Every Landsat 7 and 8 image covering any portion of our study region between 2005 and 2020 was used during analysis, totalling 4164 optical image tiles. In practice this resulted in Landsat 8 images being exclusively used beyond March 2013, with Landsat 7 images used prior to this date. Images were not filtered by cloud cover to maximise the chances of detecting surface water. We used Landsat Level-1 Tier-2 top-of-atmosphere (TOA) image tiles, which are directly available for analysis through the GEE data catalogue (https://developers.google.com/earth-engine/datasets/catalog/landsat, last access: 31 March 2021). TOA reflectance values are typically used for ice sheet studies in preference to raw digital numbers to ensure that pixel values are not influenced by differences in image acquisition conditions (Pope et al., 2016; Moussavi et al., 2020). Processing was performed on a yearly basis, involving 16 runs of the GEE script (i.e. 2005–2020). For each GEE run, an image collection was generated from images that fit the criteria of the specified time period and overlapped with the ROI. Images were additionally filtered to remove those with a sun elevation angle of less than 20∘. Images with a sun elevation lower than this threshold value result in misclassification errors when using a band-threshold-based approach, since in low-light conditions surface water is not sufficiently spectrally different to be separated from features such as cloud and rock shadow (Halberstadt et al., 2020; Moussavi et al., 2020).

We applied a surface meltwater detection method developed by Moussavi et al. (2020), who established threshold values to automatically identify surface water, cloud and rocks from Landsat 8 image bands (Fig. S1 in the Supplement). The thresholds used in Moussavi et al. (2020) showed an accuracy of > 95 % when identifying lake areas from Landsat 8 imagery, and results showed high levels of agreement when compared with lake area data generated from other methods (Halberstadt et al., 2020). Whilst the thresholds developed by Moussavi et al. (2020) were designed specifically for Landsat 8, we found that the thresholds are highly successful when applied to Landsat 7 imagery, despite minor differences in the band wavelengths of the two satellites. Our analysis shows that there is an average agreement of ∼ 90 % between Landsat 7 and Landsat 8 in the identification of surface water (see Figs. S2–S4 in the Supplement for a comparison between Landsat 7 and Landsat 8).

As per the method of Moussavi et al. (2020), areas of exposed bedrock and seawater were removed from image tiles using a mask based on the thermal infrared (TIR) and blue bands. Cloudy pixels were removed using a combination of the short-wave infrared (SWIR) band and the difference snow index (green - SWIR / green + SWIR). Following application of these masks (Fig. 2), we then used an ice-specific version of the normalised difference water index (NDWIice, blue - red / blue + red) to delineate areas of surface water. This is the most widely used technique for identifying water from optical imagery (Williamson et al., 2018; Arthur et al., 2020b) and has been successfully used to map SGLs on both the Greenland (Pope et al., 2016; Moussavi et al., 2016; Williamson et al., 2018) and Antarctic ice sheets (Stokes et al., 2019; Moussavi et al., 2020). See Fig. S1 in the Supplement for the threshold values used and Moussavi et al. (2020) for further details of the method. Once lake pixels were detected in each individual image tile, images were assigned to a time window (Fig. 2). Lake masks from individual images within each time window were then combined to create a single maximal lake mask for each time window.

For images affected by cloud cover, mapped lakes from optical satellite data represent minimum estimates of true lake area. Though simple metrics of cloud cover per image are informative, they do not account for variability in meltwater extent and visibility within a time window. To account for the uncertainty in lake area due to these visibility issues, we developed a novel technique which estimates the potential maximum lake area likely if clouds were not present. To evaluate meltwater visibility over the duration of each time window, we therefore needed to assess two key aspects: (i) a spatial assessment of the amount of ice visible within the intersection of each optical satellite image and each ROI, achieved by calculating an “image visibility score” (IVS) for every optical image (Fig. 3), and (ii) a temporal assessment of the differences in meltwater extent between images within each time window. This second stage was achieved by calculating a “lake pixel contribution score” (LPCS) for images within each time window (Fig. 3), enabling quantification of which images within any given time window contributed the most lake pixels to the overall output. These two metrics were then combined to estimate a “lake visibility percentage” (LVP) for each time window and ROI (Fig. 3).

