Methodology:

We used the Normalised Difference Water Index adapted for ice (NDWI; Yang and Smith 2013) to classify pixels as 'water' or 'non-water' (i.e. ice, snow, exposed bedrock or ocean) using the red and blue bands. All processing was conducted in Arcmap 10.5.

We applied a physically-based radiative transfer model to calculate the water depth of all pixels classified as lake (Pope et al., 2016; Sneed and Hamilton 2007). This method calculates lake water depth (z) using the rate of light attenuation in water, lake bottom albedo, and optically-deep water reflectance (Philpot, 1989).

A detailed description of the data collection, quality control, processing and analysis is given in: Arthur, J.F, Stokes, C.R., Jamieson, S.S.R, Carr, J.R, Leeson, A.A (2020) Distribution and seasonal evolution of supraglacial lakes on Shackleton Ice Shelf, East Antarctica. doi: 10.5194/tc-2020-101.

Data quality:

We used a minimum size threshold of two pixels, in order to remove very small SGLs likely comprised solely of mixed pixels (i.e. 200 m2 for Sentinel 2, 450 m2 for Landsat 7/8, 1800 m2 for Landsat 4/5, 7200 m2 for Landsat 1), following previous studies (Moussavi et al., 2020; Pope et al., 2016, Stokes et al., 2019).

We quantified the uncertainties in our mapping technique by quantitatively comparing SGL areas derived from NDWI with those derived from manual digitisation in four sample areas. The smaller SGLs (<0.01 km2) generated the largest percentage differences, and we found a mean absolute error of 0.007 and a root mean square error of 0.029 between manually-digitised SGL areas and those derived from NDWI. We attribute this to manual digitisation being less conservative at 'diffuse', less well-defined SGL edges. Despite these differences, we found very close agreement between the two methods (R2 = 0.979, p = 0.006). We also quantified the impact of sensor resolution on lake detection and lake extents, and found there is generally a very good agreement between lake areas derived from Landsat 8 and Sentinel 2 (Supplementary Fig. 3). We therefore assign a conservative uncertainty of 1% to our total SGL area for each time step, following Stokes et al. (2019).

We checked final lake masks against RGB composites and manually removed any false positives, including small bedrock outcrops not included in the bedrock mask and dark crevasses. To further reduce noise and improve the visual clarity of the dataset, we then applied the dissolve function in ArcMap to clean up overlapping pixels and the aggregate tool to combine pixels within a distance of two pixels.

A detailed description of the data collection, quality control, processing and analysis, as well as full references, is given in:

Arthur, J.F, Stokes, C.R., Jamieson, S.S.R, Carr, J.R, Leeson, A.A (2020) Distribution and seasonal evolution of supraglacial lakes on Shackleton Ice Shelf, East Antarctica. doi: 10.5194/tc-2020-101.