Feature Attribution-Driven Flood Susceptibility Assessment: an Integrated Gradients Approach in Seti Gandaki River Basin

This research introduces a sophisticated framework for flood susceptibility analysis in theSeti Gandaki watershed that integrates multiple geospatial features with deep learning method-ologies. The proposed neural network model leverages ten critical geomorphological, hy-drological, and environmental features to predict flood-prone areas with high precision. Ourmethodological approach encompasses comprehensive geospatial data preprocessing, fea-ture alignment, and threshold-based balanced sampling to ensure robust model training. Thedeveloped deep learning architecture incorporates multiple hidden layers with dropout regu-larization and batch normalization, achieving exceptional performance metrics exceeding98% accuracy, 99.65% precision, and 97.22% recall. Feature attribution analysis usingIntegrated Gradients identified Topographic Wetness Index (TWI), distance to rivers, andrainfall as the primary contributors to flood susceptibility prediction. The study generatedhigh-resolution flood susceptibility maps with quantified prediction uncertainty, providingcritical spatial intelligence for flood risk management. Comparative analysis with traditionalAnalytical Hierarchy Process (AHP) revealed significant methodological differences, withonly 8.52% spatial agreement between approaches. This research demonstrates the efficacyof integrating advanced machine learning techniques with traditional geospatial analysis toenhance flood prediction capabilities, offering valuable decision support for disaster man-agement and mitigation strategies in hydrologically complex Himalayan watersheds.Keywords: Flood susceptibility, Deep learning, Geospatial analysis, Feature attribution, SetiGandaki watershed, Disaster management, Neural networks, Prediction uncertainty