Mapped lake polygons and visibility statistics were exported as geoJSON files from GEE. Several post-processing stages were then undertaken in MATLAB to convert the data into shapefiles, merge lake polygons between ROIs and attach metadata. Shapefiles were firstly created (using the Antarctic polar stereographic projection) for every ROI tile and time window. ROI-specific shapefiles were then merged across the entire study region, to create one single dataset per time window. As part of this step, lakes split over ROI boundaries were joined together (Union), and inner polygons were “cut” from outer lake boundaries in instances where an “island” (typically an ice lid) was present within a lake. We then calculated the area and geometric centroid of each cleaned polygon and applied an area threshold of two pixels, giving minimum lake areas of 1800 m2 based on a Landsat resolution of 30 m. This filtered out noise from the raw output, likely associated with crevasse shadows or slush, whilst retaining enough data to include small lakes, especially those at high elevations that would have been missed with a higher area threshold value. Unlike some other studies (e.g. Stokes et al., 2019), we decided not to aggregate lake polygons in close proximity to each other, as tests showed this sometimes resulted in the false identification of large lakes in areas of meltwater-filled crevasses. Finally, we attached selected metadata to each identified lake based on the geometric centroid of lake polygons. The Depoorter et al. (2013) grounding line dataset was used to label lakes as either “grounded” or “floating”, whilst the elevation and surface slope of lake centroids were extracted from the Reference Elevation Model of Antarctica (REMA) database (100 m resolution) (Howat et al., 2019). All post-processing steps were automated in MATLAB, with each melt season taking approximately 2–5 h to run.

To provide an initial test of the extent to which climatic modelling can simulate surface meltwater ponding, we compared our lake area results with modelled snowmelt outputs from the Regional Atmospheric Climate Model version 2.3p2 (RACMO2.3p2) (van Wessem et al., 2018). RACMO2.3p2 has a horizontal resolution of 27 km and is coupled to an internal snow model which calculates surface melt production, refreezing, percolation, retention and runoff into the ocean. The model is forced by ERA-Interim (∼ 80 km horizontal resolution) reanalysis data (van Wessem et al., 2018). Monthly RACMO2.3p2 melt values were summed across the study region and then divided by the total number of pixels to provide monthly mean melt values. RACMO2.3p2 snowmelt outputs serve as an upper bound for meltwater availability, as the model does not specifically account for surface meltwater ponding. Moreover, it should be noted that RACMO2.3p2 locally resolves meltwater production based on model grid boxes and hence does not account for the process of meltwater flowing from higher elevations (Spergel et al., 2021). Our analysis therefore offers a preliminary comparison between the two datasets rather than a full evaluation, which would require quantification of lateral meltwater transfer and biases highlighted in van Wessem et al. (2018). Given the catchment scale of this study, the lack of lateral meltwater transport is of less importance than for smaller-scale studies (e.g. Spergel et al., 2021).

To explore the potential role of large-scale atmospheric circulation in surface meltwater ponding in the study region, we investigated the influence of the Southern Annular Mode (SAM). The SAM is the main mode of extratropical climate variability across the Southern Hemisphere and represents changes in the strength and position of the Southern Hemisphere westerly winds and storm tracks (Marshall and Thompson, 2016). We chose to compare our lake area results with the SAM because of its known influence on Antarctic temperatures (Marshall and Thompson, 2016; Fogt and Marshall, 2020), and hence surface melting. We compared our results with austral summer values of the SAM index of Marshall (2003), obtained from http://www.nerc-bas.ac.uk/public/icd/gjma/newsam.1957.2007.seas.txt (last access: 31 March 2021).

As shown by Moussavi et al. (2020), we find that the application of a band-thresholding technique within GEE is highly successful at rapidly identifying surface meltwater features over large areas and time periods. The thresholds applied were effective at masking out areas of rock and cloud over the whole study area, whilst successfully identifying surface meltwater (Fig. 5). Manual checking of mapped lakes against satellite imagery (from approximately ∼ 10 % of randomly selected time windows) identified very few false positives, and the technique performed well when differentiating lakes from areas of blue ice and shadow (Fig. 5). This is consistent with the findings of Moussavi et al. (2020), who used the same thresholds and found overall accuracies of > 95 % when mapping from Landsat-8 imagery. There was no particular spatial pattern to false positives, such as clustering around bedrock or shadow areas. False negative results were rare and mainly occurred where surface water was much darker in colour, presumably either due to sediment suspended within the water column or where lakes appeared to be very deep. Instances of sediment-laden water were confined to the immediate vicinity of rock outcrops, whilst lake depths very rarely exceed 4 m in the study region (Spergel et al., 2021). These misclassification errors thus had a minimal influence on results. As highlighted in Fig. 5, we found minimal difference in the performance of the method between Landsat 7 and 8 imagery (Fig. S3 in the Supplement).

