Automated Mapping of Glacial Lakes in the Poiqu River Basin and Bhutan Region from SAR and Optical Images Using Deep Learning

Abstract:

Glacial lakes are rapidly developing and expanding under climatic warming, with the potential risk of glacier lake outburst floods (GLOFs) increasing particularly in the Himalayas. Considering the inaccessibility of high mountain area, accurate and efficient extraction of glacial lakes from remote sensing images is the basis for monitoring the lake evolution and detecting potential danger to the surroundings. Previous glacial lake mapping methods, which are manually designed for optical images or synthetic aperture radar (SAR) images, require prior knowledge and complicated pre- and post- data processing procedures. In this study, we automatically mapped glacial lakes using a deep learning (DL) technique based on Landsat-8, Sentinel-2 optical images, and Sentinel-1 SAR images. DeepLabv3+, an advanced semantic segmentation algorithm, was applied to delineating glacial lakes in the Poiqu River basin, central Himalaya. Twenty k-fold (k=5) cross-validation experiments were implemented for each dataset to test the robustness of this DL method. Moreover, a series of experiments on images obtained at a different time and a different region were also conducted to verify the transferability of our method. The mapping results were validated based on precision, recall and F1 scores. The average F1 scores when mapping the trained images are 0.82, 0.91, and 0.76 for Landsat-8, Sentinel-2, and Sentinel-1 imagery, respectively. The F1 scores of the test on images obtained on the different dates are 0.76 (Landsat-8), 0.52 (Sentinel-2) and 0.70 (Sentinel-1). In addition, a test was performed based on images acquired in the Bhutan Himalaya, with the F1 score of 0.40 (Landsat-8), 0.64 (Sentinel-2), and 0.55 (Sentinel-1). The lake extraction accuracy can be further improved with the high-quality input images and minimal post-processing with the assistance of topographic contributes (e.g., slope angles) from a digital elevation model. The results demonstrated successful applicability of our approach in identifying glacial lakes from optical and SAR imagery and it can be potentially transferred to multi-temporal images taken in other regions. This study increases the temporal frequency and diversity of glacial lake datasets, which provides important data support for monitoring glacial lakes and assessing GLOFs risks.