Amir Naghibi

Diffuse nitrate pollution remains a persistent threat to drinking-water resources, yet current predictive frameworks rarely capture how the spatial composition (land-cover fractions) and configuration (landscape fragmentation) of land systems jointly modulate leaching. This study presents a novel multiscale framework integrating hydro-climatic, soil and anthropogenic predictors with land-cover fractions and seven fragmentation metrics calculated at nested grids of 50, 100, 200, and 500 m. This AI-based framework implemented feature selection techniques to identify the top 30 input features (out of 116 initial factors) and then modelled nitrate vulnerability in two settings: i) only using hydroclimatic, anthropogenic and soil factors; ii) the same factors plus land-cover fractions, and landscape fragmentation. The classification was made in an ordinal manner using a multi-layer perceptron encoder (MLP-encoder) benchmarked with proportional-odds ordinal regression with 100 repetitions on different data split configurations. When including the introduced land-cover fractions and landscape fragmentation at multiscale, the accuracy of the model significantly improved by 11%. The results confirm that incorporating both land-use composition and fragmentation at an intermediate scale markedly enhances nitrate-vulnerability mapping and provides spatially explicit, policy-relevant levers for intervention design. These levers include limiting large contiguous cropland blocks and conserving forest-wetland complexes to support catchment management and groundwater-protection planning at municipal and governmental levels.