Application of hybrid model-based deep learning and swarm‐based optimizers for flood susceptibility prediction in Binh Dinh province, Vietnam

Flooding is a serious natural hazard. It causes considerable impact on human life, the environment, and property, worldwide. Building a highly accurate flood susceptibility map can reduce disaster damage; it has become the main approach in flood risk management. The objective of this research is to build flood susceptibility maps for Binh Dinh province in Vietnam, applying modern machine learning and remote sensing methods, namely deep neural networks (DNN) and swarm-based optimization algorithms such as aquila optimizer algorithm (AO), sea lion optimization (SLnO), elephant herding optimization (EHO), the naked mole-rat algorithm (NMRA), Stochastic Gradient Descent (SGD). The geospatial distribution analysis approach was used to construct the input data, including 1883 sample points and 12 conditioning factors. Several well-known algorithms were used as reference models to compare the accuracy of each proposed model. The statistical indices root mean square error (RMSE), area under curve (AUC), mean absolute error (MAE), accuracy, and F1 score were used to validate the proposed model. The results show that five optimization algorithms successfully in buiding flood susceptibility maps and these models performed well with an AUC value of more than 0.97. The DNN-NMRA model came first (RMSE = 0.16, AUC = 0.99), followed by DNN-SLnO (RMSE = 0.39, AUC = 0.99), DNN-EHO (RMSE = 0.41, AUC = 0.99), DNN-AO (RMSE = 0.46, AUC = 0.97), and DNN-SGD (RMSE = 0.49, AUC = 0.93) respectively. Furthermore, the results show that the precision of new models (DNN-NMRA, DNN-SlnO, DNN-EHO, DNN-AO) surpassed the standard model DNN-SGD. The results of this study are useful in the construction of appropriate flood management strategies in at-risk regions.