Automatic Extraction Model of Lake Remote Sensing Images in Cold-Dry Regions Using Improved U-net Fusion CRFs

Lakes in cold and arid regions play an important role in the daily lives of people. However, the health of the lake ecosystem is seriously affected by climate change, especially human activities. Remote sensing satellites provide effective means for lake ecological health monitoring. How to extract lakes in cold and arid regions from remote sensing images has become a research hotspot. There exist some issues in most of the traditional methods regarding extraction accuracy as well as susceptibility to interfer-ence. Therefore, this paper proposes a novel method named AU-net, which is based on the improved U-net model and Fully Connected Conditional Random Field (FCCRF). Firstly, semantic segmentation datasets are established using Landsat 8 OLI remote-sensing images for the intended research objects, i.e., Wulliangsuhai Lake and Hulunhu Lake. Then, the U-net model is constructed. To improve the model’s adaptive ability, the Convolutional Block Attention Module is added after the feature map is extracted from the model. Finally, post-processing is conducted based on the FCCRF. The experimental results depict that the AU-net extraction accuracy is superior to the Normalized Difference Water Index, Deeplab v3+ model, and the initial U-net model. Moreover, AU-net can eliminate the interference of complex backgrounds and small holes. Further-more, AU-net has achieved good extraction results for Wuliangsuhai Lake and Hulunhu Lake in different seasons. © 2024, Taiwan Ubiquitous Information CO LTD. All rights reserved.