Abstract: Crevasses are critical indicators of glacier dynamics, and studying them can effectively monitor glacier changes. Nevertheless, conventional remote sensing techniques for crevasse detection are typically hampered by significant manual effort, poor efficiency, and limited applicability in intricate glacial terrains. To overcome these limitation, GeoScene Pro 3.1 software and high-resolution remote sensing imagery are utilized in this research, selecting the Geladandong Glacier region at the Yangtze River source as the study area. The application of deep learning techniques, specifically the U-Net pixel classification model, HED edge detector, and BDCN edge detector, is explored in the detection and extraction of crevasses. First, by adjusting different combinations of training sample sets and training parameters, the three models mentioned above were trained and analyzed for crevasse detection for five times. The detection results show that both the U-Net pixel classification model and the HED edge detector exhibit suboptimal detection performance, however, the BDCN edge detector can effectively suppress the interference of detection noise, and efficiently and accurately detect and identify the distribution location and width of crevasse. Next, by using the BDCN edge detector to detect crevasses on the four glaciers of Geladandong, the detection rates for crevasses reached 90%, 92%, 89%, and 93%, the minimum difference between the model detected width of crevasses and the manually sampled width was only 0.3 m, this fully demonstrate that the model exhibits good robustness and generalization ability when applied in different glacier regions. The BDCN edge detector is considered more suitable for the detection and identification of ice crevasses, which provides a fast and effective technical support for the periodic monitoring of ice crevasses in the Geladandong glaciers and simi-lar regions.