Integrating Machine Learning and Participatory GIS with Multi-Temporal Remote Sensing for Flood Susceptibility and Vulnerability Mapping in Nkhotakota, Malawi

This study presents an integrated approach for assessing flood susceptibility and vulnerability in Nkhotakota District, Malawi, by combining remote sensing, field-validated observations, and machine learning techniques. Physical flood susceptibility was estimated using the Soil Conservation Service–Curve Number (SCS-CN) method to model runoff potential, alongside Sentinel-2 imagery processed in Google Earth Engine to delineate inundation areas. Machine learning algorithms, including Support Vector Regression (SVR), Random Forest (RF), and K-Nearest Neighbour (KNN), were trained on 300 flood and non-flood points derived from ground-truth verification, with SVR providing the most accurate predictions (MAE = 0.041, MSE = 0.042, RMSE = 0.205, AUC = 0.98). Sub-watersheds were classified into high, moderate, and low susceptibility zones, highlighting flood-prone areas along major rivers and low-lying floodplains. Socio-economic vulnerability was assessed through 130 household surveys covering economic, physical–infrastructural, institutional–policy, and social–cultural dimensions. Analysis showed the highest absolute vulnerability in the economic dimension (0.48) and greatest relative sensitivity in the social dimension (rank = 2.94), reflecting the interplay of low income, agricultural dependence, and dense settlements near rivers. Combining susceptibility and vulnerability revealed hotspots where high physical exposure coincides with high social sensitivity, providing priority areas for targeted interventions. The study offers practical guidance for flood risk management, including structural measures, early warning, and community-based resilience strategies, and demonstrates the effectiveness of integrating remote sensing, ML, and participatory socio-economic data in flood risk assessments.