For automated mapping of Antarctic supraglacial lake extents, we jointly use Sentinel-1 and Sentinel-2 satellite imagery. The Sentinel-1 SAR data used in this study were acquired in ground range detected (GRD) format and interferometric wide (IW) swath mode, delivering data products at 10 m pixel spacing and in horizontal transmit-and-receive polarization (HH). While the general repeat interval of Sentinel-1 corresponds to 6 d at the Equator, up to daily revisit frequencies can be achieved in high latitudes. For each study region, we acquired all available Sentinel-1 data covering the period 1 November to 31 March in 2015–2021 within pre-defined boundaries, including the six ice shelves as well as surrounding grounded area. Over Riiser-Larsen, Nivlisen and Amery ice shelves, data availability during the 2015–2016 melting season was not sufficient, restricting the Sentinel-1 data selection to the period 2016–2021. In contrast, Sentinel-1 data over George VI, Bach and Wilkins ice shelves were available during all melting seasons 2015–2021. Overall, this resulted in a total number of 3075 Sentinel-1 acquisitions covering all six study regions.

Regarding optical Sentinel-2 data, we selected all imagery with a scene cloud cover < 75 % available within the same pre-defined time frame and boundaries as defined for Sentinel-1. This comparatively high cloud threshold was chosen in order to obtain the most complete record of supraglacial lake occurrence during months when high cloud cover prevails in parts of the image scenes. Moreover, cloud masking procedures over glaciated regions are often impacted by the appearance of ice similar to clouds, thus indicating higher cloud coverage than is present in the actual acquisitions and resulting in an incomplete data record. Considering these filtering criteria as well as the high revisit frequency of Sentinel-2 over polar regions (<5 d), a total of 2167 Sentinel-2 acquisitions were gathered for all six study regions within the period November to March in 2016–2021. Similarly to Sentinel-1, data availability was restricted during melting season 2015–2016.

For analysis of climatological controls on Antarctic supraglacial lake occurrence in 2015–2021, we use hourly ERA5-Land atmospheric reanalysis data (Munoz Sabater, 2019). ERA5-Land replaces the former ERA-Interim product of the European Centre for Medium-Range Weather Forecasts (ECMWF) and provides reanalysis data for a variety of land variables at a spatial resolution of 0.1∘ (∼9 km) as well as a temporal coverage dating back to 1981. The use of large-scale reanalysis data compared to ground-based observations, e.g. from local weather stations, is particularly beneficial over Antarctica as its vast size and remote geographical location complicate detailed ground-based monitoring efforts.

The ERA5-Land variables used in this study include 2 m air temperature, surface net solar radiation, wind, total precipitation and snowmelt. Data were retrieved for the period 1 October to 1 March of each melting season in 2015–2021. October and November data were included for evaluation of potential time lag effects of climate variables. For precipitation data, we retrieved hourly estimates for all remaining months in 2015–2020 in order to allow the evaluation of links between previous-year precipitation and supraglacial lake formation.

To start with, ERA5-Land temperature data provide the best estimate of the seasonal cycle of near-surface temperature over coastal Antarctica and particularly the API (Gossart et al., 2019; Tetzner et al., 2019). ERA5-Land temperature data outperform other reanalysis datasets considering its low bias and low mean absolute error of 2 ∘C over Antarctica (Gossart et al., 2019). On the other hand, surface net solar radiation represents the difference between downward solar (or shortwave) radiation incident upon the Earth's surface and the amount of radiation reflected back to space. To date, data on surface net solar radiation have not been validated over Antarctica, while studies over China and Europe point towards an improved performance of ERA5 radiation components compared to other reanalysis datasets (Jiang et al., 2020; Urraca et al., 2018). ERA5-Land wind data are provided at 10 m above the surface as the horizontal speed of air moving towards the north and east, respectively. ERA5-Land wind data were revealed to capture the typical annual cycle of wind with a mean underestimation during winter, when katabatic wind forcing is strongest (Gossart et al., 2019). Similarly, a slight underestimation of wind magnitude was observed over coastal Antarctica during summer, resulting in an overall mean absolute error of 2.8 m s-1 (Gossart et al., 2019). Estimates of ERA5-Land total precipitation are provided in metres depth and include all liquid and frozen water accumulated at the Earth's surface. Overall, ERA5-Land precipitation data perform well in detecting precipitation events over the API even though a bias may exist, e.g. due to the negligence of effects such as the redeposition of snow (Tetzner et al., 2019). Finally, ERA5-Land modelled snowmelt data are available in metres of water equivalent. As no evaluation of the ERA5-Land snowmelt product exists for the Antarctic continent, the data should be interpreted with care.

