Assessment of a family of recurrent neural network models for flood susceptibility Mapping: An explainable glass-box approach

초록

Owing to climate change, the frequency of flood events has increased, causing substantial harm to society. Consequently, identifying flood-prone areas is a fundamental step toward effective flood management and mitigation. This study aims to compare recurrent neural network (RNNs) deep learning families for flood susceptibility mapping in the Gorganrood watershed in the Golestan Province, a hotspot for floods area in Iran. In the present study, an Elman-RNN model was developed, and the results were compared with those of regular RNN, gated recurrent unit (GRU), and long short-term memory (LSTM) models. Geodata comprising 230 historical flood events collected by analyzing Sentinel-1 and 14 geo-environmental characteristics were considered for the modeling. The SHapley Additive exPlanations (SHAP) approach was used to convert the black box results to an explainable glass box. Finally, the effectiveness of the models was assessed using the area under the receiver operating characteristic curve (AUC). The SHAP results revealed that altitude, distance from rivers, normalized differences in vegetation index, rainfall, and drainage density were the most important factors influencing flooding in the Gorganrood Watershed. The validation results revealed that the Elman RNN, which had an AUCtesting of 0.969, possessed the highest generalization power, followed by the GRU (AUCtesting = 0.952), LSTM (AUCtesting = 0.947), and RNN (AUCtesting = 0.943) models. The majority of the flood-prone areas were located downstream of the watershed, and approximately 19.92 % of the Gorganrood Watershed was very susceptible to future flooding. These results demonstrate the superiority of the Elman RNN model over other recurrent models, making it a practical and reliable flood modeling tool.

키워드