Exploring Contrastive Representation for Weakly-Supervised Glacial Lake Extraction

Abstract

Against the background of the ongoing atmospheric warming, the glacial lakes that are nourished and expanded in High Mountain Asia pose growing risks of glacial lake outburst floods (GLOFs) hazards and increasing threats to the downstream areas. Effectively extracting the area and consistently monitoring the dynamics of these lakes are of great significance in predicting and preventing GLOF events. To automatically extract the lake areas, many deep learning (DL) methods capable of capturing the multi-level features of lakes have been proposed in segmentation and classification tasks. However, the portability of these supervised DL methods need to be improved in order to be directly applied to different data sources, as they require laborious effort to collect the labeled lake masks. In this work, we proposed a simple glacial lake extraction model (SimGL) via weakly-supervised contrastive learning to extend and improve the extraction performances in cases that lack the labeled lake masks. In SimGL, a Siamese network was employed to learn similar objects by maximizing the similarity between the input image and its augmentations. Then, a simple Normalized Difference Water Index (NDWI) map was provided as the location cue instead of the labeled lake masks to constrain the model to capture the representations related to the glacial lakes and the segmentations to coincide with the true lake areas. Finally, the experimental results of the glacial lake extraction on the 1540 Landsat-8 image patches showed that our approach, SimGL, offers a competitive effort with some supervised methods (such as Random Forest) and outperforms other unsupervised image segmentation methods in cases that lack true image labels.