Improving automatic glacier mapping accuracy, robustness, and reliability across the Third Pole Region

Most automatic glacier mapping methods applied in the Third Pole Region (TPR), exhibit suboptimal accuracy for debris-covered glaciers, high sensitivity to interferences, and large uncertainty in model outputs, thereby limiting their operational utility. These challenges can be mitigated by deploying advanced deep learning models and enhancing dataset representativeness across complex imaging conditions. Accordingly, we compiled three large-scale datasets used for model development, glacier-level mapping evaluation, and cross-sensor generalisation assessment, respectively. We then evaluated several state-of-the-art deep learning networks and found that the adapted TransUNet performed best. On the test dataset, the model achieves a glacier Intersection over Union (IoU) of 90% and 83% in patches featuring debris-free and debris-covered glaciers, respectively, while maintaining robust under challenging conditions. Glacier-level validation further confirms the effective detection of glacier counts and precise outlines, with both metrics improving with glacier area and the global IoUTPR > 80%. Uncertainty quantification shows high output reliability, with expected calibration errors of 0.029 and 0.009 for the test and glacier-level test datasets, respectively. Our fully automated 2024–2025 Karakoram glacier inventory, comprising 12,981 glaciers covering a total area of ~19,484 km2, exhibits high agreement with earlier inventories and reflects current glacier state. The primary contribution of this study is the development of a scalable framework that, by optimising dataset representativeness and algorithmic architectures, significantly elevates the accuracy, robustness, reliability and operational value of glacier mapping across the TPR.