Spatiotemporal Assessment of Flood Susceptibility in Sylhet Division, Bangladesh Under CMIP6 Projections

Introduction: The haor wetlands in Sylhet Division, which are flooded by the flash floods from the Meghalaya hills, are among the most flood-prone regions in Bangladesh, but the extent of the flood hazard is poorly quantified in the context of the future climate. This study aims to map the current flood susceptibility in Sylhet and estimate its changes for contrasting different SSPs from the CMIP6 model.
Methods. Using Google Earth Engine, historical flood extent (1990-2024) was outlined from Sentinel-1 SAR backscatter and JRC Global Surface Water. Random Forest technique was used to train a flood-susceptibility model using terrain predictors derived from SRTM/HydroSHEDS (elevation, slope, Height Above Nearest Drainage, topographic wetness index and distance to rivers), CHIRPS rainfall and land cover over these flood masks. Susceptibility was projected for future using the precipitation change factor from CMIP6 NEX-GDDP projection for the 2050 and 2100 horizons using the SSP2-4.5 and SSP5-8.5 scenarios. The area under the ROC curve (AUC) and overall accuracy were used to evaluate the skill of the models, while the trend in rainfall was tested using the Mann-Kendall test and the Sen's slope.
Results:There was an increase in flood susceptible area under both pathways, with the largest increase in flood susceptible area under SSP5-8.5 by 2100 ([X] % above the baseline 1990-2024, compared to [Y] % under SSP2-4.5). Random Forest classification of susceptibility achieved an AUC of [AUC] and overall accuracy of [OA]%, with [dominant factor] being the most important factor. The central part of the haor heartland and low lying floodplain emerged as the most vulnerable regions and there was a pronounced increasing tendency for the pre-monsoon rainfall (Mann-Kendall p ≈ [p]).
Conclusions: When comparing SAR-based flood mapping with machine learning susceptibility and rainfall change based on the scenarios of the climate models for the twenty first century, it is revealed that, the flood hazard in the Sylhet area will increase and the rate of change would be maximum under the high-emission scenario. The way is a low-cost, replicable framework for haor-specific flood adaptation and early warning/anticipatory action in monsoon dominated Bangladesh.