Multi-temporal flood susceptibility mapping using Sentinel-1 SAR time-series and machine learning: case study of Sebkhet Ariana (NE Tunisia)

Floods represent a recurrent natural hazard in Mediterranean coastal environments, driven by climate variability, low-lying topography, and land-use change. This study proposes a multi-temporal flood monitoring and susceptibility framework using Sentinel-1 Synthetic Aperture Radar (SAR) data for the Sebkhet Ariana coastal wetland, Tunisia. Monthly SAR composites were generated for September from 2019 to 2024. Flooded areas were extracted using histogram-based minimum error thresholding on VV-polarised backscatter. Multi-temporal observations were used to derive flood frequency indicators. A Random Forest model integrated SAR-derived flood frequency with environmental conditioning factors from Copernicus DEM and ESA WorldCover to produce a flood susceptibility map. Results show strong interannual variability, with recurrent flooding concentrated in low-lying and urban–agricultural transition areas. The Sentinel-1-derived patterns showed generally good correspondence with independent Sentinel-2/MNDWI observations, although the degree of agreement varied among monitoring periods, particularly during the anomalous 2023 conditions. Under a class-balanced independent buffered spatial validation design, the Random Forest model achieved an Overall Accuracy of 89.2% and a ROC-AUC of 0.953, demonstrating discrimination between flooded and non-flooded, confirming the effectiveness of combining SAR time series and machine learning for flood recurrence analysis and susceptibility mapping. The study provides a reproducible approach for flood monitoring in data-scarce coastal environments.