Urban flood susceptibility assessment based on convolutional neural networks

In this study, a convolutional neural network (CNN)-based approach is proposed to assess flood susceptibility for urban catchment. Nine explanatory factors covering precipitation, topographical, and anthropogenic aspects were selected and two CNNs, SCNN and LeNet-5, were implemented to identify the relationship between the explanatory factors and flood inventory between 2004 and 2014 in the Dahongmen catchment in Beijing, China. The performance of the CNNs was compared with that of support vector machine (SVM) and random forest (RF) models for three model input strategies (point-based, array-based and imaged-based strategies). The results showed that: (1) The two CNNs performed better than the SVM and RF, with an accuracy of 0.90 for SCNN and 0.88 for Lenet-5 in the testing period. (2) The CNN-based approach provided more reliable flood susceptibility maps than the comparative models with the highest area under the curve (AUC) index of 0.9 validating by another flood inventory. (3) The upstream inundations induced by pluvial flood cannot be accurately identified by using point-based and array-based strategies. These errors were corrected by considering topographical information using the imaged-based CNN approach. (4) The fixed architecture LeNet-5 produced satisfactory results and avoided the time-consuming process of architecture selection of CNN. We conclude that the proposed CNN-based approach is a valid approach and provides a high-quality susceptibility map for urban mitigation and flood management.