Supraglacial lakes form annually in surface depressions around the periphery of Greenland. Here, a deep learning method based on U-Net was employed to segment supraglacial lakes in Northeast Greenland on Zachariæ Isstrøm and 79N Glacier over the 2016 to 2022 summer melt seasons (April to September). A deep learning-based cloud segmentation model specifically trained for polar regions was also developed and utilized to remove cloudy images from the supraglacial lake time series. While the deep learning lake segmentation model is able to overcome some of the downfalls of traditional thresholding methods such as false inclusion of shadows and inadaptability to new scenes and atmospheric conditions, it still encounters some problems with large, extensive shadows and with slushy light blue ice on peak melt days.

Each .zip folder contains shapefiles associated with the year in the name (yyyy.zip). The shapefiles are labeled with the date of the Sentinel-2 image acquisition (yyyy-mm-dd_pred_vector.shp) along with associated geospatial files (.cpg, .dbf, .prj, .shx). These shapefiles contain polygons representing supraglacial lake area. The days with cloud cover larger than 10% were removed from the time series. These shapefiles are direct outputs from the model and have not been manually modified. Thus, they could contain false positives, especially from topographic or cloud shadows.