Detecting Heterogeneous Lakes Using Contrastive Representation From Remote Sensing Imagery

Lakes are among the most frequently occurring features observed in remote sensing imagery on Earth, which construct the part of water resource and supplying a great deal of life downstream. To assess water coverage and its regional or global impacts, identifying lakes in remote sensing imagery is a prerequisite as well in such lake-related studies. However, the diverse environmental conditions under which lakes form, the presence of mixed sediments leads to considerable spatial and spectral heterogeneity, making it challenging for existing methods to accurately detect heterogeneous water bodies. In this work, we proposed a weakly supervised contrastive representation method for finding the inherent universal natures of lakes with sediments. Our model comprised two modules, spatial segmentation module and spectral contrastive module for capturing the spatial and spectral features of lakes jointly. Then the patterns of lakes were compared with similarities to unlabeled samples and final lake area was extracted by decision-making process. We tested our model on Landsat-8 OLI imagery, and our model yielded $F1$ score of $79.86~\pm ~0.94$ and IoU of $70.85~\pm ~0.91$ , which showed great superiority to other competitors, especially in these lakes with sediments.