Deep learning algorithms to develop Flood susceptibility map in Data-Scarce and Ungauged River Basin in India

Flood is considered the most extensive natural disaster around the globe. Kunur River, a riverine landscape of Rarh Bengal, was selected as the study area because this basin has undergone several floods. This research work applied deep learning and benchmark machine learning methods for preparing the flood susceptibility maps (FSMs) at a basin scale. For this work, sixteen flood controlling factors were applied. These predisposing factors were chosen based on field knowledge, previous researchs, and data availability. The FSMs were produced for the better palling and management of natural resources of Kunur River Basin, applying one deep learning model (DLM) includes convolution neural network (CNN) model and three benchmark machine learning methods (BMLMs) including multilayer perceptron (MLP), Bagging, and random forest (RF). The differences in prediction capacity between the models were assessed by applying the Friedman rank test and Wilcoxon test. Performance of the FSMs, evaluated through the precision, accuracy, AUC (area under the curve), and statistical measures revealed that CNN has the highest AUC values (0.934) followed by MLP (0.927), Bagging (0.897), and RF (0.900) respectively. The CNN model's prediction capacity is slightly better than Bagging, RF, and MLP models. Finally, we can conclude that the deep learning model is more robust than the benchmark MLMs (RF, MLP and Bagging) and CNN is excellent alternative for FSMs considering the used variables.