Climate change is increasing the risk of glacial lake outburst floods (GLOFs) in many of the world’s most vulnerable and high mountain regions. Simultaneously, remote sensing technologies now facilitate continuous monitoring of glacial lake evolution around the globe, although accurate and reliable automated glacial lake mapping from satellite data remains challenging. In this study, a Second-order Attention Network (SoAN) is devised for the automated segmentation of lakes from satellite imagery. In particular, a novel Second-order Attention Module (SoAM) is proposed to capture the long-range spatial dependencies and establish channel attention derived from the covariance representations of local features. Furthermore, as the dimensions of the input and output tensors are identical and it simply relies on matrix calculations, the proposed SoAM can be embedded into different positions of a given architecture while maintaining similar reference speed. The designed network is implemented on Landsat-8 imagery and outputs are compared against representative deep learning models, demonstrating improved results with a Dice of 81.02% and a F2 Score of 85.17%.

在世界上许多最脆弱的高山地区，气候变化正在增加冰川湖溃决洪水 (GLOF) 的风险。同时，遥感技术现在促进了对全球冰川湖演变的持续监测，尽管从卫星数据中准确可靠地自动绘制冰川湖地图仍然具有挑战性。在这项研究中，设计了一个二阶注意力网络 (SoAN)，用于从卫星图像中自动分割湖泊。特别是，提出了一种新颖的二阶注意力模块（SoAM）来捕获远程空间依赖关系并建立从局部特征的协方差表示派生的通道注意力。此外，由于输入和输出张量的维度是相同的，并且它仅依赖于矩阵计算，所提出的 SoAM 可以嵌入到给定架构的不同位置，同时保持相似的参考速度。设计的网络是在 Landsat-8 图像上实现的，并将输出与代表性的深度学习模型进行比较，结果显示改进的结果为 81.02% 的 Dice 和 85.17% 的 F2 分数。