	Reduces outlier influence	May obscure genuine extreme values

Pitfall 2: Orthogonality Assumption

Traditional PCA assumes principal components are orthogonal linear combinations of original variables. In groundwater systems, this constraint may force artificial separation of processes that are naturally correlated or produce components with evenly distributed variance, complicating interpretation [60].

Advanced Methodologies

Option A: Varimax Rotation

• Apply orthogonal rotation to maximize variance of squared loadings
• Enhances interpretability by creating simpler component structures
• Implementation protocol:
  • Perform standard PCA to extract initial components
  • Retain components with eigenvalues >1 (Kaiser criterion)
  • Apply varimax rotation to retained components
  • Interpret rotated loadings >|0.5| as significant

Option B: Orthogonal Nonlinear PCA (O-NLPCA) For complex groundwater systems with non-linear correlations:

• Use Gram-Schmidt orthogonalization within neural network-based NLPCA
• Ensures components capture maximum variance sequentially
• Maintains orthogonality while modeling non-linear relationships [60]

Option C: Factor Analysis

• Employs a specific model with error terms: X = LF + ε
• Does not require component orthogonality
• Allows for correlated factors that may better represent natural processes

Table 2: Comparing Methods to Address Orthogonality Constraints

Method	Key Mechanism	Interpretability	Implementation Complexity
Varimax Rotation	Axis rotation to maximize loadings variance	High	Low
O-NLPCA	Orthogonalization in non-linear feature space	Moderate	High
Factor Analysis	Explicit error model with correlated factors	Moderate	Moderate

Pitfall 3: Non-Linear Data Relationships

Conventional PCA employs linear transformations, failing to capture complex non-linear relationships common in groundwater systems (e.g., mineral saturation indices, redox thresholds, and biological degradation pathways) [60] [25].

Nonlinear PCA (NLPCA) Framework

Option A: Autoassociative Neural Networks

• Five-layer neural network architecture with bottleneck layer
• Protocol for groundwater applications:
  • Network Structure: Input → Encoding → Bottleneck → Decoding → Output
  • Training: Minimize reconstruction error between input and output
  • Component Extraction: Bottleneck layer activations represent non-linear PCs
  • Validation: Use independent test set to prevent overfitting

Option B: Kernel PCA

• Maps data to higher-dimensional feature space using kernel functions
• Implementation protocol:
  • Select appropriate kernel (polynomial, RBF, sigmoid)
  • Compute kernel matrix K where Kij = k(xi, xj)
  • Solve eigenvalue problem for kernel matrix
  • Project original data onto kernel principal components

Option C: Multi-scale Nonlinear Strategy

• Combine wavelet decomposition with NLPCA [60]
• Separate deterministic features (approximation coefficients) from stochastic noise (detail coefficients)
• Apply orthogonal NLPCA to approximation coefficients for robust feature extraction

NLPCA Network Architecture

Integrated Case Study Application

Study Background

Application of the proposed framework to the Qujiang River Basin, China, where groundwater chemistry is influenced by natural rock weathering, agricultural activities, and industrial discharges [7].

Implementation Workflow

Integrated PCA Workflow for Groundwater

Results and Interpretation

The integrated approach successfully identified three contamination sources with quantifiable contributions:

• Natural rock weathering (26.3%)
• Agricultural and domestic activities (38.5%)
• Industrial wastewater discharges (35.2%)

Spatial validation using Mantel tests confirmed strong correlations between identified sources and land use patterns, demonstrating the framework's effectiveness in resolving complex groundwater contamination scenarios [7].

The Scientist's Toolkit

Table 3: Essential Research Reagent Solutions for Groundwater PCA

Tool/Category	Specific Examples	Function/Purpose
Statistical Software	R (FactoMineR, vegan), Python (scikit-learn), The Unscrambler [61]	Multivariate analysis implementation
Data Preprocessing	Z-score standardization, Min-Max scaling, Robust scaling, Log-transformation	Address scale variance and outlier effects
Non-linear Extensions	Kernel PCA (polynomial, RBF), Autoassociative Neural Networks, Multi-scale PCA [60] [25]	Capture complex non-linear relationships
Interpretation Aids	Varimax rotation, Promax rotation, Factor analysis	Enhance component interpretability
Validation Methods	Mantel test [7], Cross-validation, Bootstrap resampling	Verify spatial patterns and model robustness
Visualization Tools	Biplots, Scree plots, Piper diagrams [62], Spatial distribution maps	Communicate findings and identify patterns

This protocol provides a comprehensive framework for addressing three fundamental PCA limitations in groundwater chemistry research. Key recommendations include:

• Always preprocess data using appropriate scaling methods to mitigate scale variance effects
• Assess data linearity before analysis and implement NLPCA when non-linear relationships are suspected
• Apply factor rotations to enhance interpretability while being mindful of geological plausibility
• Validate results through spatial analysis and domain knowledge integration

The integrated PCA-PMF-Mantel framework demonstrated in the Qujiang River Basin case study provides a transferable template for groundwater source identification in complex hydrogeological settings [7]. By systematically addressing these common pitfalls, researchers can significantly enhance the reliability and interpretability of PCA applications in groundwater chemistry.

In the field of groundwater chemistry research, Principal Component Analysis (PCA) is a pivotal statistical tool for simplifying complex datasets, identifying contamination sources, and understanding hydrochemical processes. A critical step in PCA is determining the number of significant principal components to retain, as this decision directly influences the interpretation of underlying environmental factors. Retaining too many components can lead to model overfitting by including noise, while retaining too few can result in the loss of meaningful information [63]. This article examines three established methods for determining component retention: the Scree Plot, the Eigenvalue Greater Than One criterion (Kaiser-Guttman test), and the Broken Stick Model. We frame this discussion within the context of groundwater chemistry, providing a structured protocol to help researchers select the most appropriate method for their specific studies, thereby ensuring robust and interpretable results in the analysis of water quality and contamination sources.

Theoretical Background of Component Retention Methods

The Principle of Principal Component Analysis in Groundwater Studies

PCA is a dimensionality reduction technique that transforms a set of potentially correlated variables into a set of linearly uncorrelated variables called principal components. These components are ordered so that the first few retain most of the variation present in the original dataset [63]. In groundwater research, where datasets often include numerous chemical parameters (e.g., major ions, nutrients, trace metals), PCA helps identify the dominant processes controlling water chemistry, such as water-rock interactions, anthropogenic contamination, or agricultural runoff [20] [10]. The core challenge lies in distinguishing the components that represent true environmental signals from those that merely represent background noise.

Several heuristic and statistical methods have been developed to address the challenge of component retention. These methods aim to balance model accuracy with simplicity and interpretability [63]. No single method is universally ideal, and each has tendencies to either over-estimate or under-estimate the true dimensionality of a dataset [63]. The following sections detail the three focal methods of this article, but researchers should be aware that other techniques exist, including cumulative percentage of total variance, Bartlett's test for equality of eigenvalues, and cross-validation [63].

Comparative Analysis of Retention Methods

The following table summarizes the key characteristics, advantages, and limitations of the three primary component retention methods.

Table 1: Comparison of Methods for Determining Significant Components in PCA

Method	Theoretical Basis	Decision Rule	Advantages	Limitations
Scree Plot [63]	Visual inspection of the rate of change in eigenvalues.	Retain components before the point where the slope of the eigenvalue curve markedly levels off ("elbow").	Intuitive and graphical; allows for subjective expert judgment.	Subjective; inter-observer variability in identifying the "elbow".
Eigenvalue >1 (Kaiser-Guttman) [63]	Each retained component should explain at least as much variance as a single standardized variable.	Retain all components with an eigenvalue greater than 1.0.	Simple, objective, and easily computable; widely used.	Tends to over-estimate the number of components in datasets with many variables.
Broken Stick Model [63]	Compares observed eigenvalues to those expected from a random distribution of variance.	Retain components for which the observed eigenvalue exceeds the value expected from the random model.	Provides a objective statistical threshold; effective at identifying components explaining non-random variance.	Can be conservative, potentially under-estimating dimensions in some cases.

Protocol for Determining Significant Components in Groundwater Chemistry Studies

The following diagram illustrates the recommended sequential workflow for applying and reconciling the three component retention methods in a groundwater chemistry study.

Step-by-Step Experimental Protocol

Step 1: Data Pre-processing and PCA Execution

• Data Preparation: Compile your groundwater quality dataset. Ensure data completeness and consider appropriate methods for handling non-detects or missing values. Common parameters include major ions (Ca²⁺, Mg²⁺, Na⁺, K⁺, Cl⁻, SO₄²⁻, HCO₃⁻), nutrients (NO₃⁻, NO₂⁻, NH₄⁺), trace metals, and physical parameters (pH, EC, TDS) [64] [10] [65].
• Standardization: Given that groundwater parameters are measured in different units, standardize the data (e.g., z-scores) to have a mean of zero and a standard deviation of one. This prevents variables with larger scales from disproportionately influencing the components [63].
• Perform PCA: Execute the PCA on the correlation matrix (not the covariance matrix) using statistical software (e.g., R, Python, SPSS, SAS). Extract the eigenvalues for all possible components.

Step 2: Apply the Three Retention Methods

• Generate the Scree Plot:

  • Using your software, create a plot with the component number on the x-axis and the corresponding eigenvalue on the y-axis.
  • Visually inspect the plot to identify the "elbow" – the point where the steep decline in eigenvalues begins to level off. The components before this point are candidates for retention. Figure 1 shows an illustrative example.

• Apply the Eigenvalue >1 Criterion:

  • From the PCA output, list all eigenvalues in descending order.
  • Count and note the number of components that have an eigenvalue greater than 1.0.

• Apply the Broken Stick Model:

  • Calculate the expected eigenvalues under the broken stick model. For the k-th component in a p-dimensional dataset, the expected eigenvalue is: EV_k = (1/p) * Σ(1/i) from i=k to p
  • Compare each observed eigenvalue to its corresponding broken stick expected value.
  • Retain all components for which the observed eigenvalue is greater than the expected value.

Step 3: Reconciliation and Final Decision

• Compare Results: Tabulate the number of significant components suggested by each of the three methods.
• Seek Consensus:
  • If all three methods suggest the same number of components (e.g., 3), this provides strong evidence for that being the appropriate dimensionality.
  • If the methods disagree (a common occurrence), proceed to the next step.
• Interpret Divergent Components:
  • Scree Plot vs. Eigenvalue >1: If the eigenvalue criterion suggests more components than the scree plot, examine the variables loading highly on the extra components. If these variables form a interpretable pattern related to a known hydrochemical process (e.g., a component high on NO₃⁻ and K⁺ related to agricultural fertilizer [20]), consider retaining it.
  • Broken Stick as Arbiter: The Broken Stick Model often provides a more parsimonious estimate. If it aligns with one of the other methods, this can be used as a tie-breaker.
  • Consider Study Goals: For an exploratory study, a slightly higher dimensionality might be acceptable. For creating a robust Water Quality Index (WQI) with reduced parameters, a more conservative approach (leaning towards the Scree Plot or Broken Stick) may be preferable [66].
• Finalize and Report: Document the final number of components chosen and justify the decision based on the methodological comparison and interpretability within the groundwater context.

Validation and Post-Hoc Analysis

• Interpret Component Loadings: For the final set of retained components, interpret the factor loadings. High loadings indicate a strong contribution of a variable to that component. Try to assign a hydrogeochemical meaning to each component (e.g., "Component 1: Salinity and water-rock interaction," "Component 2: Agricultural nitrate contamination") [20] [10] [65].
• Check Sample Adequacy: Prior to PCA, ensure your data is suitable for factor analysis. Use the Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy (should be >0.6) and Bartlett's test of sphericity (should be significant, p < 0.05) [66].
• Integrate with Other Techniques: Strengthen your findings by integrating PCA with other multivariate statistical methods, such as Hierarchical Cluster Analysis (HCA) or self-organizing maps (SOM), to see if the component-based grouping corresponds to spatial clusters of water samples [64] [65].

The Scientist's Toolkit: Essential Reagents and Materials

Table 2: Key Research Reagent Solutions for Groundwater Chemistry Analysis

Item Name	Function/Explanation
Standard Solutions for Ion Chromatography	Used for the quantification of major anions (Cl⁻, NO₃⁻, SO₄²⁻) and cations (Na⁺, K⁺, Ca²⁺, Mg²⁺) in groundwater samples, which form the core dataset for PCA.
Reference Materials (CRMs)	Certified Reference Materials are essential for quality assurance/quality control (QA/QC) to calibrate instruments and verify the accuracy of analytical results for metals and ions.
Hydrochloric Acid (HCl) / Nitric Acid (HNO₃)	High-purity acids are used for sample preservation, particularly for preventing precipitation of metals and carbonates, and for digesting samples for total metal analysis.
Atomic Absorption Spectrophotometer (AAS)	An instrument for quantifying trace metal concentrations (e.g., As, Pb, Cd, Fe) which are critical parameters for identifying anthropogenic or geogenic contamination sources [67] [65].
Portable Multi-Parameter Meter	Used for in-situ measurement of physical parameters like pH, Electrical Conductivity (EC), Total Dissolved Solids (TDS), and Dissolved Oxygen (DO), which are key input variables for PCA [10] [65].

Application in Groundwater Research: A Case Study Context

The application of these protocols is best illustrated with a hypothetical case study based on real-world research. Imagine a project investigating groundwater contamination in a region with a history of phosphate mining and intensive agriculture, similar to studies in Southern Tunisia [20] or the Sichuan Basin [65].