The satellite data were pre-processed using sensor-specific processing tools. For Sentinel-1 SAR data, we used the open-source Sentinel Application Platform (SNAP). Pre-processing of Sentinel-1 SAR imagery includes orbit correction, thermal noise removal, radiometric calibration, speckle filtering and terrain correction. For the latter, the TanDEM-X PolarDEM 90 m of Antarctica (Wessel et al., 2021) was used. In order to prepare the data for prediction of supraglacial lakes (Sect. 3.2.2), each pre-processed Sentinel-1 image was normalized and tiled into 480×480 px patches with 200 px overlap (Dirscherl et al., 2021). In contrast, optical Sentinel-2 data required correction for atmospheric effects using Sen2Cor, a processor for retrieval of surface reflectance (Louis et al., 2016). Moreover, selected indices needed for image classification (Dirscherl et al., 2020) were calculated, and all bands and indices were resampled to the highest resolution of available Sentinel-2 bands (10 m).

Pre-processing of ERA5-Land data included a range of individual processing steps for each variable. Hourly estimates of ERA5-Land air temperature were resampled to daily averages, and temperature measured in kelvin was converted to degrees Celsius (∘C). Hourly data on surface net solar radiation were converted to average daily energy fluxes in watts per square metre (W m-2). Hourly zonal (u) and meridional (v) wind stress components at 10 m above the surface were used to derive daily averages of zonal and meridional wind stress to be converted to wind magnitude (wm=u2+v2) and direction (wd=180+(arctan⁡(v,u)⋅180/π)) following aggregation to bi-weekly averages (Sect. 3.2.3). For calculation of daily total precipitation, we considered the total accumulated value of a day and converted to millimetres depth. Finally, daily snowmelt information was obtained from hourly snowmelt data through consideration of the total accumulated value of a day and conversion to millimetres of water equivalent. All obtained daily ERA5-Land data were reprojected to the Antarctic Polar Stereographic projection in order to allow the comparison to lake extent mappings.

For prediction of supraglacial lake extents in normalized Sentinel-1 image patches, we use a convolutional neural network (CNN) based on residual U-Net (Dirscherl et al., 2021). In detail, a standard U-Net architecture was complemented with residual connections as well as an Atrous Spatial Pyramid Pooling (ASPP) module for improved performance. In this context, the choice of a deep learning network over traditional image processing techniques was mainly made due to the similar appearance of Antarctic supraglacial lakes and surface features such as wet snow, blue ice or topographic shadow in single-polarized SAR imagery as well as due to effects such as wind roughening and speckle noise (Dirscherl et al., 2021). Model training was performed on a graphics processing unit (GPU) cluster using 21 200 Sentinel-1 image patches extracted from 58 manually labelled Sentinel-1 image scenes covering melting seasons 2018–2019 and 2019–2020 over 13 different Antarctic regions. Following semantic segmentation of the image patches, the resultant prediction probability maps were reshaped into their original image size and thresholded. In particular, all pixels above a probability threshold of 0.5 were accepted as supraglacial lake pixels. As part of post-processing, remaining errors in the binary classification maps were reduced by applying morphological filtering as well as topographic and coastline masking using a resampled version of the TanDEM-X PolarDEM as well as a dilated Sentinel-1 coastline product of 2018 (Baumhoer et al., 2021). The performance of the Sentinel-1 classification approach was evaluated by means of 23 000 random point samples collected over 10 independent test sites that were not presented to the model during training and that contained surface features that are difficult to discriminate from lakes as well as lakes of varying appearances. Their evaluation as part of a confusion matrix suggests the good functionality of the algorithm with an average F1 score of 93 % for all regions. The F1 score is a common statistical accuracy metric reflecting the harmonic mean of the rate of true positive pixels with respect to all true class samples and the rate of true positive pixels with respect to all predicted class samples (Jolly, 2018; Müller and Guido, 2016). Remaining limitations are mostly related to false positive lake classifications over radar shadow, wet snow or blue ice.

