Hydrogeochemical assessment and groundwater fluoride prediction in Bathinda district using deep learning.

Fluoride contamination in groundwater is a serious public health concern, especially in semi-arid regions like Bathinda in Punjab, where people rely heavily on groundwater for drinking and daily use. Despite several studies on fluoride contamination, research integrating uniform spatial sampling, hydrogeochemical assessment, and advanced predictive modeling remains limited. This study addresses that gap by automating groundwater fluoride prediction using deep learning techniques and evaluating seasonal hydrochemical variations in the Bathinda district. The study collected 226 groundwater samples across the pre-monsoon and monsoon seasons using GIS-based sampling at approximately 5-km intervals. Hydrochemical parameters were analyzed following APHA standards, and the Water Quality Index (WQI) was calculated. Fluoride concentrations were spatially mapped using GIS and modeled using both machine learning and deep learning approaches, specifically the Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Deep Neural Network (DNN), and hybrid CNN-LSTM models. To enhance model robustness, data augmentation was applied using the nearest-neighbor interpolation, creating 30,000 synthetic points. Among all models, the DNN outperformed the others, with an R² of 0.92 (pre-monsoon) and 0.91 (monsoon), followed by the hybrid CNN-LSTM. Spatial analysis revealed fluoride hotspots exceeding WHO limits (>1.5 ppm), strongly associated with specific lithological units, land use land cover (LULC), and geomorphological features. This integrated approach enables accurate fluoride prediction in unsampled areas, supporting early risk identification and informed decision-making. These findings are highly relevant to strategies for groundwater management, environmental monitoring, and public health planning in regions affected by fluoride.