• Objective: Identify the primary sources and processes affecting groundwater chemistry.
• Data: 33 groundwater samples analyzed for 12 parameters, including Ra-226 (radioactivity), NO₃⁻ (nitrate), Ca²⁺, Mg²⁺, Na⁺, HCO₃⁻, Cl⁻, SO₄²⁻, pH, and EC.
• PCA & Retention Analysis:
  • PCA is run on the standardized dataset.
  • The Scree Plot shows a distinct elbow after the third component.
  • The Eigenvalue >1 rule suggests retaining four components.
  • The Broken Stick Model suggests retaining three components.
• Reconciliation: The fourth component has an eigenvalue of 1.1. Upon inspection, it shows high loadings for K⁺ and a weaker loading for NO₃⁻. Given the agricultural context, this could represent a secondary, less pervasive nutrient leaching signal. The researcher might choose to retain this fourth component for a more detailed model, justified by the Eigenvalue >1 criterion and its interpretability.
• Outcome: The final PCA model identifies four meaningful contamination sources: 1) Phosphate mining (high Ra, SO₄), 2) Agricultural runoff (high NO₃), 3) Geothermal water influence (high Na, HCO₃), and 4) Secondary agricultural activity, providing a robust foundation for targeted mitigation strategies [20].

Determining the number of significant components is a critical, non-automated step in PCA that requires careful consideration. In groundwater chemistry studies, no single retention method is infallible. By applying the Scree Plot, Eigenvalue >1 Criterion, and Broken Stick Model in a sequential protocol, researchers can make an informed and defensible decision. The reconciliation of results from these methods, guided by hydrogeochemical expertise, ensures that the final PCA model is both statistically sound and environmentally meaningful, ultimately leading to more accurate identification of contamination sources and processes governing groundwater quality.

In the analysis of groundwater chemistry using Principal Component Analysis (PCA), the initial extracted components are often not easily interpretable because they represent mathematical combinations of all original variables, such as major ions (Ca²⁺, Mg²⁺, Na⁺, K⁺, Cl⁻, SO₄²⁻, HCO₃⁻) and physicochemical parameters (pH, EC, TDS). Factor rotation is a crucial step that transforms these initial components into a more interpretable structure without altering the underlying statistical relationships. This transformation enhances the scientific utility of PCA by aligning factors with plausible geochemical processes, thereby allowing researchers to more accurately identify contamination sources, including natural water-rock interactions, agricultural inputs, industrial discharges, and domestic wastewater.

The primary objective of rotation is to achieve a "simple structure," a concept formalized by Thurstone that describes an ideal pattern where each variable loads highly on a single factor and has near-zero loadings on others [68]. This clarity is particularly valuable in groundwater studies, where distinguishing between correlated anthropogenic influences (e.g., agricultural and domestic sewage) can be challenging. Without rotation, factors often remain complex, with many variables exhibiting moderate cross-loadings, thereby obscuring the distinct geochemical processes they represent and complicating the identification of pollution sources.

Theoretical Foundations of Rotation Techniques

Orthogonal Rotation (Varimax)

Varimax is the most widely used orthogonal rotation method in geochemical studies. It operates under the constraint that the resulting factors remain uncorrelated, simplifying the mathematical model and its interpretation. The algorithm maximizes the variance of the squared loadings within each factor, which tends to polarize loadings toward either high or low values. This effect enhances interpretability by creating a clear distinction between variables that are strongly associated with a factor and those that are not. For example, in a study of the Qujiang River Basin, PCA with Varimax rotation successfully isolated distinct factors representing natural rock weathering, agricultural and domestic activities, and industrial wastewater discharge [7]. The orthogonal assumption is suitable when the underlying geochemical processes are believed to be independent, such as when a specific anthropogenic source affects the aquifer without influencing other concurrent processes.

Oblique Rotation (Promax, Oblimin, Quartimin)

In contrast, oblique rotation methods (e.g., Promax, Oblimin) relax the constraint of factor independence, allowing the rotated factors to be correlated. This approach often provides a more realistic representation in environmental systems where geochemical processes are frequently interrelated. For instance, agricultural runoff (containing nitrates and potassium) and domestic sewage (containing chlorides and sodium) often co-occur and are hydrologically connected, leading to correlated factors. Oblique rotations can achieve a conceptually simpler structure by permitting these natural correlations, potentially offering a more accurate model of complex groundwater systems. The choice between specific oblique methods (Promax, Quartimin) often depends on the algorithmic approach to achieving simple structure and the nature of the correlation matrix [69] [68].

Comparative Analysis of Rotation Methods

Table 1: Comparison of Orthogonal and Oblique Rotation Methods

Feature	Varimax (Orthogonal)	Oblique Methods (e.g., Promax)
Factor Correlation	Assumes factors are uncorrelated	Allows factors to be correlated
Theoretical Basis	Simplifies model structure; good for independent processes	More realistic for interrelated geochemical processes
Result Interpretation	Simpler; interpret factors directly from loadings	Requires examining both pattern matrix (loadings) and factor correlations
Ideal Use Case	Identifying distinct, independent sources (e.g., a single industrial point source)	Differentiating correlated sources (e.g., agricultural and domestic sewage)
Sample Application	Discriminating between volcanic and siliciclastic components in Campania [69]	Resolving complex anthropogenic mixtures in the Qujiang River Basin [7]

The practical difference between these methods is evident when evaluating their achievement of "simple structure." A comparative application of orthogonal rotations (Varimax, Quartimax, Equamax) and oblique rotations on the same dataset revealed that oblique rotation often satisfies more conditions of simple structure, particularly when the underlying geochemical factors are expected to be correlated [68]. Brown's five criteria for simple structure provide a framework for this evaluation, including requirements that each variable should have at least one near-zero loading, and for each factor pair, most variables should load significantly on only one factor [68].

A key practical consideration is that the threshold for defining a "significant" loading (often ≥0.3 or ≥0.5) can influence which rotation method appears optimal. Higher thresholds (e.g., ≥0.5) tend to produce cleaner, more interpretable factor patterns regardless of the rotation method used, though the choice of threshold should be guided by the specific research context and dataset characteristics [68].

Practical Protocols for Rotation in Groundwater Studies

Workflow for Applying and Selecting Rotation Methods

The following protocol outlines a systematic approach for applying factor rotation in groundwater chemistry studies using PCA:

Step 1: Data Preparation and Initial PCA

• Normalize the groundwater chemistry dataset to address skewed distributions and extreme outliers. Common transformations include Normal Score Transformation (NST) or log-transformation, which stabilize variance and improve multivariate analysis performance [69].
• Perform initial PCA extraction using components with eigenvalues greater than 1 (Kaiser criterion) or based on scree plot interpretation.

Step 2: Apply Multiple Rotation Techniques

• Implement Varimax rotation to achieve an orthogonal solution.
• Implement at least one oblique rotation method (e.g., Promax or Direct Oblimin).
• Most statistical software (R, SPSS, Python) supports these rotation options through built-in functions.

Step 3: Evaluate Simple Structure

• Apply Brown's five criteria to assess which rotation method achieves the best simple structure [68]:
  • Each variable has at least one loading between -0.10 and 0.10.
  • Each factor has at least as many near-zero loadings as there are factors.
  • For each factor pair, most variables have significant loadings on one factor and near-zero loadings on the other.
  • For four or more factors, most factor pairs should have many variables with near-zero loadings on both.
  • For each factor pair, only a few variables have significant loadings on both ("complex variables").

Step 4: Examine Factor Correlations and Interpretability

• If using oblique rotation, examine the factor correlation matrix. Correlations exceeding |0.3| suggest oblique rotation is more appropriate.
• Evaluate the geological and environmental plausibility of each solution. The optimal rotation should yield factors that align with recognizable hydrogeochemical processes.

Step 5: Select and Report the Final Method

• Document the selected rotation method and justifications based on statistical and conceptual grounds.
• Report both the rotated factor loadings and, for oblique rotation, the factor correlations.

Decision Framework for Rotation Selection

Application in Groundwater Chemistry Research

Case Study: Qujiang River Basin, China

An integrated PCA-PMF-Mantel test framework was applied to identify groundwater pollution sources in the Qujiang River Basin. PCA with rotation successfully identified three primary sources influencing groundwater chemistry: natural rock weathering (26.3% contribution), agricultural and domestic activities (38.5%), and industrial wastewater discharge (35.2%) [7]. The rotated factors clearly distinguished these sources, with the anthropogenic factors (agricultural/domestic and industrial) showing potential correlation, suggesting that oblique rotation might have been appropriate. The spatial distribution of these factors, validated using Mantel tests, demonstrated how rotated PCA results could be linked to specific land-use patterns (e.g., farmland in midstream areas and industrial zones downstream).

Case Study: Campania Region, Italy

In the highly anthropized Campania region, PCA was applied to over 7,000 topsoil samples to discriminate between natural and anthropogenic contamination sources. The application of Varimax rotation successfully isolated four independent sources controlling geochemical variability: two distinct volcanic districts, a siliciclastic component, and an anthropogenic component [69]. The orthogonal assumption was appropriate here, as the volcanic sources are geologically distinct from both the siliciclastic formations and the anthropogenic influences. The spatialization of rotated PCA scores created composite maps that effectively visualized the coexistence and predominance of each component across the region, providing valuable insights for environmental risk assessment.

Case Study: Amargosa Desert Region, USA

Multivariate statistical methods, including PCA with rotation, were applied to groundwater chemistry data from the Amargosa Desert region to identify hydrochemical processes and facies. The rotated factor loadings and scores were presented as biplots, demonstrating relationships between variables and sampling locations [70]. This approach revealed a distinct groundwater chemical signature along the extended flowpath of Fortymile Wash, suggesting potential interaction with a fault line and representing a relic of water that infiltrated during past pluvial periods. This case demonstrates how rotated PCA can reconstruct paleohydrological conditions and identify groundwater flow paths.

The Scientist's Toolkit: Essential Reagents and Materials

Table 2: Essential Analytical Tools for Groundwater PCA Studies

Tool/Category	Specific Examples	Function in Analysis
Statistical Software	R (psych, FactoMineR packages), SPSS, Python (scikit-learn)	Performs PCA extraction and multiple rotation options
Normalization Methods	Normal Score Transformation (NST), Box-Cox, log-transformation	Addresses skewed data distributions and extreme outliers
Rotation Algorithms	Varimax, Quartimax, Equamax (orthogonal); Promax, Direct Oblimin (oblique)	Enhances interpretability of factor patterns
Geochemical Modeling	PHREEQC, Geochemist's Workbench	Validates interpreted factors against known hydrochemical processes
Spatial Analysis Tools	GIS (ArcGIS, QGIS), inverse distance weighting interpolation	Maps factor scores to visualize spatial patterns of sources
Field Sampling Equipment	Multiparameter water quality analyzer (e.g., HANNA HI9828), submersible pumps	Collects accurate field measurements of pH, EC, temperature for robust input data

The strategic application of Varimax and oblique rotations significantly enhances the interpretability of PCA results in groundwater chemistry studies. While Varimax provides a simpler, more constrained solution ideal for independent sources, oblique methods offer greater flexibility for modeling correlated anthropogenic influences commonly encountered in contaminated aquifers. The systematic protocol outlined in this document—emphasizing comparative application, evaluation of simple structure, and geological plausibility—provides researchers with a robust framework for selecting the optimal rotation method. When properly executed, factor rotation transforms complex multivariate data into actionable insights about contamination sources, supporting effective groundwater resource management and remediation strategies.

Principal Component Analysis (PCA) is a foundational technique in groundwater chemistry research for dimensionality reduction and source apportionment. However, its fundamental limitation lies in its inherent linearity assumption, which often fails to capture the complex, non-linear interactions that characterize hydrochemical systems [25]. These non-linear relationships arise from intricate water-rock interactions, redox processes, mixing behaviors, and anthropogenic influences that govern groundwater composition.

Kernel PCA (KPCA) addresses this limitation by applying a nonlinear mapping to transform input data into a higher-dimensional feature space where linear separation becomes possible [25] [71]. This sophisticated approach enables researchers to uncover latent structures and patterns in hydrochemical data that conventional PCA cannot detect, providing more accurate insights into pollution sources, mixing processes, and geochemical evolution along flow paths.

Theoretical Foundation

The Kernel Method Concept

The mathematical foundation of KPCA centers on mapping original data points from their input space to a higher-dimensional feature space using a nonlinear function φ, then performing standard PCA in this new space [72]. The "kernel trick" enables this operation without explicitly computing the coordinates in the feature space, instead relying on the inner products between all pairs of data points [73].

For a dataset with n observations, KPCA involves solving the eigenvalue problem for the kernel matrix K, where each element K(i,j) represents the inner product between transformed data points [72]. The resulting principal components in feature space capture directions of maximum variance while accounting for nonlinear relationships, making it particularly suitable for complex groundwater systems where parameters like salinity, ion concentrations, and redox indicators exhibit interdependent, non-linear behaviors [25].

Comparative Advantages in Groundwater Context

In groundwater chemistry, KPCA offers distinct advantages over linear PCA. Traditional PCA may oversimplify the complex relationships between hydrochemical parameters, particularly in coastal aquifers where seawater intrusion creates strong nonlinear gradients, or in contaminated sites where multiple pollution sources mix non-uniformly [25] [74]. KPCA effectively models these complex interactions, providing more accurate representations of the underlying hydrochemical processes.

The capability of KPCA to handle nonlinear relationships makes it particularly valuable for identifying end-member compositions in groundwater mixing models, tracing contaminant plumes with non-conservative behavior, and understanding the complex interplay of geogenic and anthropogenic factors controlling groundwater quality [25] [4].

Practical Implementation for Groundwater Chemistry

Kernel Selection and Parameterization

Selecting an appropriate kernel function is critical for successful KPCA application in groundwater studies. Different kernel types capture different types of nonlinear relationships, and their performance varies depending on the specific hydrochemical context.

