PCA Analysis
For Water Quality Data

Data Reduction

A monitoring program tracking 25 parameters across 50 locations generates thousands of data points. PCA reduces this to a handful of components that explain the majority of the variation — typically 3 to 6 components capture 70–85% of the total variance. This makes it possible to visualize and interpret datasets that would be overwhelming to analyze parameter by parameter.

Source Identification

PCA groups parameters that vary together into components, and each component often corresponds to a distinct pollution source or environmental process. By examining which parameters load heavily on each component, you can identify the sources influencing your water quality without knowing them in advance.

Spatial and Temporal Pattern Detection

By plotting component scores for each sample, you can identify which monitoring locations are most influenced by each source, how source contributions change seasonally, and whether conditions are improving or worsening over time — all from a single analysis rather than reviewing dozens of individual parameter trends.

Step 1: Data Preparation

Water quality data requires careful preparation before PCA:

• Standardization — Parameters measured in mg/L, µg/L, NTU, and pH units must be standardized (z-score normalization) so that scale differences do not dominate the analysis
• Non-detect handling — Values below detection limits must be substituted (typically half the detection limit) or handled with censored-data methods
• Outlier review — Extreme values can distort PCA results; validated data with outliers flagged and investigated produces more reliable components
• Sample size — As a general rule, you need at least 5–10 samples per parameter for statistically meaningful PCA results

Step 2: Compute the Correlation Matrix

PCA begins by calculating the correlation between every pair of parameters. Parameters that are strongly correlated (e.g., conductivity and total dissolved solids) will load on the same component. Parameters that are uncorrelated or negatively correlated will appear on different components.

Step 3: Extract Principal Components

The analysis extracts components in order of the amount of variance they explain. The first principal component (PC1) captures the largest share of total variance, PC2 captures the next largest share (uncorrelated with PC1), and so on. Each component is defined by its loadings — the weight each original parameter contributes to the component.

Step 4: Determine the Number of Components

Not all components are meaningful. Standard methods for selecting how many to retain include:

• Kaiser criterion — Retain components with eigenvalues greater than 1
• Scree plot — Plot eigenvalues and look for the “elbow” where the curve flattens
• Cumulative variance — Retain enough components to explain 70–80% of total variance

Step 5: Interpret and Apply

Examine the parameter loadings on each retained component to identify what each component represents. Apply varimax rotation to simplify the loading structure and make component interpretation clearer. Use component scores to classify sampling locations, track temporal changes, and support decision-making.