mLSTM deep learning model to map flood hazard in an arid catchment

Flood hazard map is necessary to develop strategies for mitigation of flood damages and sustainable management of catchments especially in drylands with flash flood and megafloods. Here, we applied multiplicative long short-term memory (mLSTM) deep learning model to map flood hazard in an arid catchment – Shamil-Minab plain – in southern Iran. In order to, variables controlling flood hazard consisting of variables extracted from digital elevation model (DEM) (e.g., curvature, plan curvature, profile curvature, slope, stream power index (SPI), topographic position index (TPI)), normalized difference vegetation index (NDVI), hydrological variables (e.g., river density, distance from river), land use, lithology and soil types were mapped spatially. An inventory map for flood was generated according to field survey and historical data. Inventory map provides training and test datasets for building the predictive flood models. Finally, mLSTM model used to map flood hazard in the study area, and its performance was assessed by accuracy measures. The results shown that 27%, 19.7% and 26% of total area were belonged to very low, low and moderate hazard classes, whereas high and very high hazard classes were occupied 15.9% and 11.4% of total study area, respectively. The combination of our suggested methodology with MCDM models can be useful to map flood risk, and to mitigate destructive consequences of floods in drylands.