Table 1: Kernel Functions for Groundwater Chemistry Applications

Kernel Type	Mathematical Form	Hydrochemical Applications	Advantages
Polynomial	K(x,y) = (xᵀy + c)ᵈ	Coastal aquifer systems with seawater intrusion [25]	Effective for complex hierarchical data structures
Radial Basis Function (RBF)	K(x,y) = exp(-γ		x-y		²)	General hydrochemical differentiation [25] [72]	Handles complex nonlinear patterns; good default choice
Sigmoid	K(x,y) = tanh(αxᵀy + c)	Specific ion relationships and redox processes	Similar to neural network activation functions
Linear	K(x,y) = xᵀy	Baseline comparison and simple systems [25]	Standard PCA equivalent; useful for benchmarking

Parameter optimization is essential for achieving meaningful results. For polynomial kernels, the degree parameter (d) must be carefully selected, while for RBF kernels, the gamma parameter (γ) controls the influence of individual samples. In groundwater quality assessment, the polynomial kernel has demonstrated superior performance in preserving variance and reducing dimensionality for coastal aquifer systems [25].

Experimental Protocol for Groundwater Source Apportionment

The following step-by-step protocol outlines a standardized approach for implementing KPCA in groundwater chemistry research:

Step 1: Data Collection and Preprocessing

• Collect groundwater samples following standardized procedures (e.g., USEPA protocols) [25]
• Analyze for major ions (Ca²⁺, Mg²⁺, Na⁺, K⁺, Cl⁻, SO₄²⁻, HCO₃⁻), nutrients (NO₃⁻, NO₂⁻, NH₄⁺), trace elements, and field parameters (pH, EC, temperature) [25] [75]
• Perform min-max normalization to scale all parameters to a [0,1] range to ensure uniform contribution without discarding natural hydrochemical variability [25]

Step 2: Kernel Selection and Model Training

• Test multiple kernel types (linear, polynomial, RBF, sigmoid) using a subset of data
• Evaluate kernel performance based on variance preservation and clustering coherence
• Select optimal kernel parameters through cross-validation, focusing on the cumulative contribution rate (typically target ≥85-95% variance) [72]

Step 3: Model Application and Interpretation

• Project complete dataset using selected KPCA model
• Identify significant principal components capturing major hydrochemical patterns
• Interpret loadings to relate transformed variables to original hydrochemical parameters

Step 4: Validation and Integration

• Validate KPCA results with complementary methods (e.g., isotopic tracers, geochemical modeling) [74] [45]
• Integrate findings with spatial analysis and hydrogeological context for comprehensive interpretation

Groundwater KPCA Analysis Workflow: This diagram illustrates the comprehensive protocol for applying Kernel PCA to groundwater chemistry data, from sample collection through to decision support.

Case Studies in Groundwater Research

Water Quality Assessment in Coastal Aquifers

A recent study demonstrated KPCA's effectiveness for assessing groundwater quality in the coastal aquifers of Al-Qatif, Saudi Arabia [25] [76]. Researchers collected 39 groundwater samples from shallow and deep wells and analyzed them for key physicochemical parameters. After testing six kernel types, the polynomial kernel proved most effective in preserving variance and reducing dimensionality while capturing the non-linear relationships caused by seawater intrusion and anthropogenic activities.

The KPCA-based Water Quality Index (WQI) successfully classified wells into 'Very Bad,' 'Bad,' and 'Medium' categories, with specific wells scoring WQI = 25.51 ("Very Bad"), 46.7 ("Bad"), and 56.75 ("Medium") [25]. The analysis revealed that salinity and electrical conductivity (EC) presented poor Sub-Index scores, reflecting the impact of seawater intrusion and over-extraction, while pH consistently showed high SI values (100), indicating natural buffering capacity [25]. This approach provided more nuanced understanding of groundwater quality dynamics compared to traditional linear methods.

Hybrid Models for Source Identification

KPCA has been successfully integrated with other machine learning techniques to enhance groundwater source identification. A hybrid KPCA-ISSA-SVM model demonstrated superior performance in identifying sources of mine water inrush using hydrochemical indicators [72]. The model achieved accuracy of 90.75%, precision of 0.90, recall of 0.88, and Kappa coefficient of 0.89, significantly outperforming standard SVM and other comparative models.

In this application, KPCA served as a dimensionality reduction tool, processing nine conventional hydrochemical indicators (Ca²⁺, Mg²⁺, Na⁺+K⁺, HCO₃⁻, Cl⁻, SO₄²⁻, total hardness, alkalinity, and pH) to eliminate redundancy between discriminant indices, simplify model structure, and enhance computational efficiency [72]. This case highlights KPCA's value in preprocessing complex hydrochemical data before application of classification algorithms.

Table 2: Performance Metrics of KPCA-Based Models in Groundwater Applications

Application Context	Model Type	Key Performance Metrics	Advantages over Traditional Methods
Coastal Aquifer Quality Assessment	KPCA-WQI (Polynomial kernel)	Effective classification of wells into quality categories; identification of salinity/EC as key discriminators [25]	Captures non-linear seawater intrusion gradients; more nuanced quality classification
Mine Water Inrush Source Identification	KPCA-ISSA-SVM	Accuracy: 90.75%, Precision: 0.90, Recall: 0.88, Kappa: 0.89 [72]	Enhanced classification performance; effective dimensionality reduction
Industrial Process Monitoring	Reduced KPCA (Fractal dimension)	Reduced storage space and execution time; maintained fault detection performance [73]	Handles large-scale nonlinear data efficiently

The Researcher's Toolkit

Essential Analytical Framework

Implementing KPCA in groundwater research requires both standard hydrochemical analysis tools and specialized computational resources:

Table 3: Essential Research Reagents and Solutions for Groundwater KPCA Studies

Category	Specific Items	Function in KPCA Analysis
Field Measurement Equipment	Multiparameter meter (pH, EC, T), submersible pumps, flow-through cells	In-situ parameter measurement; representative sample collection [25]
Laboratory Analysis	ICP-OES/MS, ion chromatography, spectrophotometers	Major ion, trace element, and nutrient quantification [25] [71]
Sample Preservation	0.45μm membrane filters, refrigeration at 4°C, chemical preservatives	Maintain sample integrity between collection and analysis [25]
Computational Tools	MATLAB, Python (scikit-learn), R with kernel methods library	KPCA implementation, kernel parameter optimization, visualization [73]
Validation Methods	Stable isotope analysis, geochemical modeling, spatial interpolation	Verify KPCA results against independent methods [45] [75]

Computational Considerations and Optimization

Groundwater datasets often present computational challenges for KPCA implementation, particularly with large sample sizes or numerous parameters. The time complexity of standard KPCA is O(n³), while storage complexity is O(n²), creating potential bottlenecks with extensive monitoring networks [73].

Reduced KPCA (RKPCA) approaches address these limitations through data reduction techniques that retain the most informative observations. The fractal dimension method has proven effective for this purpose, significantly reducing storage space and execution time while maintaining detection performance [73]. For the Tennessee Eastman Process benchmark dataset, RKPCA achieved approximately 80% reduction in execution time and 65% reduction in storage requirements while maintaining fault detection capabilities [73].

Kernel PCA represents a significant advancement in the multivariate analysis toolbox for groundwater chemistry research. By effectively handling the nonlinear relationships inherent in hydrochemical systems, it enables more accurate source apportionment, quality assessment, and process understanding than traditional linear methods. The integration of KPCA with other machine learning techniques, as demonstrated in hybrid models, further expands its utility for addressing complex groundwater challenges.

Future developments in KPCA applications for groundwater research will likely focus on several key areas: (1) improved kernel functions specifically designed for hydrochemical data structures, (2) enhanced computational efficiency for large-scale monitoring networks, (3) tighter integration with process-based geochemical models, and (4) development of standardized protocols for different groundwater contexts. As these methodological advances continue, KPCA is poised to become an increasingly valuable tool for researchers and practitioners working to understand and protect vital groundwater resources.

Principal Component Analysis (PCA) is a cornerstone of multivariate data analysis in environmental science, used for dimensionality reduction, exploratory data analysis, and identifying latent patterns in complex datasets. However, conventional PCA (cPCA) possesses a critical vulnerability: extreme sensitivity to outliers and missing data. This limitation is particularly problematic in groundwater chemistry research, where datasets frequently contain anomalous measurements arising from technical errors, sampling irregularities, or genuine but extreme geochemical conditions. These outliers can disproportionately influence the principal components, potentially yielding misleading interpretations about hydrogeochemical processes and contaminant sources [77] [78] [79].

Robust Principal Component Analysis (rPCA) addresses this fundamental weakness by employing statistical techniques that are resistant to outliers and missing data. The core conceptual framework of rPCA involves decomposing a data matrix ((X)) into two distinct components: a low-rank matrix ((L)) representing the true underlying structure of the data, and a sparse matrix ((S)) capturing the outliers and noise [80] [77]. This separation can be represented as (X = L + S). Unlike cPCA, which is highly sensitive to anomalous observations, rPCA algorithms ensure that the extracted principal components reflect the covariance structure of the majority of the data, thereby providing a more accurate representation of the genuine hydrogeochemical signals [78].

Within the context of groundwater chemistry research, where data quality is paramount for distinguishing between natural and anthropogenic influences on water quality, rPCA offers a statistically sound framework for identifying outlier samples and managing missing data. This ensures that subsequent analyses, such as contaminant source apportionment and hydrochemical facies classification, are based on a reliable foundation.

Theoretical Foundations: Bridging Convex and Nonconvex Optimization

The theoretical development of Robust PCA is grounded in advanced optimization theory. Early formulations framed the problem as a convex optimization challenge, aiming to recover the low-rank matrix (L) from highly corrupted measurements (X). This convex approach leverages the nuclear norm (a surrogate for rank) and the (\ell_1)-norm (to enforce sparsity in the outlier matrix (S)) [80].

Subsequent research has bridged convex and nonconvex optimization approaches, delivering improved theoretical guarantees for the convex programming approach in low-rank matrix estimation, even in the presence of random noise, gross sparse outliers, and missing data [80]. When the underlying matrix (representing the true, clean data) is well-conditioned, incoherent, and of constant rank, convex programs can achieve near-optimal statistical accuracy in terms of both Euclidean loss and the (\ell_{\infty}) loss. This robustness holds even when a significant fraction of observations are corrupted by outliers of arbitrary magnitude [80].

A key analytical insight involves bridging the convex program and an auxiliary nonconvex optimization algorithm. This connection helps explain why rPCA can effectively separate the low-rank signal (e.g., the dominant geochemical processes governing groundwater composition) from the sparse anomalies (e.g., contamination events or sampling errors) with high probability, provided the outliers are sufficiently sparse and the singular vectors of the low-rank matrix are sufficiently spread out [80].

Comparative Analysis: Robust PCA vs. Classical PCA

The following table summarizes the critical differences between classical and Robust PCA, highlighting why rPCA is superior for managing data quality issues.

Table 1: Comparative characteristics of Classical PCA and Robust PCA.

Feature	Classical PCA (cPCA)	Robust PCA (rPCA)
Sensitivity to Outliers	Highly sensitive; a single outlier can skew components [77] [79].	Resistant to outliers; components represent majority data structure [78].
Handling Missing Data	Requires pre-processing (e.g., imputation), which can introduce bias.	Some algorithms can handle missing data directly within the optimization framework.
Underlying Assumption	Data is from a Gaussian-like distribution without extreme outliers.	Data comprises a low-rank structure corrupted by sparse, large-magnitude noise.
Core Mathematical Approach	Singular Value Decomposition (SVD) of covariance matrix.	Optimization to decompose data into low-rank (L) and sparse (S) matrices [80].
Result Stability	Unstable in the presence of outliers [79].	Stable; provides consistent results even with corrupted data.
Primary Use Case	Clean, well-behaved datasets.	Noisy, real-world datasets with outliers and missing values, like groundwater chemistry [81] [78].

Application Notes: rPCA for Groundwater Chemistry Data

Groundwater chemistry datasets are inherently complex, influenced by multiple natural and anthropogenic factors. The application of rPCA in this domain is exemplified by research in the Dawen River Basin, which aimed to identify sources of chemical constituents. While this study used Self-Organizing Maps (SOM) and Positive Matrix Factorization (PMF), it highlighted the challenge of quantitatively differentiating overlapping influences from geological processes, agricultural activities, and industrial pollution [81]. rPCA serves as a powerful preliminary tool for such analyses by first cleaning the dataset of outliers that could otherwise confound these finer source apportionment techniques.

In practice, rPCA can effectively identify groundwater samples that deviate significantly from the dominant hydrochemical patterns. These outliers could represent:

• Technical Artifacts: Sample contamination or analytical errors.
• Geochemical Anomalies: Localized contamination plumes (e.g., from a point source like a leaking tank).
• Unique Hydrogeological Conditions: Samples influenced by a unique water-rock interaction not common in the rest of the aquifer.

By flagging or down-weighting these samples, rPCA ensures that the subsequent clustering or factor analysis (e.g., with SOM or PMF) is more robust and interpretable, leading to a more accurate quantification of source contributions, such as the 26.8% from agricultural activities and 13.6% from mining operations identified in the Dawen River study [81].

Experimental Protocols and Workflow

Protocol 1: Outlier Detection in a Groundwater Chemistry Dataset

This protocol details the steps for using rPCA to detect outlier samples in a groundwater geochemical dataset, using methods adapted from RNA-seq data analysis which also deals with high-dimensional data and small sample sizes [78].

1. Data Preparation and Pre-processing

• Data Matrix Construction: Assemble your data into an (n \times p) matrix, where (n) is the number of groundwater samples and (p) is the number of measured variables (e.g., pH, EC, Na⁺, Ca²⁺, Cl⁻, NO₃⁻, etc.).
• Standardization: Center and scale all variables to unit variance. This is critical when variables are measured on different scales (e.g., ppm for ions and µS/cm for EC).