Supraglacial lake classification in optical Sentinel-2 imagery was performed using an advanced version of the machine learning algorithm published in Dirscherl et al. (2020). In detail, the algorithm is based upon a random forest (RF) classifier trained on a high-performance computing (HPC) cluster with selected Sentinel-2 bands and indices covering 14 different Antarctic regions during three austral summers as well as manually created labels and topographic variables derived from the Antarctic TanDEM-X digital elevation model (DEM). Similarly to Sentinel-1, post-classification involved bands, indices, topographic and coastline masking, and morphological erosion. The main modifications in the advanced version of the RF algorithm include (i) the integration of more training data over shadow; (ii) the improvement of post-classification, e.g. through fine-tuning of variable thresholds; and (iii) the use of prediction probability maps for efficient slush elimination. Overall, this classification strategy returned an average F1 score of ∼95 % for a random selection of 28 000 points that were sampled over 14 independent Antarctic test sites not presented to the model during training. Remaining limitations of the classification method include difficulties in detecting mixed pixels at lake edges as well as errors over shadow. Yet, errors due to shadow mainly affect acquisitions towards the end of a summer season, when the incidence angle of the sun is lower. Furthermore, the discrimination between open water and slightly frozen lakes (e.g. Fig. 2d and e) remains challenging. For this reason, we restrict our analysis to supraglacial lake extent derivation and do not calculate lake volumes. To support analyses of local controls on lake ponding, we also derive mappings of rock outcrop using the modified RF algorithm.

Following supraglacial lake extent mapping, we aggregated all data to bi-weekly time series. To start with, Sentinel-1 and Sentinel-2 classification products were combined through decision-level fusion, thus the mosaicking of all classifications to maximum lake extent classifications for each bi-weekly time interval between November and March of a melting season. To interpolate data gaps during specific months over Riiser-Larsen and Nivlisen ice shelves (Fig. S1d and e in the Supplement), we used the respective long-term mean of the period 2016–2021 at bi-weekly scale. Since data gaps were mostly present for months of low lake coverage, this interpolation strategy was considered acceptable. For analysis of environmental controls on Antarctic supraglacial lake occurrence, we converted the bi-weekly lake extent time series to fractional lake extent considering ERA5-Land grids as AOIs (see Fig. 1). Similarly, the daily ERA5-Land data were aggregated to average bi-weekly time series, and the grid values over the AOIs were extracted. For temperature, we also calculated bi-weekly maxima to capture temperature extreme events and positive degree days, resulting in seven climate variables used for statistical analysis. In order to investigate drivers of anomalous meltwater ponding and to analyse where above- or below-normal climatic conditions result in above- or below-normal lake formation with respect to the investigation period 2015–2021, we additionally converted the bi-weekly fractional lake extent and climate time series to time series of anomalies. In this study, anomalies are calculated with respect to the long-term mean 2015–2021 of each bi-weekly time interval. Finally, for joint analysis of January lake extents and total precipitation in a year preceding lake coverage, we extracted the corresponding aggregates and grid values over each study region. In this study, previous-year precipitation was used as an indirect indicator for the FAC of snow, thus for analysis of relationships between supraglacial lake formation and the state of the snowpack. Additionally, January lake extents were considered for analysis of links with annual Southern Annular Mode (SAM) (Marshall, 2003, 2018) as well as for evaluation of the inter-annual recurrence of lakes, both spatially and temporally supporting the investigation of local controls.

