Remote Sensing Image River Segmentation Method Based on U-Net

Abstract

River segmentation based on remote sensing images plays an important role in water conservancy business work, water wading monitoring work, and flood disaster prevention. In actual remote sensing images of rivers, most of the backgrounds are complex, and there is no public remote sensing image dataset specifically for the study of river segmentation. The traditional river segmentation methods have rough edge information and serious noise. To solve the above problems, this paper firstly preprocesses the Gaofen Image Dataset (GID) and Remote Sensing Image Block Segmentation Dataset (BDCI), and creates two datasets for river segmentation in high-resolution remote sensing images respectively (GID-river and BDCI-river) and then proposed a river segmentation method based on U-Net. On the basis of the original U-Net, the ResNet34 and VGG16 structures were combined to strengthen the feature extraction ability of the network, so as to achieve more accurate river edge details. The experimental results shows the mIoU of the ResNet34-UNet network on the GID-river dataset reaches 93.6%, and the mPA of the VGG16-UNet network on the BDCI-river dataset reaches 82.1%.