Machine Learning Applications to Hydrological Variables: Modelling, Estimation and Forecasting Under Changing Climate

Abstract

Groundwater overexploitation and waterlogging represent two sides of Punjab’s intensifying hydrological crisis. This dissertation addresses the dual challenge of declining water tables and surface saturation in one of India’s most intensively irrigated and agriculturally critical states. The research is structured into two integrated themes: (i) fine-resolution modelling and forecasting of reference evapotranspiration (ET₀) across India using hybrid deep learning, and (ii) the development of a physically consistent groundwater vulnerability framework—specifically, the Soil Water Logging Index (SWLI)—to hindcast, back-cast, and forecast waterlogging risk in Punjab. The first part of the thesis develops ADAQNet, a novel AI-driven architecture combining autoencoders, convolutional neural networks, and attention layers, constrained by physical laws of evapotranspiration. CMIP6 climate projections, ERA5 reanalysis, and Indian Meteorological Department datasets were used to downscale humidity and compute ET₀ across diverse agroclimatic zones. ADAQNet outperformed baseline models, showing strong agreement with EEFlux-based ET₀ (error: −1.2% to 2.5%) and validated RH estimates (MAE: 2.1–5.7%). Detailed seasonal and spatial analyses revealed critical ET₀ hotspots, temporal declines driven by atmospheric stilling and increased humidity, and implications for irrigation planning under climate change. The second part presents the SWLI, a machine learning–physical hybrid index built from four hydrologically meaningful sub-indices: Aquifer Productivity Index (API), Water Balance Index (WBI), Soil Infiltration Potential Index (SIPI), and Surface Vegetation Dynamics Index (SVDI). These indices incorporate 15 hydro-environmental predictors derived from Earth observation, field data, and symbolic regression, calibrated and validated across Punjab (2000–2014). SWLI demonstrated excellent accuracy (AUC = 0.91, F1 = 0.91), outperforming traditional indices like DRASTIC and GOD, and was robust under forecasted CMIP6 SSP3-7.0 climate scenarios (2026–2030). A novel back-casting framework extended the analysis to 1970–1990 using bias-corrected reanalysis data, revealing previously undocumented trends and quantifying policy impacts, such as the estimated 0.25 m/year slowdown in depletion following the 2009 Subsoil Water Act. The study concludes that groundwater vulnerability in Punjab is a nonlinear, multi-scalar phenomenon driven by anthropogenic recharge, aquifer heterogeneity, and climatic variability. The combined use of physics-informed feature engineering and hybrid AI models enables accurate, scalable, and interpretable forecasting of both evapotranspiration and groundwater trends. Key recommendations include targeted crop diversification in high ΔR zones, canal modernization to reduce seepage-induced waterlogging, managed aquifer recharge in high-Sy areas, and IoT-based governance linking real-time monitoring to dynamic policy enforcement. This thesis contributes a novel modelling toolkit, validated for both historical and future scenarios, with strong policy relevance for groundwater sustainability not only in India but across global agroecosystems facing similar hydrological stress.