To establish a link between environmental variables and fractional lake extents, we use the Pearson correlation. Pearson's correlation coefficient (r) measures the linear relationship between two variables, where r lies in the range of -1 to +1, reflecting a strong negative and strong positive linear relationship, respectively. Calculation of r also raises the p value (p), an indicator for statistical significance of correlations. In this study, significance levels of p<0.05(*), p<0.01(**) and p<0.001(***) were defined. In order to obtain more robust correlation values, the bi-weekly statistical analysis was restricted to summer months (December, January and February), when lake coverage is highest, and data gaps and classification errors are lowest. Apart from correlating multi-temporal observations of fractional lake extents and climate variables at identical time steps (lag 0), we performed cross-correlation with climate data at time lags 1–4, with each individual lag representing a shift of half a month with respect to the previous lag. These lags capture climatic conditions up to 2 months preceding a summer season. Considering the investigation periods December to February 2015–2021 and 2016–2021, each individual time series consisted of 36 (API) and 30 (EAIS) observations. Correlation of multi-temporal observations was performed with spatial averages, anomalies of spatial averages and pixel-based values. Additionally, fluctuations in previous-year precipitation and SAM were correlated with January fractional lake extents based on six (API) and five (EAIS) observations. For completeness, further local controls including melt–albedo feedbacks and topographical factors were investigated based on the visual comparison of lake density to the TanDEM-X PolarDEM, Sentinel-2 rock classification products and blue ice occurrence in optical imagery.

As shown in Fig. 3a–c, supraglacial lake extents on the API varied strongly both inter-annually and intra-annually. For George VI, Bach and Wilkins ice shelves, supraglacial lake extents were highest during melting seasons 2019–2020 and 2020–2021, reaching peak values of ∼805, ∼110 and ∼45 km2 during January 2020. Comparatively low lake extents were observed for melting seasons 2015–2019. For years with high supraglacial lake coverage, the peak of lake formation was in late January, reducing significantly during the first or second half of February. Moreover, the calculation of intra-annual anomalies with respect to the reference period 2015–2021 (Fig. 4a–c) revealed that lake extents during the 2019–2020 and 2020–2021 melting season were above average during their peak and slightly below average in parts of early 2019–2020 and late 2020–2021.

Figure 5a–c illustrate the annual recurrence of lakes during January 2016–2021. As can be seen, supraglacial lakes tend to form at low elevations and low surface slopes downstream of the grounding line. On Bach and Wilkins ice shelves, lakes are clustered closer to the grounding line than on George VI Ice Shelf, where lakes accumulated in flow stripes over vast parts of the flat ice shelf during the last two melting seasons. Near Wilkins Ice Shelf, lakes extended furthest onto the adjacent grounded ice, where rock outcrop is present (Fig. 5c). Further, supraglacial lakes on George VI Ice Shelf are clustered closer to the grounding line and rock outcrop in years with low supraglacial lake coverage (see Fig. 5a and b), as reflected in higher recurrence times of lakes up to 5–6 years.

Supraglacial lake extent dynamics on the EAIS (Fig. 3d–f) differ from the pattern observed on the API. Over all three ice shelves, comparatively high lake coverage was observed during the 2016–2019 melting seasons, while particularly low lake coverage was found for the 2020–2021 melting season. This pattern is also reflected in the calculation of supraglacial lake extent anomalies for the overall period 2016–2021 (Fig. 4d–f). The highest supraglacial lake extents over Riiser-Larsen (∼85 km2), Nivlisen (∼107 km2) and Amery (∼1373 km2) ice shelves were observed during early January 2020, late January 2019 and early January 2019, respectively. The average onset of EAIS supraglacial lake formation was in late December, with lakes persisting longer throughout the melting seasons compared to API lakes.

The annual recurrence of lakes during January shows clustered supraglacial lake occurrence near the grounding line of Riiser-Larsen Ice Shelf, where a blue ice region is present (Figs. 2d and 5d); lakes spreading across Nivlisen Ice Shelf, with a progressive advance in a north-easterly direction (Figs. 5e and S3); and widespread lake occurrence on the southern and eastern Amery Ice Shelf (Fig. 5f). During years of peak lake occurrence, lakes expanded further onto the flat ice shelves, while years with lower lake coverage were characterized by lakes clustering closer to the grounding line, rock outcrop or blue ice (Figs. 1d–f, 2d, 5d–f and S3). Over all three ice shelves, lakes formed at low elevations and low surface slopes. At the same time, the January recurrence of lakes reached up to 4–5 years in local depressions upstream of the grounding line (Fig. 5d–f).