2. Applying Robust PCA

• Algorithm Selection: Select an appropriate rPCA algorithm. The PcaGrid and PcaHubert algorithms (available in the rrcov R package) are highly effective. PcaGrid is noted for its low false positive rate, while PcaHubert has high sensitivity [78].
• Execution: Run the chosen rPCA function on the standardized data matrix. Specify the number of components to retain, often determined by cross-validation or scree plots.

3. Identifying Outlier Samples

• Distance Plot: rPCA outputs a robust distance for each sample. Plot the orthogonal distance (distance to the robust PCA space) against the score distance (distance within the robust PCA space).
• Outlier Classification: Samples with significantly large distances in either measure are classified as outliers. The rPCA algorithm provides statistical cut-offs for flagging outliers [78].

4. Post-Outlier Analysis

• Inspection: Manually inspect the flagged outlier samples. Examine their original geochemical parameters to understand the reason for their divergence.
• Decision Making: Decide whether to remove, correct, or retain outliers based on the investigative findings. If the outlier is due to a technical error, removal is justified. If it represents a genuine, rare geochemical environment, it may be retained but noted for separate analysis.
• Downstream Analysis: Proceed with your primary analysis (e.g., standard PCA, clustering, PMF) on the cleaned dataset.

Protocol 2: A Simple Iterative rPCA Algorithm

For environments where specialized packages are unavailable, the following iterative algorithm provides a sound, heuristic approach to robust PCA [79].

1. Initial PCA and Projection

• Compute classical PCA on the standardized data matrix (X) and retain the top (d) principal components.
• Project the data onto this (d)-dimensional subspace.

2. Outlier Rejection Loop

• Calculate the squared Euclidean norm (or Mahalanobis distance) of the projection for each data point.
• Identify data points whose projection norm is "too large." A common heuristic is to flag points beyond the 95th percentile.
• At random, remove one of the flagged outlier points. The random removal helps avoid bias from a single influential outlier [79].
• Repeat this process for a pre-defined number of iterations (e.g., (\bar{T} = n-1), where (n) is the sample size) or until a stopping criterion is met.

3. Final Estimation

• After the iterative removal of potential outliers, perform a final classical PCA on the remaining subset of "clean" data points.
• This final PCA model, based on the majority of the data, provides a robust estimate of the principal components.

Table 2: The Scientist's Toolkit: Essential Computational Reagents for rPCA.

Tool / Reagent	Function / Description	Application Context
R Statistical Software	Open-source environment for statistical computing and graphics.	Primary platform for implementing rPCA algorithms.
rrcov R Package	Provides functions for robust statistical analysis, including PcaGrid and PcaHubert.	Core engine for performing rPCA with proven algorithms [78].
Standardized Geochemical Data	Centered and scaled concentrations of major ions and parameters.	The pre-processed input matrix for rPCA to ensure variables are comparable.
Robust Distance Metrics	Statistical measures (score and orthogonal distance) to quantify a sample's deviation from the robust model.	Objective criteria for classifying a groundwater sample as an outlier [78].
Visualization Tools (e.g., ggplot2)	Libraries for creating high-quality graphs, such as robust distance plots and PCA biplots.	Critical for exploratory data analysis, outlier inspection, and result communication.

Case Study Integration and Data Visualization

The power of rPCA is demonstrated in its application to real-world data. In one case study, researchers applied both cPCA and rPCA (using PcaGrid and PcaHubert) to an RNA-seq dataset profiling gene expression in mouse cerebellum. Both rPCA methods unanimously detected the same two outlier samples, whereas cPCA failed to detect any [78]. This result underscores the objective superiority of rPCA in outlier detection.

Subsequent differential expression analysis was performed before and after the removal of these outliers. The analysis following outlier removal successfully identified biologically relevant genes that were obscured in the analysis of the full, outlier-containing dataset. This was validated by quantitative PCR, confirming that outlier removal significantly improved the performance of downstream analysis [78]. While this example is from genomics, the methodological principle translates directly to groundwater chemistry, where accurate detection of differentially influenced samples (e.g., polluted vs. background) is equally critical.

The following diagram illustrates the logical relationship between data quality issues and the analytical solutions provided by rPCA, framing it within a generalized environmental data analysis workflow.

Robust PCA is not merely a statistical refinement but an essential tool for ensuring data integrity in groundwater chemistry research. By explicitly modeling and separating outliers from the dominant low-rank structure of hydrogeochemical datasets, rPCA mitigates the vulnerability of conventional multivariate methods to data quality issues. The provided protocols for outlier detection and the iterative algorithm offer researchers a practical pathway to implement rPCA, fostering more accurate and reliable identification of contaminant sources and natural hydrogeochemical processes. Integrating rPCA as a standard pre-processing step in the analytical workflow significantly strengthens the foundation upon which critical water resource management decisions are based.

Ensuring Accuracy: Validating PCA Findings and Comparing Methodological Efficacy

In groundwater chemistry research, particularly within a thesis framework utilizing Principal Component Analysis (PCA) for source apportionment, initial data exploration and validation are paramount. Piper and Schoeller diagrams serve as essential ground-truthing tools, providing an intuitive visual representation of hydrochemical facies and processes before applying multivariate statistical methods [82]. These diagrams help researchers identify natural water types, mixing processes, and potential anthropogenic influences, thereby informing the interpretation of PCA factors which condense numerous variables into key components explaining variance in water chemistry [21] [83] [10]. For instance, PCA might reduce 12 groundwater parameters to 4 significant components explaining 68% of data variance [83], with these components gaining meaningful context when referenced against Piper and Schoeller classifications.

Hydrochemical Data Fundamentals

Core Parameters and Units

For meaningful Piper and Schoeller diagrams, groundwater samples must be analyzed for major ions, with concentrations converted to milliequivalents per liter (meq/L) for plotting [82]. The table below summarizes the essential parameters.

Table 1: Essential Hydrochemical Parameters for Diagram Construction

Parameter	Symbol	Typical Units	Notes
Calcium	Ca²⁺	mg/L, meq/L	Key cation for hardness and geochemical processes
Magnesium	Mg²⁺	mg/L, meq/L	Key cation for hardness
Sodium	Na⁺	mg/L, meq/L	Often linked to anthropogenic sources or saltwater intrusion [84]
Potassium	K⁺	mg/L, meq/L
Chloride	Cl⁻	mg/L, meq/L	Conservative ion, tracer for pollution [20]
Sulfate	SO₄²⁻	mg/L, meq/L	Can be from gypsum dissolution or anthropogenic sources [20]
Bicarbonate	HCO₃⁻	mg/L, meq/L	Derived from carbonate weathering
Carbonate	CO₃²⁻	mg/L, meq/L	Significant in high-pH waters
Nitrate	NO₃⁻	mg/L, meq/L	Key indicator of agricultural pollution [21] [20]
Electrical Conductivity	EC	μS/cm	Measure of total mineralization [21]
pH	pH	-	Controls solubility of minerals
Total Dissolved Solids	TDS	mg/L	Total mineralization

Data Quality Pre-processing

Prior to analysis and plotting, data must undergo rigorous quality checks.

• Ionic Balance: Ensure the sum of cations (meq/L) ≈ sum of anions (meq/L). An error of ±5% is generally acceptable. Calculation: Error (%) = [ (Σcations - Σanions) / (Σcations + Σanions) ] * 100.
• Detection Limits: Replace values below the detection limit with a consistent value (e.g., DL/√2) to avoid statistical bias.
• Unit Consistency: Verify all units are consistent before calculating meq/L. Conversion to meq/L: Concentration (meq/L) = Concentration (mg/L) / Equivalent Weight, where Equivalent Weight = Molecular Weight / Ionic Charge.

Application Notes & Experimental Protocols

Protocol 1: Constructing and Interpreting a Piper Diagram

Objective: To classify water types and identify dominant hydrochemical processes.

Procedure:

• Data Preparation: Convert concentrations of Ca²⁺, Mg²⁺, Na⁺, K⁺, Cl⁻, SO₄²⁻, HCO₃⁻, and CO₃²⁻ from mg/L to meq/L.
• Calculate Percentages:
  • Cation Percentage: Calculate the relative percentage of (Ca²⁺ + Mg²⁺) and (Na⁺ + K⁺) from the total cations.
  • Anion Percentage: Calculate the relative percentage of (HCO₃⁻ + CO₃²⁻) and (Cl⁻ + SO₄²⁻) from the total anions.
• Plotting:
  • The Piper diagram consists of three distinct regions: two triangular fields and one diamond-shaped field.
  • Plot the cation percentages on the lower-left ternary diagram.
  • Plot the anion percentages on the lower-right ternary diagram.
  • The corresponding points are projected vertically and horizontally, respectively, to the central diamond-shaped field, where the combined water type is displayed.

Interpretation Guide:

• Water Type Classification: The diamond field reveals the overall hydrochemical facies (e.g., Ca-HCO₃, Na-Cl, Mg-SO₄) [85] [81]. For example, a study in the Dawen River basin identified Cl·SO₄·Ca as the predominant water type [81].
• Process Identification:
  • Ca-HCO₃ Type: Typically indicates recharge water interacting with carbonate rocks (calcite, dolomite) [85].
  • Na-Cl Type: Suggests seawater intrusion, salt dissolution, or anthropogenic pollution [84].
  • Evolution Trajectories: Points trending from Ca-HCO₃ towards Na-Cl or other types can indicate salinization, ion exchange, or mixing with wastewater [81].
• Integration with PCA: The water types and groups identified on the Piper diagram should correspond to clusters of samples in the PCA score plot. A PCA factor with high loadings on Na⁺ and Cl⁻ should be represented by samples plotting in the Na-Cl region of the Piper diagram, providing a hydrochemical ground-truth for the statistical model [10].

Protocol 2: Constructing and Interpreting a Schoeller Diagram

Objective: To visualize and compare absolute ion concentrations across multiple samples and identify dilution, concentration, or hydrochemical affinity.

Procedure:

• Data Preparation: Use the meq/L concentrations of the major ions. Using a logarithmic scale is standard as it allows comparison of a wide concentration range [82].
• Axis Setup: Create a parallel vertical axis for each ion. The order is typically: Ca²⁺, Mg²⁺, Na⁺, K⁺, HCO₃⁻, Cl⁻, SO₄²⁻ (or similar). The concentration scale is logarithmic.
• Plotting: For each water sample, plot the concentration of each ion as a point on its respective axis. Connect all points for a single sample with straight lines, creating a unique "signature" or fingerprint for that sample.

Interpretation Guide:

• Signature Shape Analysis:
  • Parallel Signatures: Indicate a common hydrochemical origin or genesis, with vertical displacement representing a simple dilution or concentration effect [82].
  • Divergent Signatures: Highlight different hydrochemical processes affecting the samples (e.g., different sources of pollution, distinct water-rock interactions) [85].
• Standardized Schoeller Diagrams: For advanced comparison, a single sample (e.g., a background water or a suspected end-member) can be selected as a pivot. The logarithmic axes are then shifted so that this sample's signature becomes a horizontal line. All other samples are plotted relative to this standard, making deviations exceptionally clear [82]. This process can be efficiently implemented using specialized MATLAB tools [82].
• Process Identification: The diagram can reveal ion exchange (e.g., a deficit of Ca²⁺ relative to a standard), pollution from agricultural nitrate, or the influence of geothermal waters [20].

The following workflow illustrates the integrated use of these diagrams within a PCA-based groundwater study.

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 2: Key Reagents, Software, and Analytical Tools for Hydrochemical Studies

Item Name	Function/Application	Technical Notes
Ion Chromatography (IC) System	Quantification of major anions (Cl⁻, NO₃⁻, SO₄²⁻) and cations (Na⁺, K⁺, Ca²⁺, Mg²⁺).	Provides high-precision data essential for reliable diagrams and PCA input.
Inductively Coupled Plasma Spectrometer	Measurement of major and trace metal cations.	An alternative to IC for cation analysis.
Total Alkalinity Kit	Gran titration for accurate HCO₃⁻ and CO₃²⁻ measurement.	Critical parameter; field titration is often preferred to avoid sample degradation.
MATLAB with Custom Scripts	Advanced data processing and generation of specialized plots like standardized Schoeller diagrams [82].	Offers high flexibility for customizing diagrams and performing multivariate statistics.
R or Python (with pandas, sklearn)	Open-source platforms for performing PCA, data normalization, and generating basic plots.	Widely used for statistical analysis in research.
Gibbs Diagram	Supplementary plot to distinguish dominance of precipitation, rock weathering, or evaporation-crystallization processes.	Used alongside Piper and Schoeller for a comprehensive view.
Geochemist's Workbench	Commercial software suite for creating various hydrochemical diagrams and modeling.	User-friendly option for standard diagram creation.
Self-Organizing Map (SOM)	Unsupervised machine learning for pattern recognition and clustering of water samples [81].	Used to qualitatively classify groundwater groups before quantitative source apportionment with PMF/PCA.

Data Presentation and Synthesis

The following table synthesizes quantitative findings from case studies that successfully applied these tools, demonstrating the link between diagram interpretation and identifiable sources.