LVPs ranged from 0 %–99.9 %, with a mean LVP of 50.4 % and a median LVP of 52.7 % across the whole dataset. However, there were large differences between LVPs from Landsat 7 and Landsat 8 images, largely due to data gaps present within Landsat 7 images as a result of the failure of the scan line corrector (SLC). The median LVP from time windows using Landsat 7 imagery was 43.5 %, compared to 61.6 % when Landsat 8 images were used. By using LVPs to generate maximum lake area estimates, we were able to account for lake area underestimations resulting from data gaps in Landsat 7 imagery. On average, incorporating LVPs into lake area estimates resulted in a 58 % increase in lake area per ROI and time window when using Landsat 7 and a 42 % increase when using Landsat 8 images. When results were aggregated to generate cumulative lake area estimates per melt season, maximum potential lake area estimates were 42 % greater than mapped values on average across the entire study period.

We find that SGLs form on inland areas of the AIS where the ice shelf is narrowest and on portions of grounded ice within close proximity to the grounding zone (Fig. 6). On average, ∼ 70 % of total lake area within the study region exists on the ice shelf and ∼ 30 % on grounded ice. In high-melt years, SGLs are widespread across the width of the ice shelf between ∼ 72–73∘ S and along the Prince Charles Mountains side of the ice shelf to around 71∘ S. Very few lakes form on the ice shelf interior further north than this latitude, although a cluster of lakes sometimes form in a sub-inlet of the ice shelf near the Prince Charles Mountains (Fig. 6). Lakes on the ice shelf most frequently form on the southeast side of the Clemence Massif and on the eastern side of the Fisher Massif (Fig. 6). SGLs in these locations are typically elongate in shape and are connected by surface streams and channels to form a distributed surface drainage network. During high-melt years, the largest lakes are found along the central flow line of the ice shelf below 71∘ S; the largest mapped lake in our study had an area of 107 km2 in January 2005. However, these central lakes vary greatly in size and occurrence between melt seasons, whilst lakes nearer the grounding zone and next to areas of exposed bedrock form more frequently (Fig. 6).

SGLs on grounded ice predominantly form within approximately 20 km of the grounding zone and are particularly abundant along a 200 km stretch of the Princess Elizabeth Land ice shelf boundary between 70–72∘ S (Fig. 6b). Lakes in this region, which can be up to 6 km2 in area, typically form in the same location on an annual basis. Whilst the spatial extent of lakes varies between years, we noted several lakes in this region that formed in the same location during all 14 of the complete melt seasons studied (Fig. 6b). No large lakes form on the three main glaciers which feed the southernmost portion of the ice shelf, but extensive areas of meltwater-filled crevasses are often observed on Lambert Glacier.

Surface meltwater is found up to elevations of ∼ 1500 m, with the highest confirmed lake (with a minimum area threshold of 1800 m2) existing at 1591 m above sea level (m a.s.l.). Lakes are most common at low elevations, with the greatest lake area totals identified between 100–200 m a.s.l. This is the elevation band that covers the majority of the southern part of the ice shelf. The majority of the northern half of the ice shelf lies below 100 m a.s.l., but there is low runoff and ponding in this region (Fig. 6). Average lake size decreases with an increase in elevation, with the majority of surface meltwater above ∼ 600 m a.s.l. existing in the form of small, isolated ponds within crevasse fields (mostly on Lambert Glacier). However, larger SGLs (up to ∼ 5 km2 in area) are common at elevations up to 500 m a.s.l. on sections of grounded ice in Princess Elizabeth Land. Lake areas are greatest between 100 and 200 m a.s.l. during all 5 months of the melt season (Fig. 7), regardless of annual variations in absolute melt supply. We do, however, notice slight differences in the distribution of lake area across elevation bands between high- and low-melt years. During low-melt years, total lake area is more evenly distributed across elevations ranging between 100–400 m a.s.l. (Fig. 7b), whereas in high-melt years, lake surface areas are more concentrated between 100–200 m a.s.l. (Fig. 7a).

The seasonal and multi-year evolution of lakes for the Amery region is shown in Fig. 8. The highest cumulative number of lakes was observed during the 2016/17 melt season, during which the cumulative total number of lakes exceeded 100 000 (Fig. 8a). By contrast, fewer than 30 000 lakes were cumulatively observed during both the 2010/11 and 2011/12 melt seasons. Lake numbers were relatively low between 2006 and 2013; cumulative seasonal lake numbers remained below 50 000 for every melt season during this period, whereas five out of the six subsequent melt seasons had seasonal cumulative totals of more than 75 000 lakes.

