Continuous change monitoring of supraglacial lakes during melt and non-melt seasons with multi-source satellite imagery and deep learning

The surface melting of ice sheets is gradually intensifying with rising global temperatures. Supraglacial lakes (SGLs), an essential component of the ice-surface hydrological system, are significant for the ice sheet mass balance. SGLs have been shown to persist year-round, with some lakes retaining as buried lakes within the ice sheet subsurface throughout the winter, rather than completely draining and freezing after the melt season. Accordingly, continuous change monitoring of SGLs during melt and non-melt seasons is crucial for understanding ice sheet storage and drainage processes. Herein, we used optical and SAR satellite data to monitor the SGLs variability in the southwest Greenland from 2020 to 2022. Three deep learning models, namely, U-Net, attention U-Net (AU-Net), and attention recurrent residual U-Net (AR2U-Net), are compared. AR2U-Net showed the optimal performance in SGL extraction in SAR images. However, SGLs are not precisely extracted from SAR images in areas with high water content due to the influence of wet snow, so optical imagery extraction results are referenced in the study of melting seasons. The results of the SGL changes from 2020 to 2022 indicated that, the changes in the SGL area between years are insignificant, and the main differences are attributed to extreme climatic events occurring in Greenland. Majority of the newly formed SGLs drained at the end of the melt season, with about 16.52% forming buried lakes, which are concentrated in the mid-high elevation. However, the storage of newly buried lakes is not sensitive to an intense melting season, with the new meltwater retained in comparable amounts at the end of the three melt seasons. Buried lakes that remained unopened throughout the year are typically found at elevated altitudes, and their areas is around 2–3 times that of the retained liquid water from the melt season. Overall, the extraction methods and analytical approaches used in this study, based on a longer time series and a wider study scope will aid the further discussion of the seasonal area changes in SGLs.