D-CDA:A Denoise And Change Detection Approach For Flood disaster
location From SAR Images
Abstract
Flood disasters are among the most devastating natural disasters
worldwide. The occurrence of such disasters is often accompanied by
strong precipitation and other weather factors, making it more difficult
to identify the affected areas. Moreover, Synthetic Aperture Radar (SAR)
technology can capture images in 24-hour window and penetrate through
clouds and fog. The change detection technology, based on SAR images, is
generally utilized to locate disaster-stricken areas by analyzing the
differences between pre- and post-disaster images. However, this method
mainly faces two challenges: the presence of speckle noise reduces the
accuracy of the difference detection and the lack of a suitable SAR
dataset for flood disaster change detection. Therefore, this research
proposes a novel two-stage approach for locating the flood disaster
area, named Denoising-Change Detection Approach (D-CDA). The first stage
consists of a nine-layer denoising network with an encoder-decoder
structure, called SAR Denoising Network (SDNet). It utilizes a
multi-residual block and parallel convolutional block attention module
to extract features during the encoding process to suppress the noise
component. In the second stage, a novel convolution neural network is
proposed to detect the changes between bitemporal SAR images, namely
Coordinate Attention Fused Network (CAFNet), which combines the Siamese
network and UNet++ as the backbone and fuses multi-coordinate attention
modules to enhance the change features. Moreover, a change detection
dataset (ZhengZhou Flood - ZZF-dataset) is constructed using Sentinel-1
SAR images based on the flood disaster in Zhengzhou of China in 2021.
The simulations verify the effectiveness of the proposed method.