Performance comparison of two deep learning models for flood susceptibility map in Beira area, Mozambique

Abstract

Floods are potentially devastating natural hazards that can threaten human life and ecosystems. The serious impact caused by floods is the reason for the urgent need to provide an appropriate flood susceptibility map (FSM). Floods can be triggered by various meteorological events, such as storms in coastal areas. On 14 March 2019, Cyclone Idai produced massive and severe flooding inundation in the coastal city of Beira, Mozambique. The extensive flood damage in Beira necessitated providing an accurate FSM and analyzing the factors associated with flooding. This study integrated the data of synthetic aperture radar (SAR), geographic information system (GIS), and two deep learning models, namely, the group method of data handling (GMDH) and convolutional neural network (CNN), to generate a precise FSM. We compared the performance of CNN and GMDH models in generating the FSMs for a case study of the March 2019 flood in the Beira Area, Mozambique. The result showed that the models produced a similar pattern of FSM, where the areas near the river, having a lower slope and lower altitude, have a higher flood risk. The FSM generated by CNN and GMDH models were compared to find a better model for flood risk assessment. The prediction performance of CNN model (AUC = 0.90 and RMSE = 0.022) was better than the result of GMDH model (AUC = 0.87 and RMSE = 0.089). The study should be used to mitigate the flood disaster and reduce flood risk in the future.