Unlicensed groundwater pumping threatens aquifer sustainability, particularly in water-stressed, data-scarce regions where conventional ground-based monitoring is no longer viable. Without physical well-monitoring, regional authorities lack the data required to track and manage groundwater depletion. To address this gap, we propose an interpretable machine learning (ML) framework that uses Earth Observation (EO) data as a proxy for groundwater monitoring.
Moving beyond static, distance-based GIS methods, our approach detects illegal extraction by measuring divergences between expected hydroclimatic stress and observed dry-season vegetation health. Specifically, we detect dynamic anomalies in evapotranspiration and greenness. These remote sensing signals are combined with static constraints—such as topography and distance to known surface canals—to train interpretable spatial ML algorithms. The framework incorporates strict spatial validation to prevent data leakage, alongside socio-economic triangulation to support causal validity.
Because transparent evidence is needed for legal enforcement, 'black box' models are avoided. Instead, we prioritize explainability to assure policymakers that risk predictions are driven by physical hydrology and infrastructural isolation rather than statistical noise. The final model outputs a continuous spatial probability surface and discrete risk tiers, which are ethically aggregated to protect unrecorded infrastructure locations while still providing decision-grade intelligence to field inspectors. This framework offers a scalable tool to help prioritize water policy enforcement in data-scarce regions.