Abstract:

The interference of spectral foreign objects such as vegetation, shadows and clouds leads to the low integrity and poor extraction effect of remote sensing images. A remote sensing image water extraction model MAP_UNet(A UNet of Combining Multi Atrous Convolution and Pooling Model) is proposed in this study, which combines multi-level cavity convolution and pooling model. The model uses UNet as the reference codec network to extract different dimensional features of water bodies, introduces double recursive residual module to prevent gradient disappearance of deep network, and uses multiple modules to integrate spatial cavity convolution and maximum pooling to capture a larger range of feature information and further strengthen the feature semantic relationship of adjacent scales. In order to verify the effectiveness and advance of the proposed method, experiments are carried out using high-resolution visible remote sensing image data and the results are compared with open deep learning semantic segmentation algorithm. The experimental results show that the MAP_UNet model has achieved good results in extracting accuracy and preventing misdetection of different objects of the same spectrum. The accuracy rate, recall rate, F1-Score and MIoU(Mean Intersection over Union) are 96.20%, 92.64%, 87.27% and 89.10%, respectively. Compared with UNet(U-shaped Network), UNet++ and UNet ASPP(UNet with Atrous Spatial Pyramid Pooling Network) networks, the performance of proposed method has significant improvement.

Key words: water extraction, deep learning, neural network, atrous convolution, semantic segmentation, UNet, remote sensing image