The highest lake area totals during an individual time window were identified during the 2004/05 and 2005/06 melt seasons (Fig. 8b). During the first half of January 2005, surface meltwater covered an estimated maximum total area of 2814 km2. This was almost 3 times greater than the average total lake area for the first half of January (963 km2 for maximum estimates) throughout the study period. As observed with lake numbers, the 7-year period between late 2006 and early 2013 was characterised by low lake area coverage (Fig. 8b). The average estimated cumulative lake area per season during this time period was 1062 km2. This was around 3 times lower than the equivalent average of 2997 km2 between 2014 and 2020 (excluding 2018/19 due to incomplete data availability), despite the 2015/16 melt season having very low areas of lake coverage.

Although there is high variability in both the number and total areas of lakes observed between melt seasons, we do not observe an overall increasing or decreasing trend. A strong correlation (r= 0.81, p= 2.1 × 10-32) is observed between lake numbers and total lake area for individual time windows. In addition to having the highest number of lakes, the 2016/17 season also had the highest cumulative lake area, with an estimated (based on lake visibility corrected scores) maximum lake area total of 5179 km2. High lake area totals were recorded during the 2005/06 season, despite only having the sixth highest number of lakes.

Clear seasonal patterns of lake numbers and areas can be observed within each melt season (Figs. 8 and S8). Between October and early December, total lake areas were typically very low, with any meltwater forming in crevasses or pooling in small depressions close to exposed bedrock. For all studied years, there was a sharp increase in total lake area during the second half of December, including in melt seasons when absolute lake area was relatively low. On average, total lake area increased by an order of magnitude during this time window compared to the first half of December. Lake area coverage typically continued to increase into the first half of January, when maximum lake areas for the melt season were most commonly observed. Peak lake area totals were experienced during the first half of January on 8 out of the 14 occasions for which data were generated throughout the entire melt season (Table 2). In low-melt years, it was more common for lake areas to peak later in the melt season, usually during the second half of January and on one occasion (2009/10) during the first half of February. In most years, total lake area decreased through late January and early February, and by the second half of February, most lakes had frozen over. The average estimated total lake area for late February was 97 km2, compared with 348 km2 during the first half of the month. Despite these seasonal trends in total lake area, we did not observe a shift in meltwater cover to higher elevations throughout each melt season (Fig. 7).

We compared our lake area results with monthly surface snowmelt rates from RACMO2.3p2 to investigate the relationship between observed and modelled results. There is strong positive correlation between the seasonal totals of the two datasets (r= 0.76, p= 0.002), showing that the RACMO model captures the temporal variations in melting indicated by lake observations reasonably well (Fig. 9). The two melt seasons with the highest cumulative total lake area (2016/17 and 2005/06) also had the highest mean seasonal snowmelt estimates. However, the mean seasonal melt total for 2005/06 was 23.7 mm w.e. greater than the 2016/17 estimate, despite displaying very similar cumulative lake areas. The biggest discrepancy between the two datasets was in 2014/15 when modelled melt rates were low, whereas the cumulative lake area was the third highest throughout the study period.

Figure 8c reveals minor inter-annual variations in both the spread and the maximum estimates of modelled melt rates. Mean monthly RACMO melt was highest during December in most of the study years, but peak melt was modelled to have occurred during January in six melt seasons. In years when maximum melt was modelled to have occurred during December, total lake area typically (75 % of the time) peaked during the first half of January, indicating a lag between peak melt and peak lake storage of ∼ 15–30 d. Similar lag times were observed in years when modelled melt values were highest in January, with total lake area in these years most commonly peaking in either the second half of January or early February (Table 2). The duration of high (> 30 mm w.e.) melt rates also varied between years. In 2005/06, high melt rates were experienced over a single month (December), whilst remaining very low during other months of the melt season. This matches well with the lake area data for that year, where a sharp increase in total lake area was observed between mid-December and mid-January, before rapidly dropping again by the end of January. In some years, maximum melt rates were sustained over both December and January, although absolute values of melt rate were usually lower in these years. In 2012/13, for example, the maximum monthly melt estimate was 21.0 mm w.e., but because this level of relatively low melt was sustained over a period of 2 months, mean seasonal melt was the fourth highest during the study period (Fig. 9).

To investigate the extent to which large-scale variability in Antarctic climate influences surface meltwater area, we correlated our lake area results against the SAM index (Fig. 10). We find that there is a significant negative correlation (r=-0.54, p= 0.029) between total lake area and the SAM index for austral summer months. Melt seasons with a negative summer SAM index correlated with years when total lake areas were greatest, whilst years with a positive summer SAM index were associated with low total lake areas. The SAM index was below -1 on two occasions throughout the study period (2005/06 and 2016/17), the same two years that we observed the greatest cumulative lake areas (excluding the 2004/05 melt season where data were only available during the second half of the melt season). Years with a positive SAM index of 2 or more were characterised by low surface meltwater cover, with the notable exception of the 2014/15 season. This melt season was associated with the highest SAM index of the whole study period, yet had the fourth highest cumulative lake area total.

