A novel entropy-based machine learning frame work for flood risk mapping in Pakistan

Pakistan, South Asia. This study presents a three-tiered framework, the District-level Flood Risk Assessment Model (D-FRAM), which uses multiple satellite datasets with national surveys to analyze spatial and temporal flood risk nationally. Satellite-derived environmental, topographical, and hydrological data were used to predict the Flood Susceptibility Index (FSI) using supervised machine learning models. A hybrid unsupervised clustering model was used to derive the Flood Vulnerability Index (FVI). These were combined with historic Max-flood extent (FM) and Flood Frequency (FF) (2010–2023) to compute the Flood Risk Index (FRI) through entropy-weighted multi-criteria analysis. XGBoost proved most effective for predicting FSI, achieving an AUC of 0.95 and accuracy of 0.89. The framework identified districts like Sujjawal, Jacobabad, Kashmore, and Jaffarabad as having highest flood risk due to low elevation, proximity to rivers, and poor infrastructure. Jaffarabad had 96.47 % of land classified as "Very High" FSI, while Sindh showed the largest provincial "Very High" susceptibility at 50.30 %. The FVI analysis revealed Sibi (1.0), Kech (0.91), Mastung (0.89), and Shahdad Kot as vulnerable districts with limited services. FRI peaked in August, with district maxima in Sheikhupura (0.80, July), Kashmore (0.86, August), and Sujjawal (0.87, October). Southern provinces, mainly Sindh and eastern Balochistan, had the highest Critical Hotspots, with Sindh comprising 48.3 %. • First nationwide flood risk mapping integrating ML and unsupervised clustering. • XGBoost achieved 89.44 % accuracy for flood susceptibility classification. • SOM-UMAP clustering quantified socioeconomic vulnerability at district level. • Monthly flood risk mapping captured dynamic monsoon-driven variations. • Entropy-based MCDA enabled objective integration of flood risk components.