Numerous supraglacial lakes form on the Greenland Ice Sheet (GrIS) during the summer, and accurately estimating their depth is crucial for understanding GrIS water storage. In this study, we estimate the depth of 35 representative GrIS supraglacial lakes using ICESat-2, Sentinel-2 imagery and ArcticDEM data. ICESat-2-derived lake depth is used to validate the performance of three remote sensing methods, namely empirical formula method (EFM), radiative transfer method (RTM) and depression topography method (DTM). EFM relies on ICESat-2-derived lake depth to construct empirical formulas, while RTM and DTM do not. The results show that (1) the green band EFM performs best; the DTM performs secondarily but tends to consistently underestimate depths; the green-band RTM has lower accuracy and overestimates depths, while the red-band RTM also has lower accuracy but underestimates depths. (2) Temporal changes of depression topography have limited impacts on the performance of DTM, whereas the uncertainties caused by lake shoreline height estimates should be considered. (3) The performance of RTM is significantly influenced by the spectral attenuation coefficient. We further identify the factors that affect spatiotemporal extrapolation of these methods and recommend prioritizing the use of the EFM when near-simultaneous ICESat-2 data are available; otherwise, DTM is recommended, yet an underestimation ratio should be used.

Supraglacial lakes (SGLs) are widespread across the Greenland ice sheet and cause transient changes in ice flow. Here, we produce the first annual ice-sheet wide database of maximum summer SGL extents spanning 1985 to 2023 using all July and August Landsat images. Lake visibility percentages were calculated to estimate the uncertainty induced by variable image data coverage. SGLs were mainly distributed between 1000 and 1600 m elevation, with large lake area observed in northwestern, northeastern and southwestern basins. Lake area increased at a rate of 50.5 km2 a−1 across the entire Greenland, and lakes advanced to higher elevations at an average rate of 10.2 m a−1 during 1985–2023. We leveraged spatiotemporally matched ICESat-2 and Landsat 8 reflectance data to develop a deep learning model for lake depth inversion for the period 2014–23. This model demonstrates the highest accuracy among all image-based methods, albeit with an underestimation of ~15% when compared to ICESat-2 data. A significant positive correlation between lake volume and area is used to up-scale the approach to the entire time period, indicating a lake volume increase of 221.9 ± 63.6 × 106 m3 a−1. Increasing air/land surface temperature, surface pressure and decreasing snowfall were the most important contributing factors in driving lake variability.