In this paper, we have overcome two key factors which previously restricted the generation of robust high-resolution time series of SGL extent from optical satellite imagery. First, by incorporating a threshold-based method for lake detection within GEE, with results generated by time windows, we have created a fully automated method for generating lake area time series that is quick and simple to run. The majority of SGL mapping studies in Antarctica have been limited in spatial and/or temporal resolution, partly due to methodological constraints relating to the computational expense of processing large imagery datasets. Despite having a relatively high spatial density of SGLs compared to most other areas of Antarctica (hence reducing the speed of processing within GEE), we were able to map an area of > 185 000 km2 over a 15-year time period in less than a week of wall-clock time. This rapid processing opens up the possibility of future studies to investigate surface meltwater evolution over vastly increased spatial and temporal scales, compared to what would be possible using manual or semi-automated methods. The method requires minimal inputs and user intervention (file transfers are required between the GEE and MATLAB automated stages), meaning it can be quickly adapted to generate lake area time series for other regions of Antarctica, and ultimately a pan-ice sheet study. By using a time window approach whereby the length of time windows can be varied (e.g. daily, monthly or yearly mapping), the method could be used to investigate surface meltwater processes at a range of temporal resolutions (depending on image availability). Whilst it is computationally simple to scale up the method to map at a continent-wide scale, it should be noted that the band reflectance thresholds may need adjusting when mapping certain regions of Antarctica. Moussavi et al. (2020) established the thresholds applied here based on spectral analysis of four ice shelves around Antarctica, covering a wide range of surface conditions and ponding characteristics. However, ice shelves with large regions of dirty ice or high debris content, such as the McMurdo Ice Shelf, are more likely to result in misclassification errors, meaning new thresholds may need to be established in such locations.

Second, our SGL mapping procedure incorporates a robust new method for assessing image visibility, enabling us to account for variability in cloud cover and image data coverage when generating time series. Whilst multiple studies have provided Antarctic SGL area and volume estimates from optical mapping (Arthur et al., 2020a; Dell et al., 2020; Moussavi et al., 2020), accounting for low image visibility from cloud cover has remained the primary limiting factor in creating a continuous and consistent time series (Moussavi et al., 2020). Furthermore, reported SGL areas and volumes based on optical mapping likely underestimate ground-truth meltwater extent, since very few optical images are entirely cloud-free. Here, we performed image visibility assessments on every image analysed, enabling us to quantify levels of uncertainty for lake area results. Maximum lake area estimates, which incorporated visibility assessments, increased mapped lake areas for time windows on average by approximately 50 %. This highlights the importance of accounting for image visibility when reporting lake area results, especially when working with Landsat 7 imagery (due to the SLC failure) or mapping frequently cloud-covered regions, such as the Antarctic Peninsula (van Wessem et al., 2016). Our method assumes that lakes have an equal chance of occurring across ice-covered areas of an ROI. In reality, lakes are often spatially clustered and occur in similar locations between years. This uneven spatial distribution is a potential source of error for our maximum lake area estimates. The sign and size of this error will be dependent upon the degree of lake clustering and the position of clustered lakes relative to cloud cover within each ROI for each time window.

Surface lakes are often widespread on inland sections of the ice shelf during austral summer months, whilst almost no SGLs form on the northern half of the ice shelf closer to the ocean. The spatial distribution of surface lakes on the AIS is strongly influenced by variations in firn air content across the study area, as similarly observed across other ice shelves in Antarctica (Lenaerts et al., 2017; Arthur et al., 2020a; Dell et al., 2020). The lack of surface meltwater ponding in the northern half of the study region (Fig. 6) is likely a consequence of high rates of snow accumulation near the calving front (Budd, 1966). A thick snowpack near the ice front has large pore spaces within the firn layer, meaning surface meltwater can percolate downwards and be accommodated within the pore spaces (Bell et al., 2018). By contrast, low accumulation rates further inland on the ice shelf likely result in a lower firn air content, meaning the firn layer becomes saturated with meltwater more quickly, causing ponding of surface water (Bell et al., 2018; Arthur et al., 2020a). Cycles of melting and refreezing increase the grain-size of particles within the firn layer, reducing the albedo of the surface compared to fine-grained fresh snow (Zwally and Fiegles, 1994; Phillips, 1998). This can induce a positive feedback whereby previously melted areas are more likely to experience further melting, due to the increased absorption of short-wave radiation associated with low-albedo surfaces (Kingslake et al., 2017). It is possible that this feedback is further enhanced by the presence of ice slabs and lenses, which can from beneath areas of intermittent pond formation (Hubbard et al., 2016). These dense layers of ice inhibit meltwater percolation and can be several degrees warmer than ice that has not undergone lateral heat fluctuations that result from the melting and refreezing of ice (Hubbard et al., 2016). Such ice slabs have been shown to have important implications for lake development over multiple melt seasons, based on modelling of the Larsen C ice shelf (Buzzard et al., 2018).

