River Extraction from Remote Sensing Images in Cold and Arid Regions Based on Deep Learning

The extraction of rivers in cold and arid regions is of great significance in the application of ecological environment monitoring, agricultural planning, and disaster early warning in cold and arid regions. In recent years, there have been more studies on river extraction, but river extraction for cold and arid regions is still in its infancy. The rapid development of deep learning provides new ideas for river extraction in cold and arid regions. A professional data set was produced based on the characteristics of cold and arid regions to provide support for river extraction in remote sensing images in cold and arid regions. Combining transfer learning and deep learning, the ResNet50 network was migrated to the Linknet network to obtain the R-Linknet network, which was used to improve the recognition accuracy of the network. At the same time, the dense atrous spatial pyramid pooling was combined with the R-Linknet network to expand the receptive field of the network, which can extract more detailed information and increase the coherence of the extracted river. A new loss function was combined with the Dice Loss function and the binary cross entropy loss function during training. The verification on the data set showed that compared with semantic segmentation networks, the proposed method had an accuracy rate of 0.216, 0.099, 0.031, 0.056 and 0.023 higher than that of FCN_8s, ResNet50, DeeplabV3, Unet and the original Linknet network, respectively, and the intersection over union was increased by 0.190, 0.142, 0.056, 0.105 and 0.028, respectively. After adding dense atrous spatial pyramid pooling, it increased the pixel accuracy by 0.023, and improved the intersection over union by 0.050. After training with the new loss function, the pixel accuracy and crossover ratios were increased by 0.019 and 0.022, respectively. The rivers extracted by this method were more clear and consistent, and can meet the needs of subsequent research. © 2020, Chinese Society of Agricultural Machinery. All right reserved.