Figure 6 presents the results of the statistical correlation analysis between climate variables and fractional lake extents (see Tables S1 and S2 in the Supplement for individual correlation values). As can be seen, bi-weekly mean and maximum air temperature correlates positively at lag 0–1 over all three ice shelves considering both spatial averages (Fig. 6a) and anomalies of spatial averages (Fig. 6b). Moreover, lake extents over George VI Ice Shelf correlate positively with spatial averages of lag 2 solar radiation, and lake extents over Wilkins Ice Shelf correlate negatively with anomalies of lag 1 solar radiation and lag 1 wind direction. Lastly, surface melt correlates positively over Bach and Wilkins ice shelves considering spatial averages and anomalies.

Linear correlation analysis over EAIS ice shelves (Fig. 6a and b) returned significant positive lag 0–3 relationships with bi-weekly mean and maximum air temperature considering both spatial averages and anomalies. Likewise, solar radiation correlates positively at varying lags over all three ice shelves. Over Riiser-Larsen Ice Shelf, significant negative lag 0 correlation was found for precipitation and wind magnitude, and significant positive lag 0 correlation was found for wind direction considering spatial averages and anomalies. Further, statistical analysis over Nivlisen Ice Shelf returned negative lag 1 correlation for wind direction. Finally, surface melt correlates positively at lag 1 over Riiser-Larsen and Amery ice shelves.

Correlation with annual SAM (Table 1) indicates a strong negative linear relationship over all three API ice shelves. Over EAIS ice shelves, r values are comparatively low, with Riiser-Larsen correlating negatively and Nivlisen and Amery returning positive correlation values. Regarding correlation with previous-year precipitation (Table 1), we find negative relationships over Amery, George VI, Bach, Wilkins and Riiser-Larsen ice shelves and low positive correlation over Nivlisen Ice Shelf.

Figures 7 and 8 illustrate selected results for pixel-based correlations, chosen on the basis of Fig. 6. The figures show only significant pixels with the highest correlations considering all performed cross-correlations as well as the corresponding lags. Remaining pixel-based correlations can be found in the Supplement (Figs. S4 and S5). Pixel-based correlation over George VI Ice Shelf (Fig. 7a–h) revealed positive relationships for temperature and solar radiation correlating at higher lags in the central ice shelf section compared to locations near the grounding line. Likewise, statistical analysis over Bach Ice Shelf (Fig. 7i–p) implies predominant positive correlations for temperature and radiation. Over both ice shelves, snowmelt correlates positively, and wind direction correlates positively in the east and negatively in the west. Spatial correlations over Wilkins Ice Shelf were rather heterogeneous (see Figs. 5c and 7q–x).

Over EAIS ice shelves (Fig. 8a–x), positive correlation was found for bi-weekly average air temperature and solar radiation. While lags were varying for each ice shelf, they were generally higher on the ice shelf area than on adjacent inland ice and near the grounding line. Moreover, correlation of lake extents with wind magnitude and precipitation over Riiser-Larsen Ice Shelf revealed a negative relationship at lag 0, and correlation with wind magnitude over Amery Ice Shelf revealed negative and positive relationships at varying lags along the south-eastern and eastern grounding zone.

Finally, Fig. 9 outlines the spatial correlation pattern between previous-year precipitation and January fractional lake extents. As correlations result from only few observations, potentially leading to unreliable values of p, correlations with p>0.05 are also shown. As can be seen, previous-year precipitation correlates negatively over regions of frequent lake recurrence (Fig. 5) on George VI, Bach, Wilkins, Riiser-Larsen and Amery ice shelves and positively over most of Nivlisen Ice Shelf.