The clustering of surface lakes around the grounding line at southern latitudes of the AIS can further be explained by the influence of katabatic winds. Near-surface air temperatures in coastal regions of East Antarctica are strongly influenced by katabatic winds which originate from the ice sheet's interior (Lenaerts et al., 2017). These winds, which are commonly strong and directionally persistent (Lenaerts et al., 2017), generate localised surface and atmospheric conditions that are conducive to surface melting. Katabatic winds warm adiabatically as they flow down surface slopes, disrupting the natural temperature inversion and resulting in warmer, more humid air adjacent to the ice surface at the break in slope of the grounding zone (Doran et al., 1996). These atmospheric conditions, combined with the occurrence of low surface slopes on the ice shelf, optimise the local environment for meltwater ponding, resulting in SGL formation around the grounding zone of Antarctic ice shelves (Arthur et al., 2020b; Elvidge et al., 2020). Particularly high numbers of lakes are observed on the narrowest part of the AIS, as this is likely the focal point for katabatic winds that are channelised, and hence strengthened, down Lambert, Fisher and Mellor glaciers (Zwally and Fiegles, 1994). Furthermore, increased numbers of flow stripes in this narrow section of the ice shelf provide greater surface roughness within which lakes can form (Ng et al., 2018). Our results show that lakes form at lower latitudes along the Prince Charles Mountains side of the ice shelf compared to the Princess Elizabeth Land margin (Fig. 6). We suggest this is because katabatic winds continue to be channelised by the Mawson Escarpment once on the ice shelf, causing them to naturally flow out along the western margin of the ice shelf. Once the ice shelf widens and is no longer as confined by topography, the winds likely weaken in strength, thus negating the localised warming effect and limiting lake growth.

Strong katabatic winds can also erode the surface snow layer within which melt could be stored, exposing highly compacted, less permeable surfaces. Continued wind scouring around the grounding zone can expose areas of blue ice, which have a lower albedo (∼ 0.57) than refrozen snow (∼ 0.7) (Lenaerts et al., 2017). The presence of blue ice, in addition to the high number of low-albedo nunataks that surround the inland portion of the AIS, increases net surface absorption of solar energy, providing a localised warming effect and enhancing surface melt rates (Kingslake et al., 2017). Surface melt rates on other ice shelves in Antarctica, such as Roi Baudouin and Shackleton, have been shown to be strongly controlled by melt–albedo feedbacks (Lenaerts et al., 2017; Jakobs et al., 2019; Arthur et al., 2020a; Dell et al., 2020). Our results support these findings, as we observe a clear spatial association between low-albedo surfaces and areas of high lake occurrence, such as the large number of lakes that form annually next to the Prince Charles Mountains (Fig. 6). The spatial distribution of surface meltwater in the study region is hence closely controlled by melt–albedo coupling between exposed bedrock, blue ice and surface melting (Kingslake et al., 2017).

On both grounded and floating sections of the study region, lakes typically form in the same location on an annual basis (Fig. 6). Surface topography controls the hydrological routing of surface water, resulting in the ponding of water in small hollows and basins (Bell et al., 2018). Longitudinal surface structures on the ice shelf surface, caused by lateral compression and longitudinal extension of ice (Glasser et al., 2015; Ely et al., 2017), channelise surface meltwater downstream, likely explaining the elongate shape of lakes observed on the ice shelf. Variations in the downstream extent of lakes between years are therefore likely to partly be a consequence of variable melt supply (Spergel et al., 2021). The distribution of surface basins on grounded ice is controlled by subglacial topography, meaning lakes can form annually in fixed surface depressions (Echelmeyer et al., 1991; Ignéczi et al., 2018).

