•  
  •  
 

Corresponding Author

Eqab Almajali

Document Type

Article

Keywords

Environmental data analytics, Explainable artificial intelligence (XAI), Intelligent decision support systems, Machine learning, Water quality index classification

Abstract

The increasing volume and complexity of environmental data have highlighted the need for advanced analytical and artificial intelligence methods to support water-resource monitoring and management. Although machine-learning (ML) models have achieved strong performance in water-quality prediction and classification, their black-box nature remains a major obstacle to their use in environmental decision-making and policymaking. This study presents an explainable artificial intelligence (XAI)-based framework for two complementary water-quality assessment tasks: three-class water quality index classification using Dataset A and binary water-potability classification using Dataset B. Eight machine-learning algorithms were evaluated: logistic regression, support vector machine, random forest, extremely randomized trees, bagging classifier, multilayer perceptron, light gradient boosting machine, and categorical boosting. Five-fold stratified cross-validation was used to improve the reliability and generalizability of the evaluation. Shapley additive explanation, local interpretable model-agnostic explanations, permutation importance, and partial dependence plots were applied to interpret model behavior and extract domain-relevant insights. LightGBM achieved the highest mean accuracy for the minimum-value subset of Dataset A (0.946 ± 0.010), whereas CatBoost achieved the highest mean accuracy for the maximum-value subset (0.956 ± 0.006). For Dataset B, Extra Trees achieved the highest mean accuracy (0.682 ± 0.015). The interpretability analyses identified pH, dissolved oxygen, biochemical oxygen demand, conductivity, and microbiological indicators as influential predictors within Dataset A, whereas pH, total dissolved solids, turbidity, and other physicochemical variables contributed prominently to the binary potability predictions in Dataset B. Evaluating two datasets with different prediction tasks enabled a broader assessment of model performance and interpretability across complementary analytical settings. The findings further indicate that combining environmental data analytics, ML, and XAI can support transparent water-resource monitoring and provide a foundation for future deployment in larger, continuously updated, and distributed environmental-monitoring systems.

Download

Share

COinS