Local-scale flash flood susceptibility assessment in northeastern Bangladesh using machine learning algorithms

Abstract

The recurring flash floods in northeastern Bangladesh have been causing significant damage to the lives and livelihoods of local people every year, underscoring the necessity of effective flood management and mitigation efforts. Identifying flood hazard areas is the very first step in achieving such ends. While machine learning algorithms have been widely used in flood susceptibility studies, there is a significant dearth of research addressing their application for local-scale assessment of flood susceptibility in Bangladesh. Besides, in most of the flood susceptibility studies, only statistical metrics are used to evaluate the prediction capability of machine learning algorithms, overlooking their ability to account for spatial consistency. Therefore, this study attempts to evaluate the performance of Random Forest (RF) and Support Vector Machine (SVM) algorithms in flash flood susceptibility assessment at local scale using statistical and spatial metrics. RF model performed better than SVM with area under the curve (0.964), accuracy (92.1 %) and kappa (0.894). Five susceptibility zones were identified using a natural breaks method: very high, high, moderate, low and very low. The overall spatial agreement between the susceptibility maps of RF and SVM was 73.3 %. For very low susceptible classes, the spatial agreement was highest for RF-SVM (97.15 %) and SVM-RF (81.54 %). The findings of this study can be useful for a more accurate assessment of flash flood susceptibility in similar local settings. Moreover, the prepared susceptibility map can be helpful for taking sustainable measures to mitigate the devastating effects of flood hazard in northeastern Bangladesh.