Table 3: Synthesis of Hydrochemical Findings from Case Studies Using Piper and Schoeller Diagrams

Study Region	Identified Water Type/Facies (Piper)	Key Signature Features (Schoeller)	Inferred Hydrochemical Processes & Sources (Linked to PCA Factors)
Gaoqiao Diluvial Fan, China [85]	HCO₃-Ca·Mg and HCO₃·SO₄-Ca·Mg	Spatial evolution of signatures from top to edge of fan.	Water-rock interaction (carbonate weathering) is primary control. Anthropogenic F⁻ pollution from historical industry locally overprints natural signal.
Dawen River Basin, China [81]	Cl·SO₄·Ca	N/A (Study used SOM clustering).	PCA/PMF quantified sources: Natural geology (29.0%), Agricultural activities (26.8%), Water-rock interaction (23.9%), Mining (13.6%), Domestic wastewater (6.7%).
Eloued, Algeria [21]	N/A (High mineralization)	High mineralization indicated by elevated points across all ions.	PCA identified mineralization from geological weathering and anthropogenic inputs (associated with NO₂⁻), and nitrification processes (linked to temperature and NO₃⁻).
Southern Tunisia [20]	N/A	Signatures with high Ra and NO₃.	PCA distinguished contamination sources: phosphate mining (radioactivity), agricultural runoff (nitrates), and fossil geothermal waters.
Abomey-Calavi, Benin [84]	Na-K-Cl (79.22% of samples)	Signatures showing Na⁺+K⁺ dominance over Ca²⁺+Mg²⁺.	Cation exchange (clays capturing Ca²⁺/Mg²⁺, releasing Na⁺/K⁺) is the dominant process. Seawater intrusion was confirmed in 4.54% of coastal samples.

Source apportionment models are critical tools in environmental chemistry, enabling researchers to identify and quantify the contributions of various pollution sources to a given sample. In groundwater chemistry research, understanding these sources is paramount for developing effective remediation strategies and policies. This analysis focuses on three prominent models: Principal Component Analysis (PCA), Positive Matrix Factorization (PMF), and MixSIAR. Each employs distinct mathematical frameworks to solve the common challenge of source attribution, with varying requirements for data input, underlying assumptions, and interpretive approaches [86]. PCA serves as a dimensionality-reduction technique, PMF as a robust receptor model that handles measurement uncertainty, and MixSIAR as a Bayesian framework ideal for stable isotope and other biotracer data. Their application within groundwater systems must account for the complex interplay of hydro-geochemical processes, anthropogenic activities, and natural hydrogeological heterogeneity [3] [7].

Theoretical Foundations and Model Comparison

The core mathematical principles and structures of PCA, PMF, and MixSIAR dictate their respective applications and performance in source apportionment studies. PCA is a multivariate statistical method that reduces the dimensionality of a dataset by transforming the original variables into a new set of uncorrelated variables, the principal components (PCs). These PCs are linear combinations of the original variables and are ordered so that the first few retain most of the variation present in the original dataset [8]. The model does not require prior knowledge of the number or composition of sources. In contrast, PMF is a receptor model that decomposes a sample data matrix (X) into two matrices: the factor contribution matrix (G) and the factor profile matrix (F), such that (X = GF + E), where (E) is the residual matrix. A key advantage of PMF is that it incorporates uncertainty estimates for each data point and applies non-negativity constraints, which provide physically realistic solutions [86]. MixSIAR represents a different paradigm as a Bayesian mixing model. It uses Markov Chain Monte Carlo (MCMC) methods to estimate probability distributions of source contributions, formally expressed as (p(\theta | y) \propto p(y | \theta) p(\theta)), where (p(\theta | y)) is the posterior distribution of source proportions ((\theta)) given the tracer data ((y)), (p(y | \theta)) is the likelihood, and (p(\theta)) is the prior distribution [87]. This framework naturally quantifies uncertainty and allows for the incorporation of prior knowledge.

Table 1: Comparative Summary of Source Apportionment Models

Feature	PCA	PMF	MixSIAR
Core Principle	Dimensionality reduction via eigenvector decomposition [8]	Least-squares factor analysis with constraints [86]	Bayesian statistical framework [87]
Data Input	Concentration data only	Concentration data and associated uncertainties [86]	Tracer data (e.g., isotopes, fatty acids), source data, prior information [87]
Key Assumptions	Linearity; sources are orthogonal (uncorrelated)	Constant source profiles; linear combinations; non-negative contributions [88]	Tracer values of sources are distinct and known; mixing is linear [87]
Handling of Uncertainty	Does not explicitly incorporate data uncertainty	Explicitly models measurement uncertainty [86]	Fully probabilistic; provides posterior distributions for all parameters [87]
Primary Output	Factor loadings and scores; qualitative source identification [4]	Quantitative source profiles and contributions [7]	Probability distributions of source proportions [87]
Typical Application in Environmental Science	Initial exploratory analysis; identifying correlated variables and potential sources [4] [89]	Quantifying contributions of specific pollution sources (e.g., industrial, agricultural) [3] [90]	Tracing nutrient flows in food webs; quantifying dietary proportions [87]

Experimental Protocols for Groundwater Source Apportionment

Protocol for PCA-based Apportionment in Groundwater

Implementing PCA for groundwater source identification involves a structured workflow. First, sample collection and analysis must be designed to capture spatial and temporal variability. A minimum of 26-94 groundwater samples is typical, collecting data for major ions (Ca²⁺, Mg²⁺, Na⁺, K⁺, Cl⁻, SO₄²⁻, HCO₃⁻), nutrients (NO₃⁻, NH₄⁺), and heavy metals (e.g., Fe, Mn, As) [3] [4] [7]. On-site parameters like pH, EC, and DO should be measured in situ. Second, data pre-processing is critical. The dataset should be checked for completeness, and missing values must be addressed, often via imputation or removal. Data is typically log-transformed if not normally distributed and standardized (e.g., z-scores) to give all variables equal weight [3]. Third, model execution involves conducting the PCA, often using software like R or SPSS. The number of significant principal components to retain is determined based on Kaiser’s criterion (eigenvalue >1) and scree plots [4] [89]. Finally, interpretation involves analyzing the factor loadings—the correlations between original variables and principal components. A high loading (e.g., > |0.6|) indicates a strong association. For example, a component with high loadings on NH₄⁺, NO₂⁻, and NO₃⁻ might be interpreted as an "agricultural" source [4]. The factor scores can then be mapped geographically using GIS to identify spatial patterns of contamination [4].

Protocol for PMF-based Apportionment in Groundwater

The US EPA PMF software is commonly used for this analysis. The initial step of data preparation is more rigorous than for PCA because PMF requires uncertainty estimates for each data point. The uncertainty ((u{ij})) for a concentration value ((x{ij})) is often calculated using the method detection limit (MDL) and an error fraction, e.g., (u{ij} = \sqrt{(0.1 \times x{ij})^2 + MDL^2}) for values above the MDL [86]. The subsequent model setup and run involves defining the number of factors and selecting model parameters. Unlike PCA, the number of sources is not determined by an eigenvalue rule but through an iterative process. The model is run multiple times with different numbers of factors, and the most physically meaningful solution is selected based on the examination of residual plots (Q robust vs. Q true) and the interpretability of the source profiles [7] [86]. The "Fpeak" parameter can be adjusted to reduce rotational ambiguity and help separate collinear sources [91]. The final and most crucial step is factor interpretation. The resolved source profile for each factor is examined—this shows the chemical composition of that source. For instance, a profile enriched in Cu, Zn, As, and Cr may be identified as "mining-related activities," while one with high Fe and Mn may be attributed to "natural geological processes" [89]. The model outputs the mass contribution of each identified source to every sample.

Protocol for MixSIAR-based Apportionment

MixSIAR is implemented in the R programming environment and is ideal for tracer data like stable isotopes. The protocol begins with formulating the model structure. The user must define the mixture data (the tracer signatures from the groundwater samples), the source data (the tracer signatures of the potential end-member sources), and optionally, discrimination factors (trophic enrichment factors) if applicable. In groundwater studies, sources could include "precipitation," "agricultural runoff," and "septic effluent," characterized by their δ¹⁵N-NO₃ and δ¹⁸O-NO₃ signatures [87]. The next step is specifying the model type. MixSIAR allows for the inclusion of "fixed," "random," or "continuous" effects. A "random" effect could be the sampling location (e.g., well ID) to account for spatial grouping, while a "continuous" effect could be the nitrate concentration itself [87]. The user must also select an error term structure ("Residual * Process" is often recommended). After running the model, which uses JAGS to perform MCMC sampling, the results must be diagnosed and interpreted. Key diagnostics include checking for MCMC chain convergence (using the Gelman-Rubin diagnostic, where values < ~1.05 indicate convergence) and examining the posterior density plots. The primary output is the posterior distribution of source proportions, which is typically summarized by its median and 95% credible interval, providing a robust estimate of contribution and its associated uncertainty [87].

Diagram Title: Source Apportionment Model Workflow

The Scientist's Toolkit: Key Reagents and Materials

Table 2: Essential Research Reagents and Materials for Groundwater Source Apportionment Studies

Item Name	Function/Application	Example from Literature
HCl-HNO₃-HF-HClO₄ Acid Mixture	Digestion of solid samples (e.g., fugitive dust) for heavy metal analysis prior to instrumentation.	Used to digest fugitive dust samples for analysis of As, Cd, Co, Cr, Cu, Mn, Ni, Pb, Zn, Fe, and Sn [89].
National Standard Soil Sample (GSS-23)	Quality control material used to validate analytical accuracy and calculate recovery rates for heavy metal analysis.	Analyzed alongside fugitive dust samples; recovery rates for all elements were between 91.2% and 108.2% [89].
Multiparameter Water Quality Analyzer	On-site measurement of critical physicochemical parameters (pH, EC, DO, Eh, Temperature) during groundwater sampling.	A HANNA HI9828 instrument was used to stabilize and record in-situ parameters in the Qujiang River Basin study [7].
Atomic Fluorescence Spectrophotometer (AFS)	Quantification of specific elements, particularly hydride-forming metals like As, in digested or aqueous samples.	Used in conjunction with an ICP-MS to measure heavy metal concentrations in fugitive dust samples [89].
Inductively Coupled Plasma Mass Spectrometer (ICP-MS)	Highly sensitive detection and quantification of a wide range of trace metals and elements in water and digested samples.	A Perkin-Elmer Elan 9000 was used to measure the concentrations of 11 heavy metals in fugitive dust [89].
MARGA (Monitoring Aerosols and Gases in Air)	Online analysis of major water-soluble ions (e.g., sulfate, nitrate, ammonium) in particulate matter or other matrices.	Used for hourly measurement of major ions (sulfate, nitrate, ammonium) in a PM2.5 source apportionment study [88].

Comparative Analysis and Model Selection

Direct comparisons reveal distinct performance characteristics among these models. A study on stormwater runoff found that while PCA-MLR identified three pollution sources, PMF resolved five, providing a more detailed source mechanism, including two additional sources [86]. Statistically, the PMF model demonstrated superior performance with a Nash coefficient of 0.86–0.99 and a lower percentage error compared to PCA-MLR [86]. This highlights PMF's enhanced ability to deconvolve complex source mixtures. Furthermore, PMF's integration of data uncertainty makes it more robust against outliers and measurement errors. However, a significant limitation of traditional PMF is its assumption of constant source profiles, which can be violated over long time series or for sporadic sources like fireworks or biomass burning [88]. To address this, advanced "rolling PMF" techniques have been developed, which allow source profiles to evolve over time by running PMF on a moving time window [88].

The choice between models is often dictated by the research question and data type. PCA serves as an excellent exploratory tool for initial data assessment and hypothesis generation, requiring only concentration data. It is particularly useful for identifying the main factors controlling hydrochemical variability, such as distinguishing between water-rock interaction and anthropogenic inputs [3] [4]. PMF is the preferred model when the goal is to obtain quantitative mass contributions from specific, relatively constant pollution sources (e.g., industrial discharge, agricultural fertilizer) and when robust uncertainty data is available [3] [7] [86]. MixSIAR is uniquely suited for studies utilizing isotopic or other biotracer data (e.g., δ¹⁵N and δ¹⁸O of nitrate) to trace the flow of specific elements or compounds through a system, and when the explicit quantification of uncertainty via probability distributions is required [87].

A powerful trend in modern environmental research is the coupling of these models to leverage their respective strengths. A common and robust framework involves using PCA for initial, qualitative source identification, which then informs the setup and factor number selection for a subsequent PMF analysis to achieve quantitative apportionment [89] [7]. This PCA-PMF combination has been successfully applied to identify and quantify heavy metal sources in fugitive dust [89]. For a comprehensive assessment, this quantitative result can be further validated with spatial analysis, such as the Mantel test, to correlate source contributions with land use patterns, creating a powerful PCA-PMF-Mantel framework [7]. This integrated approach provides a complete process from qualitative identification to quantitative apportionment and spatial validation, significantly improving the accuracy and interpretability of groundwater pollution source identification.

In groundwater chemistry research, accurately identifying pollution sources and their spatial distribution is paramount for developing effective remediation strategies. Principal Component Analysis (PCA) serves as a powerful dimensionality reduction tool, transforming complex hydrochemical datasets into principal components (PCs) that represent dominant variance patterns often associated with specific contamination sources or natural processes [20]. However, PCA alone provides limited spatial context and requires integration with complementary techniques for comprehensive environmental forensics.

The synergy of PCA with Hierarchical Cluster Analysis (HCA) and Geographic Information Systems (GIS) creates a robust framework that bridges statistical pattern recognition with spatial validation. This integrated approach allows researchers to not only identify hydrochemical facies and contamination sources but also visualize their spatial distribution and validate patterns against landscape features, anthropogenic activities, and hydrological boundaries [81] [92]. This protocol details the application of this integrated methodology specifically for groundwater chemistry source identification, providing step-by-step procedures from data collection to spatial validation.

