Flood Detection in Sar Images Based on Multi-Depth Flood Detection Convolutional Neural Network

Abstract

(1) Background: For realizing an effective flood detection of Synthetic Aperture Radar (SAR) images in water regions, a post-classification comparison algorithm with Multi-depth Flood Detection Convolutional Neural Network (MDFD-CNN) is proposed in this paper. The MDFD-CNN is used to classify and extract the water regions in the SAR image, aiming at solving the flood detection after high-resolution SAR water regions extraction in complex terrain. (2) Methods: What the MDFD-CNN owning a two-branches network with different depth is used for extracting respectively the water regions in the bi-temporal SAR images using the saliency detection. In the process of recognizing the water regions, to improve the reliability of the training samples and reduce the number of training samples, MDFD-CNN uses the SAR backscatter coefficient significance detection to obtain the salient regions of the extracted training samples. Moreover, a piecewise back-propagation is adopted to optimize the network. With the detected water regions of the bi-temporal images, a post-classification comparison is implemented for detecting the changes in water regions. (3) Results: Compared with the manually extracted features used by conventional methods, what the MDFD-CNN extracted is more robust. Meanwhile, experimental results illustrates the effectiveness, especially strong noise immunity, of the proposed method. (4) Conclusions: The proposed method offers a valid way for flood detection in water regions of SAR images in realistic environments.