Partial least squares discriminant analysis (PLS-DA) for classifying levels of urban metal contamination using analytical chemometrics and pollution indices

Abstract

Supervised machine learning has become increasingly adopted in urban contamination studies, offering higher accuracy than conventional methods in interpreting complex datasets, distinguishing contamination zones, identifying key drivers, and defining geochemical fingerprints. Accordingly, this study explores the applicability and performance of Partial Least Squares Discriminant Analysis (PLS-DA) for classifying heavy metal contamination levels in urban road dust through the ICP-AES geochemical data and pollution indices in a densely populated urban area of Menofiya Governorate (Egypt), as a case study, while also assessing the associated ecological and human health risks. Cd, Cr, Cu, Pb, and Zn exhibited mean enrichment factors above 1.5, indicating predominantly anthropogenic contributions. The overall ecological risk assessment classified the study area as high ecological risk (RI > 600), with an average RI of approximately 745. High carcinogenic risks were identified for children due to Cd and Cr exposure, with CR (via ingestion) and overall TCR values exceeding 1E10-4. The PLS-DA model achieved its lowest prediction error at two components (RM-PRESS = 0.50), explaining 84.74% of the response variance, with nearly all samples falling within Hotelling’s T² limits and showing stable residual distribution. Model evaluation showed normal, unbiased residual patterns, independent errors, and full consistency between actual and predicted classes, confirming high performance. Moreover, Variable Importance in Projection (VIP) analysis demonstrated that Fe (1.74), Mn (1.12), and Cd (1.00) were the main variables (VIP ≥ 1) driving the separation between the pollution load index (PLI) classes.

Bibliographic Information