Methodological Comparison and Selection Framework

Table 1: Comparison of Multivariate Techniques for Groundwater Source Identification

Method	Primary Function	Key Advantages	Limitations	Typical Applications in Groundwater Studies
PCA	Data dimensionality reduction	Identifies dominant variance patterns; Reveals correlated parameters; Reduces data complexity without significant information loss [20]	Linear assumptions; No spatial explicit output; Requires complementary techniques for source quantification [81]	Initial data exploration; Identifying major contamination sources; Parameter correlation analysis [93]
HCA	Sample classification based on similarity	Groups samples with similar characteristics; Identifies hydrochemical facies; No prior assumptions about group numbers needed	Results dependent on distance metrics and linkage methods; Sensitive to outliers; Computationally intensive for large datasets	Classification of groundwater samples; Identifying natural and anthropogenic influenced zones [81]
K-means	Partitioning clustering	Computationally efficient; Works well with large datasets; Creates distinct non-overlapping clusters	Requires pre-specification of cluster number (k); Sensitive to initial centroid selection; Struggles with non-spherical clusters [94]	Regional groundwater quality zoning; Aquifer characterization [81]
DBSCAN	Density-based spatial clustering	Identifies clusters of arbitrary shapes; Robust to outliers; Does not require pre-specified cluster number [94]	Sensitive to parameter selection (eps, minPts); Struggles with varying densities; Performance depends on data structure [94]	Contamination hotspot detection; Anomaly detection in groundwater quality [94]
SOM	Nonlinear dimensionality reduction and clustering	Preserves topological properties; Handles nonlinear relationships; Visualizes high-dimensional data [81]	Complex implementation; Training parameters affect results; Black box nature limits interpretability	Complex groundwater system characterization; Pattern recognition in multivariate hydrochemical data [81]

Integrated Workflow: PCA-Clustering-GIS

The following diagram illustrates the comprehensive workflow for integrating PCA, clustering techniques, and GIS in groundwater chemistry studies:

Research Reagent Solutions and Essential Materials

Table 2: Essential Research Materials and Analytical Requirements for Integrated Groundwater Analysis

Category	Specific Items/Techniques	Technical Specifications	Application Context
Field Sampling Equipment	GPS receiver or smartphone with GPS Essentials application	Sub-meter accuracy preferred for spatial mapping [92]	Precise geolocation of sampling points for GIS integration
	Water sampling bottles	HDPE, pre-cleaned, acid-washed for trace metal analysis	Preventing sample contamination during collection and transport
	Portable field meters	pH, EC, TDS, DO, temperature measurement capability	Real-time determination of unstable parameters [10]
Laboratory Analytical Requirements	Ion chromatography	Anions (Cl⁻, NO₃⁻, SO₄²⁻, F⁻), cations (Na⁺, K⁺, Ca²⁺, Mg²⁺)	Major ion analysis for hydrochemical facies identification [10]
	ICP-MS/OES	Trace metals (As, Pb, Cd, Hg, Fe, Mn) detection at ppb levels	Toxic element quantification [92]
	Stable isotope ratio mass spectrometry	δ¹⁵N-NO₃, δ¹⁸O-NO₃, δ¹³C, δ²H, δ¹⁸O	Pollution source tracking and hydrological process identification [45]
Statistical Analysis Software	R Statistical Software	FactoMineR, cluster, ggplot2 packages	PCA, HCA, and advanced statistical modeling
	Python	Scikit-learn, Pandas, NumPy, Matplotlib	Machine learning implementations and custom analysis scripts
	PAST, SPSS	User-friendly statistical interfaces	Accessible multivariate analysis for non-programmers
GIS Platforms	ArcGIS	Spatial Analyst, Geostatistical Analyst extensions	Advanced spatial interpolation and map algebra operations
	QGIS	Free, open-source with GRASS, SAGA integration	Cost-effective spatial analysis and visualization [92]
	Google Earth Engine	Cloud-based processing of satellite imagery	Land use change analysis and large-scale pattern recognition

Detailed Experimental Protocols

Groundwater Sampling and Data Collection Protocol

• Site Selection Strategy: Employ stratified random sampling based on hydrogeological units, land use types, and proximity to potential contamination sources. Include samples from varying aquifer depths (unconfined and confined) where possible, as demonstrated in a study showing different nitrate sources between unconfined (52.5% chemical fertilizers) and confined (53.9% manure & sewage) aquifers [45].

• Sample Collection and Preservation: Collect samples after purging wells until pH, EC, and temperature stabilize (typically 3-5 well volumes). Filter samples through 0.45μm membranes for cation analysis and acidify to pH <2 with ultrapure HNO₃. Keep samples chilled at 4°C during transport and storage.

• Spatial Data Compilation: Collect complementary spatial datasets including land use/cover maps, geological maps, soil maps, drainage networks, aquifer boundaries, and locations of potential contamination sources (industrial sites, agricultural areas, wastewater treatment plants) [81].

PCA Implementation Protocol

• Data Preprocessing: Standardize the dataset using z-score transformation to normalize variables with different units and magnitudes, calculated as (value - mean)/standard deviation [95].

• PCA Execution: Apply PCA to the correlation matrix of major ions (Ca²⁺, Mg²⁺, Na⁺, K⁺, HCO₃⁻, Cl⁻, SO₄²⁻, NO₃⁻) and physical parameters (pH, EC, TDS). Determine the number of significant PCs using Kaiser criterion (eigenvalue >1) and scree plot analysis [20].

• Component Interpretation: Interpret PCs based on factor loadings, considering |loading| >0.5 as significant. For example, in the Dawen River Basin study, PCA revealed five distinct factors representing natural geology (29.0%), agricultural activities (26.8%), water-rock interactions (23.9%), mining operations (13.6%), and domestic wastewater (6.7%) [81].

Clustering Analysis Protocol

• Data Preparation for Clustering: Use PC scores from significant components as input variables to reduce dimensionality and minimize multicollinearity effects.

• HCA Implementation: Apply Ward's method with squared Euclidean distance as the similarity measure to create dendrograms showing hierarchical relationships between sampling sites. Determine the optimal number of clusters using the elbow method or silhouette analysis [81].

• Alternative Clustering Approaches: For spatial clustering, implement DBSCAN with parameters (eps=0.05, minPts=3) optimized using Silhouette Score and Davies-Bouldin Index, which has shown effectiveness in identifying groundwater quality clusters and contamination hotspots in arid environments [94].

GIS Integration and Spatial Validation Protocol

• Database Development: Create a geodatabase incorporating sampling locations with associated hydrochemical data, PCA results, cluster assignments, and environmental layers.

• Spatial Interpolation: Apply Inverse Distance Weighting (IDW) or kriging to create continuous surfaces of PC scores and cluster assignments. In the Abbottabad, Pakistan study, IDW effectively visualized exceedance levels of As, Pb, Cd, CFU, and Hg, identifying high-risk zones [92].

• Spatial Pattern Analysis: Overlay interpolated surfaces with land use, geological, and hydrological layers to validate statistical patterns. Calculate spatial statistics (Moran's I) to quantify spatial autocorrelation of identified contamination patterns.

Quantitative Results and Data Interpretation

Table 3: Representative Results from Integrated PCA-Clustering-GIS Studies

Study Area	PCA Results (Major Sources Identified)	Clustering Results	Spatial Validation Findings	Management Implications
Dawen River Basin, China [81]	Five factors: Natural geology (29.0%), Agricultural activities (26.8%), Water-rock interactions (23.9%), Mining operations (13.6%), Domestic wastewater (6.7%)	SOM identified five clusters: (i) Natural geological processes, (ii) Agricultural activities, (iii) Hydrogeochemical evolution, (iv) Mining operations, (v) Domestic wastewater discharge	Clusters showed distinct spatial patterns aligned with land use: agricultural cluster along riverbanks, mining cluster in industrial zones	Targeted management: agricultural best practices in eastern regions, industrial controls in western areas
Abbottabad, Pakistan [92]	Five PCs (76% cumulative variance): PC-1: Microbial health risks (CFU, Hg, Cd); PC-2: Natural pollution; PC-3: Arsenic health risk; PC-4 & PC-5: Natural processes	Spatial clustering revealed exceedance Levels 3-5 for As, Pb, Cd, CFU, and Hg across both union councils	GIS modeling identified uniform contamination distribution suggesting widespread poor waste management practices	Urgent need for improved solid waste management and water treatment infrastructure
Al-Qatif, Saudi Arabia [94]	Kernel PCA with polynomial kernel identified salinity and seawater intrusion as primary factors	DBSCAN effectively detected contamination hotspots and spatial anomalies in groundwater quality	Higher salinity clusters in heavily urbanized and agricultural areas influenced by seawater intrusion and over-extraction	Need for managed aquifer recharge and sustainable extraction policies in coastal areas
Zhengzhou Section, Yellow River [95]	APCS-MLR identified mineral dissolution, human activities, and Yellow River recharge as controlling factors	Spatial analysis showed evolution from HCO₃-Na·Ca·Mg type near river to Cl·SO₄·HCO³ type further away	Groundwater near riverbanks and ponds showed strongest human activity impact compared to other regions	Protection measures for riverbank filtration zones and regulation of agricultural practices

Advanced Integration Techniques

Kernel PCA for Non-Linear Data Structures

For complex groundwater systems with non-linear relationships, implement Kernel PCA using polynomial kernels which have demonstrated superior performance in preserving variance and achieving effective dimensionality reduction compared to linear, RBF, sigmoid, and cosine kernels [94]. This approach is particularly valuable in coastal aquifers with complex seawater-freshwater interfaces.

APCS-MLR for Quantitative Source Apportionment

Combine PCA with Absolute Principal Component Score-Multiple Linear Regression (APCS-MLR) for quantitative source apportionment. This method uses PCA for dimensionality reduction, calculates absolute factor scores, then applies multiple linear regression to quantify the contribution of each identified source to individual water quality parameters [95].

Spatial Objective Weighting with PCA

For spatial decision-making, develop PCA-based objective weighting approaches within Multi-Criteria Decision Making (MCDM) frameworks. This method uses PCA to determine criterion weights based on variance contribution rather than subjective expert opinion, creating more data-driven susceptibility assessments [96].

Principal Component Analysis (PCA) is a powerful multivariate statistical tool widely employed in hydrogeochemical studies to identify the underlying processes and sources influencing groundwater chemistry. Its effectiveness, however, is significantly controlled by the hydrogeological setting of the aquifer, particularly whether it is confined or unconfined. This application note provides a detailed comparison of PCA performance across contrasting aquifer types, synthesizing findings from recent international case studies. The content is framed within a broader thesis on PCA applications for groundwater source research, offering standardized protocols and data interpretation frameworks for researchers and environmental scientists.

Unconfined aquifers are characterized by a permeable water table open to the atmosphere, allowing direct vertical recharge and greater susceptibility to surface contamination. In contrast, confined aquifers are bounded above and below by impermeable layers (aquitards), restricting vertical flow and recharge, often resulting in longer groundwater residence times and different evolutionary pathways [97] [98].

Theoretical Background: Aquifer Typology and Hydrochemical Evolution

A clear understanding of aquifer confinement is fundamental to interpreting hydrochemical data and the resulting PCA outputs.

• Unconfined Aquifers: Often referred to as water table aquifers, their upper surface is the water table. They are directly recharged by rainwater percolation and surface water seepage. This direct connection to the surface makes them more vulnerable to contamination from anthropogenic activities and more immediately responsive to climatic conditions such as drought [97]. The water chemistry often reflects recent interactions with the soil zone and surface contaminants.
• Confined Aquifers: These are saturated aquifers overlain by a confining layer of low-permeability material. Water within them is often under pressure, and recharge occurs slowly in distant outcrop areas. Their confined nature provides a degree of protection from surface contamination, and their chemistry is predominantly governed by prolonged water-rock interactions along extended flow paths [97] [98]. The potentiometric surface, rather than the water table, defines the head distribution.

These fundamental differences in recharge mechanism, vulnerability, and flow dynamics lead to distinct hydrochemical signatures, which PCA can help to decode.

Comparative Case Studies

The following case studies from arid and semi-arid regions illustrate how PCA performance and outcomes vary between unconfined and confined aquifer settings.

Unconfined Aquifer Case Study: Eloued, Algeria

A study of the unconfined aquifer in the arid southeastern region of Eloued, Algeria, utilized PCA and Hierarchical Ascending Classification (HAC) on 113 water samples [21].

Table 1: Key Hydrochemical Parameters and PCA Results from the Eloued, Algeria Study

Aspect	Description
Average EC	4748 μS/cm (range: 2678 - 18076 μS/cm), indicating highly mineralized water.
Key Contaminant	Nitrate concentrations in half of the samples exceeded WHO standards.
PCA Variance	Five principal components explained 83% of the total variance across eight variables.
Identified Processes	1. Mineralization from geological weathering, long residence time, and anthropogenic inputs (associated with EC, TDS, NO₂⁻).2. Acid-base equilibrium (pH, NH₄⁺).3. Nitrification processes linked to temperature and NO₃⁻.
Anthropogenic Link	Elevated nitrates were strongly attributed to human activities, including fertilizer use, wastewater discharge, and organic matter decomposition.

Performance Insight: In this unconfined setting, PCA effectively delineated the strong overlapping influence of natural geochemistry and anthropogenic pollution. The high explanatory power of the model (83% variance) successfully separated the effects of geological weathering from contamination drivers like agricultural and domestic waste, highlighting the aquifer's vulnerability.

Confined Aquifer Case Study: Debrecen Area, Hungary

Research on the confined Quaternary aquifer system in the Pannonian Basin of Debrecen, Hungary, tracked groundwater chemistry from 2019 to 2024 using PCA, SOM, and HCA [99].

Table 2: Key Hydrochemical Parameters and PCA Results from the Debrecen, Hungary Study

Aspect	Description
Dominant Water Type	Ca-Mg-HCO₃, with a temporal shift toward Na-HCO₃.
Primary Driver	Increased salinity and hydrochemical evolution were driven by ongoing rock-water interactions.
PCA & HCA Findings	HCA showed a reduction from six clusters (2019) to five (2024), indicating a gradual homogenization of water quality. PCA confirmed this trend was linked to water-rock interactions.
Anthropogenic Influence	PCA identified only limited contributions from anthropogenic activities.
Overall Quality Trend	Groundwater quality generally improved over time, with most areas meeting drinking standards, attributed to the stability of the confined aquifer system.

