Deep Learning-Based Automatic Boundary Extraction Algorithm for Glacier Remote Sensing Data

Glaciers are closely related to climate change, and effective monitoring of glacier changes can provide key technical and data support for important fields such as natural environment protection and climate change response. The study is conducted research on automatic boundary extraction algorithms for glacier remote sensing data based on deep learning, which by used Landsat satellite remote sensing images of glaciers in the Geladandong and Amumachin Mountains of the Qinghai Tibet Plateau from 2000 to 2023 as the basic data to construct a glacier semantic segmentation dataset. In response to the physical characteristics of glaciers, the DualDeepLabV3+model is proposed innovatively. It is a dual branch backbone network that combines the advantages of ResNet and Swin Transformer based on the DeepLabV3+model is adopt. In order to verify the advantages and reliability of DualDeepLabV3+in automatic extraction of glacier boundaries, the study selected five mainstream semantic segmentation models, U-Net, PSPNet, PIDNet, DeepLabV3+-ResNet18, and DeepLabV3+_Swin Transformer, as comparisons. Six indicators, including pixel accuracy (PA), kappa coefficient (Kappa), intersection over union (IoU), precision, recall, and F1 score (F1 score), were used for comparative experiments. The results indicate that the DualDeepLabV3+model is significantly better than the comparative model in five key indicators: PA (0.9903)、Kappa(0.9759)、IoU(0.9657)、Precision(0.9886)、 Recall(0.9766)With an F1 score of 0.9825. And its performance is the best in automatic extraction of glacier boundaries. Based on the algorithm for automatically extracting glacier boundaries, the study further calculated the glacier area and verified the applicability and effectiveness of the model in glacier area calculation scenarios.