Toxicity Prediction of Landfill Leachate-Contaminated Crops Using Machine Learning Models Based on PAH and Heavy Metal Concentrations

Abstract

The unregulated disposal of municipal solid waste in landfills generates leachate that contaminates the surrounding soil and crops with toxic substances, posing a major threat to food safety and human health. This study evaluated the contamination levels in agricultural fields located near five landfill sites in South India. A total of 600 samples (370 safe, 230 unsafe) comprising soil and edible crop tissues were analyzed for 16 polycyclic aromatic hydrocarbons (PAHs) and eight heavy metals using Gas Chromatography-Mass Spectrometry (GC-MS) and Atomic Absorption Spectrophotometry (AAS). Labels were assigned according to international safety thresholds, and multiple machine learning models-Artificial Neural Network (ANN), Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN)-were trained using nested, group-aware 5-fold cross-validation, with additional leave-one-site-out validation to test geographical generalization. Among the tested models, ANN achieved the highest predictive accuracy of 97.8% (AUC = 0.98), followed by RF (94.7%) and SVM (93.6%). Feature importance analysis revealed that Cd (importance = 0.214), benzo[a]pyrene BaP ( 0.187), and Pb ( 0.162) were the most influential predictors of crop safety. These findings demonstrate that integrating contaminant profiling with machine learning provides a robust framework for environmental risk assessment and supports safe agricultural practices in landfill-impacted regions.