Performance Insight: For this confined aquifer, PCA excelled at identifying the dominant natural geochemical processes and tracking slow, system-internal evolutionary trends like mineral dissolution and cation exchange. The limited anthropogenic signal in the PCA results underscores the natural protection offered by the confining layers.

Advanced PCA Application: Coastal Aquifers, Saudi Arabia

A study in the complex coastal aquifers of Al-Qatif, Saudi Arabia, which include both shallow (effectively unconfined) and deeper (confined) systems, highlighted a limitation of traditional linear PCA [25]. It struggled to resolve the non-linear relationships caused by interacting processes like seawater intrusion, reverse ion exchange, and variable abstraction. The application of Kernel PCA, a non-linear extension, was found superior for handling this complexity. Using a polynomial kernel, the model effectively preserved variance and categorized wells into "Very Bad," "Bad," and "Medium" quality classes, providing a more robust framework for management in a mixed hydrogeological setting.

Experimental Protocols for PCA in Groundwater Studies

To ensure reproducible and comparable results, the following standardized protocols are recommended.

Pre-PCA Hydrochemical Data Collection Protocol

Objective: To collect representative groundwater samples for physicochemical analysis. Materials:

• Peristaltic Pump or Submersible Pump: To purge wells and collect samples from specific depths.
• Multi-Parameter Water Quality Meter: For in-situ measurement of pH, Temperature, Electrical Conductivity (EC), Dissolved Oxygen (DO), and Redox Potential (Eh).
• Sample Bottles: Pre-cleaned HDPE bottles; some pre-acidified for cation analysis.
• On-site Filtration Kit: With 0.45 μm membrane filters.
• Cooler with Ice Packs: For sample preservation at 4°C during transport.

Procedure:

• Well Purging: Purge the well for 10-15 minutes or until pH, EC, and temperature stabilize [25] [7].
• In-situ Measurement: Calibrate the meter and record pH, Temperature, EC, DO, and Eh immediately after purging.
• Sample Collection:
  • Filter water samples for anion and cation analysis.
  • Collect duplicates and field blanks for quality control.
  • For nitrate analysis, some protocols recommend preservation with boric acid [100].
• Sample Storage: Place samples immediately on ice and transport to an accredited laboratory for analysis of major ions (Ca²⁺, Mg²⁺, Na⁺, K⁺, Cl⁻, SO₄²⁻, HCO₃⁻, NO₃⁻) and other relevant parameters.

Data Pre-processing and PCA Execution Protocol

Objective: To prepare hydrochemical data and perform Principal Component Analysis. Software: R (with FactoMineR, factoextra), Python (with scikit-learn, pandas), or SPSS.

Procedure:

• Data Cleaning: Handle missing values (e.g., by imputation or removal) and screen for obvious errors.
• Data Standardization : This is a critical step. Because hydrochemical parameters have different units and variances, data must be auto-scaled (converted to Z-scores) to have a mean of 0 and a standard deviation of 1. This prevents variables with large magnitudes from dominating the PCA.
• PCA Execution:
  • Input the standardized data matrix.
  • Extract principal components (PCs) and evaluate their eigenvalues. A common criterion is to retain PCs with eigenvalues >1 (Kaiser's criterion).
  • Examine the scree plot to visually assess the proportion of total variance explained by each PC.
• Interpretation:
  • Analyze the loadings of each variable on the retained PCs. High absolute loadings (close to -1 or 1) indicate a strong influence of that variable on the component.
  • Interpret the physicochemical meaning of each PC based on its high-loading variables (e.g., PC1: "Salinity and seawater intrusion"; PC2: "Agricultural nitrate pollution").
  • Use score plots to see how individual samples cluster based on the identified processes.

Kernel PCA Protocol for Non-Linear Data Structures

Objective: To apply Kernel PCA when traditional PCA fails to capture complex, non-linear relationships. Procedure:

• Kernel Selection: Test different kernel functions (e.g., Linear, Polynomial, Radial Basis Function (RBF), Sigmoid). The Al-Qatif study found the Polynomial kernel most effective [25].
• Model Training: Map the original hydrochemical data into a higher-dimensional feature space using the selected kernel function.
• Dimensionality Reduction: Perform linear PCA in the new feature space to extract non-linear principal components.
• Validation: Assess the performance based on the variance retained in the reduced dimensions and the interpretability of the resulting components in a hydrogeological context.

The following workflow diagram summarizes the key steps in conducting a PCA for groundwater studies, from field sampling to final interpretation.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Research Reagent Solutions and Materials for Hydrochemical and PCA Studies

Item Name	Function/Brief Explanation
Multi-Parameter Water Quality Meter	Essential for accurate in-situ measurement of unstable parameters like pH, EC, and DO, which are critical for PCA input data.
0.45 μm Membrane Filters	Used for filtering suspended solids from water samples to ensure analysis represents only dissolved constituents.
Standard Anion/Cation Solutions	Certified reference materials for calibrating laboratory instruments (e.g., Ion Chromatograph, ICP-OES) to ensure analytical accuracy.
Statistical Software (R/Python)	Platforms containing specialized libraries (FactoMineR in R, scikit-learn in Python) for performing robust PCA and related multivariate analyses.
Kernel Functions (Polynomial, RBF)	Used in advanced Kernel PCA to handle non-linear relationships in complex aquifer systems where traditional PCA may fail [25].

This comparison demonstrates that the performance and interpretation of PCA are intrinsically linked to aquifer hydrogeology. In unconfined aquifers, PCA robustly identifies a mix of natural and anthropogenic processes, often revealing a strong contamination signature from surface activities. In confined aquifers, PCA typically elucidates slower, natural geochemical evolution driven by water-rock interactions with a muted anthropogenic signal. For complex systems like coastal aquifers with non-linear relationships, Kernel PCA offers a powerful advanced alternative. Researchers should therefore select and interpret PCA methodologies within the specific physical context of the aquifer system under investigation to accurately unravel the sources and processes governing groundwater chemistry.

Application Context in Groundwater Chemistry Research

In groundwater chemistry studies, Principal Component Analysis (PCA) is a vital statistical tool for simplifying complex hydrochemical datasets. It identifies a smaller set of uncorrelated variables, the principal components (PCs), which capture the majority of the variance in the original data [25] [101]. These components are derived from the eigenvalues and eigenvectors of the data's covariance matrix. The eigenvalues are particularly crucial as they indicate the amount of variance accounted for by each PC, thereby helping researchers decide how many components to retain for interpreting underlying processes such as water-rock interaction, anthropogenic contamination, or seawater intrusion [25].

However, a significant challenge arises because eigenvalues calculated from a single dataset are point estimates. They lack an inherent measure of statistical reliability or uncertainty [101]. In the context of groundwater source research, this means that the apparent importance of a factor (e.g., salinity from seawater intrusion versus nitrate from agricultural runoff) could be misleading if the eigenvalue is unstable. Bootstrap resampling addresses this by providing a robust, data-driven method to quantify the uncertainty of these eigenvalues and construct confidence intervals around them. This process helps validate the stability of the identified components, ensuring that the hydrochemical sources and processes they represent are reliable and not merely artifacts of sampling variability [102] [101].

Core Methodology and Workflow

The bootstrap approach is a powerful resampling technique that allows for the estimation of sampling distributions from the available data alone, without stringent parametric assumptions. Its application to PCA for generating confidence intervals on eigenvalues follows a systematic protocol [101].

Conceptual Workflow

The following diagram illustrates the logical sequence of the bootstrap resampling procedure for PCA eigenvalues.

Detailed Experimental Protocol

Title: Bootstrap Resampling for Confidence Intervals on PCA Eigenvalues in Hydrochemical Datasets

Objective: To quantify the uncertainty and stability of eigenvalues derived from a PCA of groundwater chemistry data by constructing non-parametric bootstrap confidence intervals.

Materials and Reagents: Table 1: Essential Research Reagents and Materials for Hydrochemical Analysis

Item	Function in Context
Groundwater Samples	The primary source material, collected from wells representing the aquifer system of interest.
Standard Reference Materials	Certified water standards used to calibrate analytical instruments and ensure measurement accuracy of ions.
Ion Chromatography System	For accurate quantification of major anions (e.g., Cl⁻, SO₄²⁻, NO₃⁻) and cations (e.g., Na⁺, K⁺, Ca²⁺, Mg²⁺).
Statistical Software	Platforms such as R or Python with specialized libraries (e.g., boot in R, scikit-learn in Python) for performing PCA and bootstrap resampling.

Step-by-Step Procedure:

• Data Preparation and PCA on Original Sample:

  • Compile a dataset of n groundwater samples, each analyzed for p hydrochemical parameters (e.g., pH, EC, Na⁺, Ca²⁺, Cl⁻, HCO₃⁻, NO₃⁻). Address missing values and outliers appropriately [25].
  • Standardize the data (e.g., z-scores) to ensure all parameters contribute equally to the PCA, preventing dominance by variables with larger units or variance.
  • Perform PCA on the original, standardized n x p data matrix to obtain the initial set of eigenvalues, λ₁, λ₂, ..., λp.

• Bootstrap Resampling:

  • Set the number of bootstrap replicates, B (typically B >= 1000).
  • For b = 1 to B: a. Draw a Bootstrap Sample: Randomly select n rows from the original dataset with replacement. This creates a new dataset of the same size but with some original samples possibly duplicated or omitted. b. Perform PCA on the Bootstrap Sample: Standardize this new bootstrap dataset using the means and standard deviations from the original dataset. Then, perform PCA to obtain a new set of bootstrap eigenvalues, λ*₁b, λ*₂b, ..., λ*pb. c. Store the Results: Save the eigenvalues from this bootstrap iteration.

• Construct Confidence Intervals:

  • For the i-th eigenvalue (λi), you now have a distribution of B bootstrap estimates (λ*i1, λ*i2, ..., λ*iB).
  • To construct a 95% percentile confidence interval, find the 2.5th percentile and the 97.5th percentile of this bootstrap distribution for λi [101].
  • The resulting interval [λ*i,(lower), λ*i,(upper)] represents the 95% confidence interval for the true i-th eigenvalue.

Data Presentation and Interpretation

The results of the bootstrap procedure are best summarized in a table that contrasts the original point estimates with their bootstrap-derived uncertainty measures.

Table 2: Example Output of Bootstrap PCA on a Simulated Groundwater Chemistry Dataset (n=66, B=1000)

Principal Component	Original Eigenvalue (λ)	% Variance Explained (Original)	Bootstrap Mean Eigenvalue (λ*)	95% Confidence Interval for λ	Bootstrapped % Variance (Mean)
PC1	4.52	41.1	4.48	[4.15, 4.82]	40.7
PC2	2.18	19.8	2.22	[1.87, 2.59]	20.2
PC3	1.45	13.2	1.43	[1.11, 1.76]	13.0
PC4	0.98	8.9	1.01	[0.72, 1.31]	9.2
PC5	0.62	5.6	0.65	[0.41, 0.90]	5.9

Interpretation Guidelines

• Component Stability: A stable and reliable component is indicated by a narrow confidence interval. For instance, in Table 2, PC1 has a tight interval around a high eigenvalue, confirming its dominance in explaining hydrochemical variance [101].
• Component Selection: The bootstrap results can inform decisions on the number of components to retain. A scree plot can be enhanced by overlaying the confidence intervals. Components whose confidence intervals do not overlap with those of subsequent components are considered significant. A sharp drop in the lower bound of the confidence interval can serve as a more robust cut-off point than visual inspection of a scree plot alone.
• Assessing Overlap: Overlapping confidence intervals between successive eigenvalues (e.g., PC3 and PC4 in the example above) suggest that their order of importance is uncertain. This might indicate that the underlying hydrochemical processes they represent are not perfectly distinct.

Implementation Notes for Groundwater Research

• Software Implementation: This method is readily implementable in R using the boot and prcomp functions, or in Python using scikit-learn and numpy for the PCA, with custom code for the bootstrap loop.
• Data Quality: The bootstrap method estimates the uncertainty due to sampling variability. It cannot account for systematic errors in the data, such as biased well selection or analytical inaccuracies. High-quality, representative groundwater sampling remains paramount [10] [64].
• Integration with Other Techniques: In modern hydrogeology, bootstrap-resampled PCA can be integrated with other machine learning models. For example, it can be used to quantify uncertainty in input features before they are fed into an Artificial Neural Network (ANN) or a Random Forest (RF) model, thereby improving the reliability of final predictions like Water Quality Index (WQI) or contamination source apportionment [102] [103].

Conclusion

Principal Component Analysis stands as an indispensable, powerful tool in the hydrogeologist's arsenal, effectively reducing complex groundwater chemistry data into interpretable patterns of contamination sources. From foundational principles to advanced integration with models like APCS-MLR, PCA provides a robust framework for distinguishing between geogenic processes and anthropogenic impacts such as agricultural runoff, industrial discharge, and seawater intrusion. However, its success hinges on acknowledging and overcoming inherent limitations through careful data pre-processing, validation with complementary methods, and awareness of its assumptions. Future directions point towards increased use of hybrid models that couple PCA with stable isotope analysis and machine learning techniques, offering even more precise quantification of contamination sources and empowering the development of targeted, effective groundwater remediation and management policies globally.

