Geospatial and machine learning-based flood susceptibility assessment in Nagaon District, Assam, India

Abstract

Flooding is a major natural hazard in the Brahmaputra floodplain, severely impacting lives, agriculture, and infrastructure. This study presents a geospatial and machine learning-based flood susceptibility assessment for Nagaon district, Assam, integrating remote sensing and GIS with ensemble boosting models. A spatial database of thirteen flood conditioning factors and flood inventory points was developed using multi-source data. Advanced models, XGBoost and LightGBM, were applied, and their interpretability was enhanced using SHAP analysis. Results indicate that flood susceptibility is primarily controlled by Normalised difference vegetation index, land use/land cover, rainfall, and proximity to rivers. Approximately 5.29% of the study area falls under high to very high susceptibility zones. Model validation shows strong predictive performance, with AUC values ranging from 0.784 to 0.845, further improved by ensemble approaches. The study provides a robust, interpretable framework for flood susceptibility mapping, supporting effective disaster management and sustainable planning in flood-prone regions.