A Hybrid CPCB–Sentinel Framework for Temporal and Explainable Water Quality Monitoring in India

Abstract

Reliable and scalable monitoring of inland water bodies remains a critical challenge in regions where laboratory-based measurements are spatially sparse and temporally discontinuous. This study proposes a novel hybrid learning framework that integrates physicochemical observations from the Central Pollution Control Board (CPCB) with multispectral Sentinel-2 satellite features for automated binary water pollution classification across Indian states. Pollution labels are derived from regulatory thresholds on dissolved oxygen (DO) and biochemical oxygen demand (BOD). Unlike prior work that relies solely on random data partitioning, we systematically benchmark classical machine learning and temporal deep learning architectures— including Random Forest (RF), CNN-LSTM, and Transformer-based models—under three validation protocols: random, group-based (state-year), and temporal splits. This deployment-oriented evaluation is a key novelty of the proposed framework. The Transformer model achieves a temporal ROC-AUC of 0.786, while the CNN-LSTM achieves perfect recall under class-weighted training, reflecting a deliberate high-sensitivity configuration rather than overfitting. SHAP-based interpretability analysis identifies spectral bands and turbidity-sensitive indices as dominant predictors. The findings establish a practical and interpretable pathway toward scalable regulatory water quality monitoring under realistic spatial and temporal distribution shifts.