Mapping Debris-Covered Glaciers Using High-Resolution Imagery (GF-2) and Deep Learning Algorithms

Abstract

Glacier inventories are fundamental in understanding glacier dynamics and glacier-related environmental processes. High-resolution mapping of glacier outlines is lacking, although high-resolution satellite images have become available in recent decades. Challenges in development of glacier inventories have always included accurate delineation of boundaries of debris-covered glaciers, which is particularly true for high-resolution satellite images due to their limited spectral bands. To address this issue, we introduced an automated, high-precision method in this study for mapping debris-covered glaciers based on 1 m resolution Gaofen-2 (GF-2) imagery. By integrating GF-2 reflectance, topographic features, and land surface temperature (LST), we used an attention mechanism to improve the performance of several deep learning network models (the U-Net network, a fully convolutional neural network (FCNN), and DeepLabV3+). The trained models were then applied to map the outlines of debris-covered glaciers, at 1 m resolution, in the central Karakoram regions. The results indicated that the U-Net model enhanced with the Convolutional Block Attention Module (CBAM) outperforms other deep learning models (e.g., FCNN, DeepLabV3+, and U-Net model without CBAM) in terms of precision for supraglacial debris identification. On the testing dataset, the CBAM-enhanced U-Net model achieved notable performance metrics, with its accuracy, F1 score, mean intersection over union (MIoU), and kappa coefficient reaching 0.93, 0.74, 0.79, and 0.88. When applied at the regional scale, the model even exhibits heightened precision (accuracies = 0.94, F1 = 0.94, MIoU = 0.86, kappa = 0.91) in mapping debris-covered glaciers. The experimental glacier outlines were accurately extracted, enabling the distinction of supraglacial debris, clean ice, and other features on glaciers in central Karakoram using this trained model. The results for our method revealed differences of 0.14% for bare ice and 10.36% against the manually interpreted glacier boundary for supraglacial debris. Comparison with previous glacier inventories revealed raised precisions of 8.74% and 4.78% in extracting clean ice and with supraglacial debris, respectively. Additionally, our model demonstrates exceptionally high exclusion for bare rock outside glaciers and could reduce the influence of non-glacial snow on glacier delineation, showing substantial promise in mapping debris-covered glaciers.