GLA-STDeepLab: SAR Enhancing Glacier and Ice Shelf Front Detection Using Swin-TransDeepLab With Global–Local Attention
Qi Zhu, Huadong Guo, Lu Zhang, Dong Liang, Zherong Wu, Yiming Liu, Zhuoran Lv
Abstract
The understanding of glacier and ice shelf front changes is of paramount importance in analyzing their material balance and their significant contributions to global sea-level variations. Recently, the increasing availability of remote sensing images has provided ample data support for calving front detection. By combining massive amounts of data, deep learning techniques have demonstrated great potential in front extraction, as they can automatically generate task-specific features. In this study, in contrast to the traditional framework based on a pure CNN architecture, we propose a novel calving front detection network, GLA-STDeepLab, by incorporating the Swin transformer module into DeepLabv3+ to enhance the modeling capability of long-range contextual dependencies. We also introduce Global-local attention mechanisms into the module, striving to foster interactions between the corresponding coarse and fine-grained feature maps. Through extensive experimentation on a diverse and challenging SAR database, CaFFe, our network is shown to surpass the current state-of-the-art methods, achieving results with an IoU of 0.94 and a mean distance error (MDE) of 473±34 m. More meaningfully, we extracted the time-series front information of the Amery Ice Shelf from 2015 to 2023 and precisely captured the calving events of the D-28 and D-32 icebergs. The findings demonstrate that our method enables precise monitoring of the dynamic information of glacier and ice shelf fronts, as well as their far-reaching implications for ice sheet mass balance and global climate change. The code used in this study is provided at https://github.com/Tangyu35/Calving-front-detection.