Optimized ensemble techniques for nitrate concentration modelling from groundwater integrated with mRMR extraction algorithm

The study presents a comprehensive machine learning framework for the prediction of nitrate (NO 3 ) concentration in groundwater across multiple regions in India. A total of 15 physicochemical parameters were evaluated using minimum redundancy maximum relevance (mRMR) algorithm to rank the most influential features affecting NO 3 levels, whereby; residual sodium carbonate (RSC), electrical conductivity (EC), magnesium (Mg), and calcium (Ca) emerged as the most significant variables. Two categories of predictive models were developed: single models; Gaussian Process Regression (GPR-M1 to M4) and Artificial Neural Networks (ANN-M1 to M4), as well as optimized ensemble techniques. Among the single models, ANN-M1 yielded the best performance with R 2 = 0.807, RMSE = 55.47, and EN-ANN-BO optimized ensemble technique achieving an R 2 of 0.998, and MAE of 5.10 outperformed all the techniques used in the current study. More also, the optimized ensemble techniques significantly outperformed all baseline GPR and ANN model in both training and testing phases. During the training, the best baseline model (ANN-M1) achieved R 2 of 0.817 with an RMSE of 50.74, while the strongest GPR baseline (GPR-M1) produce an R 2 of 0.708 and RMSE of 64.07. In contrast, the optimized ensemble models (EN-ANN-BO and EN-GPR-BO) achieved R 2 values of 0.998–0.999 with RMSE values below 6.63, representing an RMSE reduction of approximately 87–94% relative to the best baseline models. The study demonstrates a scalable and transferable approach to groundwater monitoring, offering an effective tool for policymakers and environmental agencies to proactively address nitrate pollution.