There is a clear intra-seasonal pattern of total lake area; it remains low through the early part of the melt season, before rapidly increasing during late December and reaching a maximum in January (Fig. 8; Table 2) and then decreasing sharply during February. This matches with results from scatterometer studies which show large decreases in backscatter values over the AIS in January, indicating a rapid increase in the intensity of surface melting (Oza et al., 2011). The sudden increase in lake area (up to an order of magnitude increase within half a month) is likely a consequence of the hypsometry of the study region. Over 35 % (∼ 65 000 km2) of the study region lies at an elevation lower than 200 m a.s.l., meaning that a minor increase in temperature increases melt potential over a vast area of ice. This contrasts with the typical hypsometry of the Greenland Ice Sheet, where relatively steep slopes at the ice sheet margin mean that an equivalent rise in temperature would initiate melting over a much smaller area (McMillan et al., 2007; Sundal et al., 2009). The large lake area contribution from low elevations possibly explains why we do not observe a major elevation shift in peak area contribution throughout the melt season (Fig. 7), as the signal from the ice shelf masks any changes in total lake area contribution at higher elevations. Following the initial appearance of meltwater ponds, overall lake area is likely further enhanced by positive feedbacks, whereby lowered surface albedo from melting promotes further melting. Furthermore, the development of surface streams enables lateral transfer of surface water, rapidly increasing the spread of water across the ice shelf surface (Kingslake et al., 2017). Sharp decreases in lake area during February are presumably indicative of the widespread freezing of SGLs, although evidence of lake drainage events has also been observed in the region (Fricker et al., 2009; Pan et al., 2020; Spergel et al., 2021).

There is a strong association between annual cumulative lake area and the summer SAM index (Fig. 10), suggesting that annual ice-shelf-wide variations in lake area cover are influenced by large-scale climate variability. Phases of the SAM naturally oscillate on a multi-decadal timescale (Picard et al., 2007), possibly explaining the observed multi-year phases between periods of low and high lake area coverage (Fig. 8). When SAM is in a positive phase, air temperatures are typically higher over the Antarctic Peninsula and lower over the rest of the continent, whilst the reverse is the case during a negative SAM phase (Marshall and Thompson, 2016; Turner et al., 2020). Our results broadly support this relationship, as observed by the statistically significant negative correlation between lake area and summer SAM index (Fig. 10). For example, the 7-year period between 2006 and 2013, which was largely characterised by positive summer SAM indexes, coincided with low annual cumulative surface meltwater coverage. The only year during this period with a negative summer SAM index (where we would expect slightly warmer temperatures) was in 2009/10. This melt season had the highest cumulative lake area of this 7-year period, suggesting that the summer SAM index is linked to melt rates on an annual basis. This wider climatic control on SGL formation suggests that the AIS has an abundance of basins within which meltwater can be accommodated, resulting in a linear relationship between melt rates and SGLs (Fig. 9). This may not necessarily be the case in other regions of Antarctica, where steeper topography may limit the number and size of depressions able to host meltwater, thus resulting in enhanced surface runoff and a non-linear relationship between melt and SGL area.

There was high variability in the austral summer SAM index from 2013–2020, ranging from -1.75 in 2016/17 to 3.69 in 2014/15. In general, lake areas followed the broad pattern we would expect based on their association with the SAM throughout this time period, with the main exception being the 2014/15 melt season. Large lakes formed during this melt season, despite there being a negative SAM and low melt rates predicted by RACMO. Greater-than-expected meltwater ponding during this melt season can be explained by enhanced scouring of the ice shelf surface by strong katabatic winds. Following a snowfall event in late November 2014, large areas of low-albedo blue ice were exposed on the ice shelf by mid-December (Fig. 11a, b), suggesting strong wind scouring throughout the first half of December. Between 21 and 28 December, the ice shelf was transformed from being almost entirely lake-free to widely covered by SGLs (Fig. 11c). The following melt season, by contrast, snow cover persisted across most of the ice shelf throughout December (Fig. 11e), meaning any meltwater could be accommodated within the firn pack rather than ponding as surface water.

The formation and extent of SGLs are highly sensitive to minor fluctuations in surface air temperature (Langley et al., 2016). During December 2014, the ice shelf was pre-conditioned as a low-albedo, impermeable surface, optimising the conditions required for surface meltwater ponding. Given this, it is likely that a transient increase in air temperature, possibly induced by a strong katabatic event, could have resulted in a large change in surface meltwater characteristics. Surface melt rates depend on all terms of the surface energy balance (Oza et al., 2011), meaning air temperature is not the sole factor in determining surface melt rates. Whilst RACMO-modelled melt estimates include a surface albedo parameterisation, melt–albedo feedbacks are difficult to resolve due to the lack of representation of blue ice within the model and the relatively coarse resolution of the data (27 km). Previous studies have shown that RACMO often underpredicts meltwater production in areas of Antarctica where blue ice is warmed by katabatic winds (Trusel et al., 2013; Lenaerts et al., 2017). This possibly explains why there were such major differences in lake area coverage between 2014/15 and 2015/16, despite RACMO mean seasonal snowmelt estimates differing by only ∼ 2 mm w.e. (Figs. 8 and 9). Jakobs et al. (2019) found that surface albedo was the main difference in ice surface characteristics between high- and low-melt years on the Ekström ice shelf, supporting our hypothesis that large variations in melt extent can be caused by variations in surface reflectance characteristics. Over our entire study period, however, RACMO shows a good agreement with lake area.

