Accurate and automatic mapping of complex debris‐covered glacier from remote sensing imagery using deep convolutional networks

Abstract

Debris‐covered glacier mapping for monitoring glacier fluctuations is necessary to prevent geohazards caused by glaciers. Recently, deep learning‐based methods have been widely utilized for the identification of debris‐covered glaciers. Compared with conventional geospatial methods, deep learning‐based approaches have the advantage of large‐scale coverage and outstanding accuracies in identifying glaciers. However, there are two main difficulties when using deep learning‐based approaches: (1) object misclassification, which causes the misclassifications of the surrounding bedrock, snow, and mountain‐shadowed areas as glaciers; and (2) the inaccurate segmentation of boundaries. In addition, the sample sets for training deep learning models are generally insufficient, which leads to unsatisfactory results. To address the above problems, a deep learning‐based approach is proposed for accurate and automatic mapping of the complex debris‐covered glacier from remote sensing imagery for glacial hazard prevention. First, high‐quality remote sensing imagery is acquired and pre‐processed. Second, we adopt a weight‐optimized glacier semantic segmentation model to our approach. Then, the post‐processing procedures can be used to obtain the glacier outline. Finally, an accurate and automatic mapping of complex debris‐covered glaciers can be achieved. To demonstrate the effectiveness, our approach is applied to observe spatiotemporal changes in the glacier outline in the Nyenchen Tanglha region. Results show that most evaluation metrics of our model are higher than 90%, which demonstrates that it is reasonable for the glacier boundary extraction. We also verify that the proposed approach can be used to observe the changes in glaciers to prevent and control glacial hazards in high‐mountain regions.