WaterDetectionNet: A New Deep Learning Method for Flood Mapping With SAR Image Convolutional Neural Network

Abstract

Floods are among the world's worst natural disasters, causing significant damage to people, infrastructure, and the economy. Since synthetic aperture radar (SAR) can work in all weather and is not affected by clouds and rain, the use of SAR for flood mapping and disaster assessment has obvious advantages. However, SAR images are highly susceptible to speckle noise, shadows, and distortions, which affects the accuracy of traditional water body extraction methods. To solve this problem, we designed a new model called WaterDetectionNet (WDNet) based on SAR remote sensing images and convolutional neural network, which has a strong water extraction capability for accurate flood mapping. In order to improve the generalization ability of the model, we used a semiautomatic strategy to generate the SAR dataset S1Water containing rich semantic information with diversity. Compared with the traditional machine learning and deep learning methods, we introduced a self-attention module to increase spatial and channel attention, and adaptively update the network weights, which improved the model performance and extraction accuracy of the real case study of the Poyang Lake flood in 2020. The experimental results show that the accuracy, recall, intersection over union, and F1 score of the WDNet model were 0.986, 0.994, 0.974, and 0.987, respectively. This method is expected to provide a cost-effective alternative to global rapid flood mapping, improve the reliability of flood disaster analysis, and offer a reference for postdisaster emergency management.

Keywords

• Convolutional neural network
• flood mapping
• radar remote sensing
• self-attention
• Sentinel-1 SAR
• water body extraction