Starting with local control factors, correlation analysis with precipitation revealed API lake coverage to be peaking in years with previous-year low precipitation (Table 1, Fig. 9a–c). This highlights that supraglacial lake formation is promoted over more compact, air-depleted firn (Arthur et al., 2020a; Leeson et al., 2020), while fresh snow dampens lake evolution allowing the percolation of meltwater into the low-density firn layer before it can accumulate on the surface (Kuipers Munneke et al., 2014; Lenaerts et al., 2017; Stokes et al., 2019). In fact, low precipitation in 2017, 2019 and 2020 might have led to increased firn air depletion, potentially contributing to facilitated meltwater ponding in 2017–2018, 2019–2020 and 2020–2021. Apart from the local glaciological setting, also the local topography and ice shelf geometry control lake distribution. Over George VI Ice Shelf, meltwater accumulated in flow stripes controlled by the prevailing compressive flow regime (Fig. 5a) (LaBarbera and MacAyeal, 2011; Hambrey et al., 2015). Further, lakes on all three ice shelves appeared at low elevations and low surface slopes (Fig. 5a–c), being favourable to meltwater ponding (Arthur et al., 2020b; Stokes et al., 2019). Additionally, lake ponding was dictated by the ice shelf geometry and grounding line location where lakes tended to cluster (Fig. 5a–c). Other local factors that could have influenced lake distribution are the refreezing of meltwater in firn pore space and channelized basal melting, causing a modified ice surface topography (Spergel et al., 2021; Dow et al., 2018; McGrath et al., 2012).

In addition, cross-correlation analysis demonstrated a positive relationship between API supraglacial lake formation and air temperature, also considering temperature extremes, with lag 1 prevailing in regions where lakes form during high-melt years and lag 0 in regions where lakes form during low-melt years closer to the grounding line and rock outcrop (Figs. 5a–c, 6, and 7b, j, r). Moreover, the observed correlation pattern reveals that anomalously high or low lake extents result from anomalously high or low air temperatures (Fig. 6b). On the one hand, this indicates that meltwater ponding is initiated directly after temperature rises, particularly in regions where meltwater accumulation is promoted through melt–albedo feedbacks near rock outcrop, as present upstream of the grounding line of Wilkins and George VI ice shelves (Fig. 5a and c), while lake formation farthest from these regions occurs at slower timescales. On the other hand, meltwater transport across the ice shelves (Kingslake et al., 2017) later in the melting season can likely be attributed to higher lags downstream of the grounding lines (see connected drainage systems in Fig. 2a–c). The detected link between supraglacial lake formation and air temperature was previously reported for other Antarctic regions (Kingslake et al., 2015; Langley et al., 2016; Dell et al., 2020) and is well in line with the analysis in Banwell et al. (2021), suggesting that high meltwater ponding in 2019–2020 over George VI Ice Shelf results from sustained air temperatures ≥ 0 ∘C.

Over George VI and Bach ice shelves, solar radiation exhibited a positive relationship with supraglacial lake occurrence, particularly in regions where lakes cluster during high-melt years (Figs. 5a, b and 7c, d, k, l). This points towards the amount of incoming solar radiation being an equally important driver for supraglacial lake formation on API ice shelves. Similar conclusions were drawn by Laffin et al. (2021), who proposed that enhanced solar radiation is a main driver for foehn-induced melt on Bach and George VI ice shelves. In contrast, correlation with anomalies of solar radiation over Wilkins Ice Shelf returned a negative link at lag 1, which could indicate that lake formation results from enhanced longwave radiation, e.g. due to cloud cover and thus the higher importance of effects such as sensible heat transport, as discussed in Laffin et al. (2021) and Banwell et al. (2021). As correlations over Wilkins Ice Shelf were rather heterogeneous overall, the detected pattern could also reflect spurious correlation. However, weak negative lag 0–1 correlation was also found for George VI and Bach ice shelves (Table S2), reflecting an anomalously low availability of incoming solar radiation during the last two melting seasons (Fig. S6c and d).

