An optimized stacking ensemble framework with multi-layer explainable AI for soil heavy metal contamination...

Human health, ecosystem stability, and agricultural sustainability are all seriously threatened by soil heavy metal contamination, necessitating precise, trustworthy, and understandable assessment frameworks. The intricate nonlinear interactions, multicollinearity, and class imbalance features present in soil physicochemical data are frequently difficult for traditional statistical techniques and stand-alone machine learning models to capture. This study suggests a transformer-driven stacking ensemble framework for multiclass soil contamination classification that is optimized using Genetic Algorithms (GA) in order to overcome these difficulties. To improve generalization and avoid data leakage, the FT-Transformer, XGBoost, and Multilayer Perceptron (MLP) models were first optimized separately using a single GA-based hyperparameter search strategy. They were then combined using an out-of-fold (OOF) probability stacking technique. A multi-level Explainable Artificial Intelligence (XAI) framework has been implemented to make sure that decisions are clear and easy to understand. TreeSHAP was used to look at the global feature importance of the optimized XGBoost model. It showed that heavy metals like Zn, Pb, Cr, and Ni are the main predictors for all levels of contamination. Anchor explanations on the stacking ensemble improved model-level interpretability by helping to create high-precision decision rules. Counterfactual explanations were used to find the smallest changes to features needed to change contamination predictions. The suggested framework was better at making predictions than each base learner on its own, and it also made the predictions easier to understand and more aware of risks. This study offers a scalable, optimized, and interpretable AI-driven decision-support system for monitoring soil quality and assessing environmental risks.