Evaluation of satellite methods for estimating supraglacial lake depth in southwest Greenland

Abstract

Supraglacial lakes form on the Greenland ice sheet in the melt season (May to October) when meltwater collects in surface depressions on the ice. Supraglacial lakes can act as a control on ice dynamics since, given a large enough volume of water and a favourable stress regime, hydrofracture of the lake can occur, which enables water transfer from the ice surface to the bedrock, where it can lubricate the base. The depth (and thus volume) of these lakes is typically estimated by applying a radiative transfer equation (RTE) to optical satellite imagery. This method can be used at scale across entire ice sheets but is poorly validated due to a paucity of in situ depth data. Here we intercompare supraglacial lake depth detection by means of ArcticDEM digital elevation models, ICESat-2 photon refraction, and the RTE applied to Sentinel-2 images across five lakes in southwest Greenland. We found good agreement between the ArcticDEM and ICESat-2 approaches (Pearson's r=0.98) but found that the RTE overestimates lake depth by up to 153 % using the green band (543–578 nm) and underestimates lake depth by up to 63 % using the red band (650–680 nm). Parametric uncertainty in the RTE estimates is substantial and is dominated by uncertainty in estimates of reflectance at the lakebed, which are derived empirically. Uncertainty in lake depth estimates translates into a poor understanding of total lake volume, which could mean that hydrofracture likelihood is poorly constrained, in turn affecting ice velocity predictions. Further laboratory studies to constrain spectral radiance loss in the water column and investigation of the potential effects of cryoconite on lakebed reflectance could improve the RTE in its current format. However, we also suggest that future work should explore multi-sensor approaches to deriving lake depth from optical satellite imagery, which may improve depth estimates and will certainly result in better-constrained uncertainties.