An open-data AI–GIS–MCDM framework for consensus prioritisation of diffuse agricultural pollution susceptibility: A Delhi proof-of-concept

Diffuse run-off from cultivated land is a leading, unregulated pressure on peri-urban surface water, yet regulators must target scarce inspections across large catchments with no point-source register. We present an open-data decision-support framework that prioritises 1 km grid cells of the National Capital Territory of Delhi by their relative susceptibility to diffuse agricultural pollution. Rather than claiming to detect pollution, the framework estimates susceptibility from openly available proxies: a multi-year (2018–2025) post-monsoon Sentinel-2 median composite, terrain from the Copernicus DEM, and distance to water from HydroSHEDS and JRC layers, all summarised as agriculture-masked, area-based per-cell statistics with multi-year median, trend, and inter-annual-variability features. A Random Forest, evaluated with spatial-block cross-validation to remove spatial-autocorrelation leakage, reproduces a transparent susceptibility rule with accuracy 0.865 (95% CI 0.846–0.883), kappa 0.797 and a high-class AUC of 0.982, and yields a calibrated probability (Brier 0.057; ECE 0.063) that serves as one decision criterion. Seven literature-weighted criteria are aggregated by TOPSIS and cross-checked against VIKOR, FN-MABAC and equal-weight TOPSIS. The methods agree globally (Spearman 0.95–0.99) but share only 15–38 of the 50 highest-priority cells, so the coarse ranking is robust while the exact shortlist is method-sensitive. A global Monte-Carlo sensitivity analysis over simultaneous ± 20% weight uncertainty gives a minimum Spearman of 0.935, but the top-50 overlap stays at or above 40/50 in only 67% of runs. Eleven cells are selected by at least three methods and stay in the top 50 in at least 95% of runs, forming a robust first inspection tranche. We frame the study explicitly as a reproducible proof-of-concept: it screens inspection priority and would require field validation before operational use.