Explainable GeoAI Reveals Anthropogenic Dominance Over Terrain in Urban Flood Susceptibility Mapping: A U-Net++-Based Framework for Nagpur Urban Area, India

This repository contains the code and data supporting the manuscript “Explainable GeoAI Reveals Anthropogenic Dominance Over Terrain in Urban Flood Susceptibility Mapping: A U-Net++-Based Framework for Nagpur Urban Area, India” (Khuntia and Katpatal; submitted to Environmental Modelling & Software, ENVSOFT-D-26-03805). The study develops a validation-and-explanation protocol for deep-learning flood susceptibility mapping in the Nagpur Urban Area. It compares 25 models (13 deep learning, 12 classical or hybrid machine learning) under spatially blocked validation with three random seeds, and selects U-Net++ on mean ROC-AUC. Model behaviour is examined with SHAP, Grad-CAM, Integrated Gradients, occlusion and LIME, together with Monte Carlo uncertainty analysis, Sobol sensitivity analysis and ablation. The code package contains the workflow as numbered Python modules (data preparation, model comparison, explainability, uncertainty, sensitivity, hazard mapping, spatial statistics, validation and transferability tests) and as a single-file version. The data package contains 15 aligned predictor rasters (terrain, hydrology, built environment and vegetation; EPSG:32644, about 12.5 m, 1,574 × 1,578 cells), a SAR-derived flood target (the union of three events in September 2023, July 2024 and September 2025), the three event masks, an independent July 2026 validation mask, 100 documented flood-prone locations, and the trained U-Net++ model. Code is released under the MIT licence and author-derived data under CC BY 4.0. Third-party source datasets (Copernicus GLO-30, Sentinel-2, Esri Land Cover, Microsoft Building Footprints, OpenStreetMap, Survey of India toposheets, Maharain rainfall) remain under their original licences. The flood layers show where flooding occurred in the SAR record and are not a design-flood or hazard-zoning product. The July 2026 threshold results are event-optimized and based on a single event.