Abstract: The identification of glacial lakes is a prerequisite for understanding their response to climate change and assessing potential risks of glacial lake outburst floods (GLOFs). Although remote sensing technology enables continuous monitoring and assessment of global glacial lake evolution, accurately and reliably extracting glacial lakes in complex plateau terrain regions remains challenging. This study proposes an intelligent glacial lake identification method for complex plateau terrain based on multisource remote sensing data and an improved Mask R-CNN deep learning model. Building upon the original Mask R-CNN framework, we introduce attention mechanisms at three key components: the high-level features (Conv4 and Conv5) of the ResNet-50 backbone network, each feature map in the Feature Pyramid Network (FPN), and the Mask Head. Utilizing a multi-band dataset composed of Sentinel-2 highresolution imagery, ALOS-DEM, and Normalized Difference Water Index (NDWI) data, we conducted tests in Nyingchi City, southeastern Tibetan Plateau. Comparative analyses were performed between the enhanced Mask R-CNN model and three other models (U-Net, SegNet, and DeepLab V3) for glacial lake identification. Results demonstrate that the improved Mask R-CNN achieves superior accuracy, with precision, recall, and accuracy values reaching 91.25%, 93.69%, and 92.89% respectively. The enhanced model effectively mitigates interference from mountain shadows, lake turbidity, and freeze-thaw conditions on glacial lake identification while significantly improving detection efficiency for small glacial lakes. This research provides a reliable solution for glacial lake identification in complex plateau terrain regions and establishes a novel framework combining deep learning with multi-source remote sensing data for intelligent glacial lake extraction, offering new possibilities for related studies.