Multiple Kernel Learning with Maximum Inundation Extent from MODIS Imagery for Spatial Prediction of Flood Susceptibility

Abstract

This paper proposes a new technology of spatial prediction for flood susceptibility. Multiple kernel learning was used to build the flood susceptibility model and predict the flood inundation risk of the Sanhuajian Basin of the Yellow River. Based on the historical flow records of the Huayuankou Site and the MODIS remote sensing images of the study area, the maximum inundation range was extracted by the open water likelihood index method, and the flooded and non-flooded sample sites were selected. Considering the availability of pertinent literatures and data, ten flood conditioning factors were defined as the sample characteristics. The model performance was evaluated in terms of accuracy, F1 score, and AUC. According to the results, multiple kernel learning significantly outperforms the support vector machine adopting single kernel, and NLMKL demonstrates the best comprehensive performance. The flood susceptibility map generated by MODIS remote sensing images and multiple kernel learning, therefore, can provide effective help for researchers and decision makers in flood management.