Whilst the AIS has some of the highest concentrations of surface meltwater ponding in Antarctica, the ice shelf lies in a region that is currently thought to be largely resilient to widescale hydrofracture (Lai et al., 2020). Substantial lateral buttressing from the valley sides results in relatively low tensile longitudinal resistive stresses, meaning increased meltwater ponding is unlikely to cause the rapid breakup of the ice shelf (Lai et al., 2020). However, given the vast amount of ice that is discharged through the AIS, it is crucial that we continue to develop our understanding of how varying levels of surface meltwater can influence hydrological and ice dynamic processes in the region. Repeated cycles of melting and refreezing at the ice surface releases latent heat, weakening the ice structure and making it more prone to future climatic perturbations (Hubbard et al., 2016). Changes in temperature or precipitation patterns, in addition to predicted ocean warming, could also influence the vulnerability of the ice shelf to melt-induced fracture. Furthermore, if meltwater starts to pond at higher elevations on a regular basis, crevasses on steeper topography may start to undergo enhanced hydrofracture processes (Tuckett et al., 2019). The advection of this weakened ice structure onto the ice shelf could precondition the ice shelf to further fracturing from greater volumes of surface meltwater ponding (Dunmire et al., 2020).

The association we observe between lake area and RACMO-modelled snowmelt gives us confidence in the ability of this model to predict future melt conditions. These results show that modelled melt rates from RACMO could be used to generate first-order predictions of surface meltwater area at an annual scale for the AIS region. However, some melt conditions that lead to the formation of lakes are not currently well captured by RACMO, such as the influence of blue ice on lake formation (Fig. 11). Snowmelt–albedo feedbacks have a particularly strong influence on melt rates in East Antarctica (Jakobs et al., 2021), and further work is required to quantify this process within modelled melt estimates. Future work should also evaluate whether a similar relationship between modelled melt and lake area occurs for other areas in Antarctica. The surface characteristics of some regions may preclude the formation of surface lakes (e.g. if firn aquifers are present), resulting in a weaker association between modelled melt and observed lakes, even if modelled estimates are broadly accurate. It is also likely that variations in hypsometry and lateral meltwater transfer alter the lag we find between modelled melt and peak meltwater ponding (Fig. 8, Table 2).

The influence of the SAM on future meltwater cover in the study region will likely be influenced by trends in both stratospheric ozone levels and greenhouse gas emissions (Fogt and Marshall, 2020). Stratospheric ozone depletion has led to positive trends in the SAM in the austral summer season over recent decades, although there are signs that recovery of the stratospheric ozone hole is starting to counter this trend (Banerjee et al., 2020). Increases in greenhouse gas emissions have been shown to have a secondary influence on the SAM by strengthening the mid- to high-latitude temperature gradient, hence resulting in a more positive SAM (Arblaster et al., 2006). Future melt rates on the AIS will therefore likely be influenced by several competing climatic factors, with enhanced melt from regional warming and near-surface feedbacks potentially being offset by decreased melt associated with a positive SAM.

Large volumes of surface meltwater on grounded ice around the AIS (∼ 30 % of estimated total lake area) leave open the potential for surface-to-ice bed connections to develop via hydrofracture (Krawczynski et al., 2009). Surface-melt-induced variations in Antarctic ice flow have currently only been inferred to occur on northern parts of the Antarctic Peninsula (Tuckett et al., 2019). However, it is likely that surface-to-bed hydraulic connections will become more frequent as Antarctic-wide temperatures increase (Bell et al., 2018), and evidence of lake drainage events has already been identified in the grounding zone of the AIS region (Fricker et al., 2009; Pan et al., 2020; Spergel et al., 2021). The injection of surface meltwater to the ice sheet bed could also have implications on rates of ice shelf basal melting, as a consequence of meltwater plumes emerging at the grounding line (Jacobs et al., 1992). Future work should therefore investigate the distribution and recurrence frequency of lake drainage events and assess whether they have any impact on grounded ice flow of glaciers feeding the AIS. If such a link were found to exist, it could have significant impacts on the speed at which ice is discharged into the AIS, hence influencing rates of sea level rise.