References

• 1. Principal Component Analysis (PCA) Explained Visually ... [https://towardsdatascience.com/principal-component-analysis-pca-explained-visually-with-zero-math-1cbf392b9e7d/]
• 2. Principal Component Analysis (PCA) For Dummies [https://www.billconnelly.net/?p=697]
• 3. Apportionment and evolution of pollution sources in a ... [https://www.sciencedirect.com/science/article/abs/pii/S0169772218301104]
• 4. Groundwater pollution source apportionment using ... [https://pubmed.ncbi.nlm.nih.gov/31463754/]
• 5. 5 PCA Visualizations You Must Try On Your Next Data ... [https://darioradecic.substack.com/p/top-5-pca-visualizations]
• 6. How to read PCA biplots and scree plots | by BioTuring Team [https://bioturing.medium.com/how-to-read-pca-biplots-and-scree-plots-186246aae063]
• 7. Groundwater Pollution Source Identification Based on a ... [https://www.mdpi.com/2073-4441/17/19/2881]
• 8. Contaminant Source Attribution Using PCA [https://www.trihydro.com/news/news-details/environmental-statistics-contaminant-source-attribution-using-pca-pfas-an-example]
• 9. Source apportionment of groundwater pollutants in Apulian ... [https://bmcchem.biomedcentral.com/articles/10.1186/1752-153X-6-S2-S5]
• 10. Assessment of Groundwater Quality through ... [https://www.mdpi.com/2073-4441/16/11/1570]
• 11. A Beginner's Guide to Eigenvectors, Eigenvalues, PCA, ... [https://wiki.pathmind.com/eigenvector]
• 12. Eigenvalues and eigenvectors [https://en.wikipedia.org/wiki/Eigenvalues_and_eigenvectors]
• 13. pca - Making sense of principal component analysis ... [https://stats.stackexchange.com/questions/2691/making-sense-of-principal-component-analysis-eigenvectors-eigenvalues]
• 14. Principal Component Analysis (PCA) [https://www.geeksforgeeks.org/data-analysis/principal-component-analysis-pca/]
• 15. Eigenvalues and eigenvectors [https://www.graphpad.com/guides/prism/latest/statistics/stat_pca_process_eigenvalue_eigenvector.htm]
• 16. How does eigenvalues measure variance along the ... [https://stats.stackexchange.com/questions/346692/how-does-eigenvalues-measure-variance-along-the-principal-components-in-pca]
• 17. Chemical Characterization and Source Analysis of Shallow ... [https://www.mdpi.com/2073-4441/17/13/1959]
• 18. Water quality modelling using principal component ... [https://www.sciencedirect.com/science/article/abs/pii/S0025326X22011754]
• 19. Inconsistency of PCA-based water quality index [https://www.sciencedirect.com/science/article/abs/pii/S0048969722084571]
• 20. Using principal component analysis to distinguish sources ... [https://www.sciencedirect.com/science/article/pii/S0375674225000020]
• 21. Multivariate Analysis of Groundwater Quality Using PCA ... [https://www.sciencedirect.com/science/article/pii/S1944398625004795]
• 22. Understanding Scores and Loadings • LearnPCA [https://bryanhanson.github.io/LearnPCA/articles/Vig_04_Scores_Loadings.html]
• 23. What are principal component scores? - pca [https://stats.stackexchange.com/questions/222/what-are-principal-component-scores]
• 24. Synoptic sampling and principal components analysis to ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5508047/]
• 25. Kernel principal component analysis-based water quality ... [https://www.nature.com/articles/s41598-025-18980-1]
• 26. component analysis and hierarchical cluster analyses of ... [https://www.sciencedirect.com/science/article/abs/pii/S0009281915000045]
• 27. svd - PCA of non-Gaussian data - Cross Validated [https://stats.stackexchange.com/questions/32105/pca-of-non-gaussian-data]
• 28. Multivariate statistics [http://biometry.github.io/APES//Stats/stats50-MultivariateStatistics.html]
• 29. How to perform a principal components analysis (PCA) in SPSS Statistics | Laerd Statistics [https://statistics.laerd.com/spss-tutorials/principal-components-analysis-pca-using-spss-statistics.php]
• 30. Use and Interpret Principal Components Analysis in SPSS - Accredited Professional Statistician For Hire [https://www.scalestatistics.com/principal-components-analysis.html]
• 31. correlation - What are the assumptions of factor analysis? - Cross Validated [https://stats.stackexchange.com/questions/43304/what-are-the-assumptions-of-factor-analysis]
• 32. How to use FactorAssumptions [https://cran.r-project.org/web/packages/FactorAssumptions/vignettes/vignette.html]
• 33. Groundwater quality characterization using an integrated ... [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0294533]
• 34. Investigating Water Quality Data Using Principal ... [https://www.mdpi.com/2073-4441/13/3/343]
• 35. Guidelines for interpreting hydrochemical patterns in data ... [https://www.sciencedirect.com/science/article/abs/pii/S0022169409006386]
• 36. Principal Component Analysis (PCA) | Omics data ... [https://laboratory-of-lipid-metabolism-a.gitbook.io/omics-data-visualization-in-r-and-python/metabolites-and-lipids-multivariate-statistical-analysis-in-r/principal-component-analysis-pca]
• 37. Principal component analysis [https://en.wikipedia.org/wiki/Principal_component_analysis]
• 38. US EPA - Health & Environmental Research Online (HERO) [https://hero.epa.gov/hero/index.cfm/reference/details/reference_id/3066007]
• 39. Principal Component Analysis (PCA): Explained Step-by- ... [https://builtin.com/data-science/step-step-explanation-principal-component-analysis]
• 40. Principal Component Analysis (PCA) from Scratch [https://bagheri365.github.io/blog/Principal-Component-Analysis-from-Scratch/]
• 41. PCA, eigen decomposition and SVD [https://pages.mtu.edu/~shanem/psy5220/daily/Day04/PCA.html]
• 42. Synoptic sampling and principal components analysis to ... [https://pubs.usgs.gov/publication/70188458]
• 43. 11.4 - Interpretation of the Principal Components | STAT 505 [https://online.stat.psu.edu/stat505/lesson/11/11.4]
• 44. How to interpret PCA loadings? - Cross Validated [https://stats.stackexchange.com/questions/92499/how-to-interpret-pca-loadings]
• 45. N source apportionment in groundwater: Insights from PCA ... [https://egusphere.copernicus.org/preprints/2025/egusphere-2025-5482/]
• 46. Professional Paper 1885 [https://pubs.usgs.gov/publication/pp1885]
• 47. Application of principal component analysis in the pollution ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3575243/]
• 48. Geochemical control, and source apportionment [https://www.sciencedirect.com/science/article/abs/pii/S0269749124017159]
• 49. component analysis as a tool to indicate the origin of ... [https://www.sciencedirect.com/science/article/abs/pii/S0016706105000984]
• 50. Groundwater pollution source identification and ... [https://www.sciencedirect.com/science/article/pii/S004896972033905X]
• 51. Groundwater Quality in Agricultural Lands Near a Rapidly ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7917582/]
• 52. Geochemical Factors as a Tool for Distinguishing ... [https://www.mdpi.com/2073-445X/13/4/434]
• 53. Assessment of Groundwater Quality Using APCS-MLR Model [https://www.mdpi.com/2073-4441/15/2/225]
• 54. APCS-MLR model: A convenient and fast method for ... [https://www.sciencedirect.com/science/article/abs/pii/S0301479722006740]
• 55. Groundwater pollution source identification and ... [https://pubmed.ncbi.nlm.nih.gov/32610237/]
• 56. APCS-MLR model: A convenient and fast method for ... [https://pubmed.ncbi.nlm.nih.gov/35472839/]
• 57. Source apportionment of water pollutants in Poyang Lake ... [https://www.frontiersin.org/journals/environmental-science/articles/10.3389/fenvs.2022.924350/full]
• 58. Spatial distribution characteristics and source analysis of ... [https://pubmed.ncbi.nlm.nih.gov/37758147/]
• 59. Using PCA-APCS-MLR model and SIAR model combined ... [https://www.sciencedirect.com/science/article/abs/pii/S0883292722001585]
• 60. A multi-scale orthogonal nonlinear strategy for multi-variate ... [https://www.sciencedirect.com/science/article/abs/pii/S0959152406000199]
• 61. Multivariate statistical analysis of physicochemical parameters of... [https://www.redalyc.org/journal/429/42977076004/]
• 62. Application of statistics to correlate groundwater with land... chemistry [https://pubmed.ncbi.nlm.nih.gov/37036575/]
• 63. Component retention in principal component analysis with ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC1797006/]
• 64. PMC Search Update - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC11928377/]
• 65. A Case Study in the Northern Sichuan Basin [https://www.frontiersin.org/journals/environmental-science/articles/10.3389/fenvs.2022.907872/full]
• 66. component analysis incorporated water quality index ... [https://www.sciencedirect.com/science/article/pii/S2590252024000102]
• 67. Principal Component Analysis (PCA)–Geographic ... [https://www.mdpi.com/2071-1050/14/21/14572]
• 68. Orthogonal versus oblique rotation: A practical application ... [https://medium.com/@kavengik/orthogonal-versus-oblique-rotation-a-practical-application-part-2-376189e52bac]
• 69. Principal Component Analysis to Discriminate and Locate ... [https://www.mdpi.com/2076-3298/12/5/163]
• 70. Groundwater Chemistry Analysis by Multivariate Statistical ... [https://ui.adsabs.harvard.edu/abs/2006AGUFM.H41B0405W/abstract]
• 71. Application of kernel principal component analysis and ... [https://www.nature.com/articles/s41598-018-20121-w]
• 72. A KPCA-ISSA-SVM Hybrid Model for Identifying Sources of ... [https://www.mdpi.com/2073-4441/17/19/2859]
• 73. Improving kernel PCA-based algorithm for fault detection in ... [https://www.sciencedirect.com/science/article/abs/pii/S0957582023008212]
• 74. Source apportionment and migration characteristics of ... [https://www.sciencedirect.com/science/article/abs/pii/S0269749123015865]
• 75. Health Risk Assessment of Groundwater in Cold Regions ... [https://www.mdpi.com/2073-4441/17/20/2984]
• 76. Kernel principal component analysis-based water quality ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12508030/]
• 77. Enhancing Data Fidelity with Robust Principal Component ... [https://medium.com/the-modern-scientist/unveiling-the-true-signal-enhancing-data-fidelity-with-robust-principal-component-analysis-676dbb482196]
• 78. Robust principal component analysis for accurate outlier ... [https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-020-03608-0]
• 79. spss - How to identify outliers and conduct robust ? PCA [https://stats.stackexchange.com/questions/71899/how-to-identify-outliers-and-conduct-robust-pca]
• 80. [2001.05484] Bridging Convex and Nonconvex Optimization in Robust ... [https://arxiv.org/abs/2001.05484]
• 81. Source Analysis of Groundwater Chemical Components in ... [https://www.mdpi.com/2073-4441/17/20/2924]
• 82. Standardized Schoeller diagrams—A Matlab plotting tool [https://link.springer.com/article/10.1007/s00767-023-00556-3]
• 83. Entropy-based groundwater quality evaluation with ... [https://pubmed.ncbi.nlm.nih.gov/40293572/]
• 84. Presentation and use of a new diagram to determine the ... [https://www.sciencedirect.com/science/article/pii/S240584402304896X]
• 85. Hydrochemical Characteristics and Groundwater Quality ... [https://www.mdpi.com/2071-1050/10/12/4507]
• 86. Comparison of two receptor models PCA-MLR and PMF for ... [https://www.sciencedirect.com/science/article/abs/pii/S0048969719302554]
• 87. Bayesian Mixing Models in R • MixSIAR - Brian Stock [https://brianstock.github.io/MixSIAR/]
• 88. Short-term source apportionment of fine particulate matter with ... [https://acp.copernicus.org/articles/24/475/2024/]
• 89. The Predominant Sources of Heavy Metals in Different Types ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9602615/]
• 90. Source apportionment of PM10 particles in the urban ... [https://www.sciencedirect.com/science/article/pii/S0013935125009107]
• 91. A Comparison of Multiple Combined Models for Source ... [https://aaqr.org/articles/aaqr-14-01-oa-0024]
• 92. Principal Component Analysis (PCA)–Geographic ... [https://www.wisdomlib.org/science/journal/sustainability-journal-mdpi/d/doc1815966.html]
• 93. Chemical characteristics of groundwater and surface water ... [https://pubmed.ncbi.nlm.nih.gov/40102623/]
• 94. New perspective on density-based spatial clustering of ... [https://www.sciencedirect.com/science/article/abs/pii/S0022169425009047]
• 95. Chemical characteristics and controlling factors of ... [https://www.frontiersin.org/journals/environmental-science/articles/10.3389/fenvs.2025.1580232/full]
• 96. Leveraging Principal Component Analysis for Data-Driven ... [https://www.mdpi.com/2220-9964/14/5/195]
• 97. Unconfined vs Confined Aquifers [https://mcwec.org/2022/10/confined-unconfined-aquifers/]
• 98. 6.3 Confined Aquifers - GW Books - The Groundwater Project [https://books.gw-project.org/hydrogeologic-properties-of-earth-materials-and-principles-of-groundwater-flow/chapter/confined-aquifers/]
• 99. integration of classical and unsupervised learning methods [https://pubmed.ncbi.nlm.nih.gov/40025328/]
• 100. Evaluating groundwater quality through contaminant ... [https://www.frontiersin.org/journals/environmental-science/articles/10.3389/fenvs.2025.1636843/full]
• 101. Bootstrap Approach for Principal Component Analysis ... [https://www.preprints.org/manuscript/202307.1100]
• 102. Bootstrap approach for quantifying the uncertainty in ... [https://www.sciencedirect.com/science/article/pii/S1658077X23000851]
• 103. Bootstrapping Enhanced Model for Improving Soil Nitrogen ... [https://www.mdpi.com/2504-3129/6/2/23]