Flood susceptibility mapping using convolutional neural network frameworks

Abstract Flood is a very destructive natural disaster in the world, which seriously threaten the safety of human life and property. In this paper, the most popular convolutional neural network (CNN) is introduced to assess flood susceptibility in Shangyou County, China. The main contributions of this study are summarized as follows. First, the CNN technique is used for flood susceptibility mapping (FSM) through two different CNN classification and feature extraction frameworks. Second, three data presentation methods are designed in the CNN architecture to fit the two proposed frameworks. To construct the proposed CNN-based methods, 13 flood triggering factors related to flood historical events in the study area were prepared. The performance of these CNN-based methods was evaluated using several objective criteria in comparison to the classic support vector machine (SVM) classifier. Experiments results demonstrate that all the CNN-based methods can help produce more reliable and practical flood susceptibility maps. For example, the proposed CNN-based classifiers were 0.022–0.054 higher than SVM in terms of area under the curve (AUC). In addition, in the classification process, CNN-based feature extraction can effectively improve the prediction capability of SVM by 0.021–0.051 in terms of AUC. Therefore, the proposed CNN frameworks can help mitigate and manage floods.