Furthermore, cross-correlation of supraglacial lake extents over Wilkins Ice Shelf with anomalies of wind direction (Fig. 6b) point towards the existence of a negative link over regions of frequent lake coverage (Figs. 5c and 7u). Analysing the corresponding time series in more detail, negative correlations seem to be caused by predominant occurrence of north(-easterly) winds (i.e. north-easterly winds that are inclined more towards northerly direction) preceding recent increased lake ponding (Fig. S6). At the same time, positive correlations were visible, reflecting short-lived shifts to north-westerly winds preceding periods of high lake formation. Similar shifts in wind direction are also visible along the western and eastern grounding lines of George VI and Bach ice shelves (Fig. 7e and m). Wind conditions are a known driver for supraglacial lake occurrence on the API, with previous studies reporting that warm air temperatures over George VI Ice Shelf are promoted by sensible heat transport with warm north-westerly and north-easterly winds as well as a shift towards more frequent north-easterlies (Banwell et al., 2021). Moreover, foehn winds are known to influence surface melting on the north-eastern API (Cape et al., 2015; Datta et al., 2019; Luckman et al., 2014; Turton et al., 2020) and can equally be observed on the western API with (north-)easterly prevailing wind direction (Laffin et al., 2021). This suggests the occurrence of both foehn conditions and sensible heat transport on the API. Foehn winds could also explain the clustering of lakes near the grounding lines, where foehn wind influence is assumed to be highest (Lenaerts et al., 2017), as well as the positive link with lag 0–1 snowmelt (Fig. 7) (Cape et al., 2015; Turton et al., 2020). Further, the detected relationship is in good agreement with our observations of air temperature and solar radiation, where radiation potentially influences foehn conditions, and foehn conditions influence air temperature.

For investigation of large-scale climatic drivers, annual SAM was correlated with January lake extents. Our results imply a negative relationship for all investigated API ice shelves (Table 1), mostly due to occurrence of a negative SAM year in 2019 (Fig. 10) coincident with enhanced melting in 2019–2020 (Fig. 4a–c). Although the detected relationship seems counterintuitive because negative SAM years usually lead to cold conditions on the API, Bevan et al. (2020) suggest that teleconnections between the tropics and high latitudes in 2019 are responsible for the observed 2019–2020 melt and negative SAM. In detail, these teleconnections caused unusual perturbations of Southern Hemisphere atmospheric flow as well as sudden stratospheric warming (SSW), thus causing SAM to be negative and the ongoing positive Indian Ocean Dipole (IOD) in 2019 (BOM, 2021), represented via the dipole mode index (DMI) (Fig. 10), to bring warm air to the API via northerly winds. This overarching driver is well in line with correlations between lake extents and air temperature as well as wind and could potentially explain increased 2019–2020 meltwater ponding. As DMI returned to neutral values in 2020, other overarching drivers, including SAM returning to positive values (Fig. 10) as well as modified pressure conditions in the Amundsen Sea Low (ASL) (NSIDC, 2021), are believed to have contributed to high and low lake extents during the early and late 2020–2021 melting season, respectively.

Overall, our results indicate that the complex interplay of local, regional and large-scale environmental controls dictates supraglacial lake formation on the API. In particular, the local glaciological and microclimatic setting including melt–albedo feedbacks, sensible heat exchange or the ice shelf topography was found to play an important role for supraglacial lake distribution. Likewise, long-term precipitation conditions influence the FAC and were found to be a primary control on the location and strength of meltwater ponding over all ice shelves. On the other hand, regional climatic conditions and particularly air temperature, solar radiation and (foehn) wind conditions are among the main controlling factors for temporal dynamics of supraglacial lakes in 2015–2021 and are mainly driven by large-scale atmospheric modes including SAM, IOD or ASL. In consequence of the good agreement between lake evolution and detected driving factors on all ice shelves (Table 2), a similar response of other API ice shelves resulting in high meltwater coverage during 2019–2020 and 2020–2021 is suggested.

For EAIS ice shelves, the observed negative relationship with short-term precipitation over Riiser-Larsen Ice Shelf again highlights that supraglacial lake formation is controlled by the local glaciological setting and a reduced FAC (Figs. 6 and 8g, h). At the same time, weak negative correlation (p>0.05) was found over Nivlisen and Amery ice shelves (Figs. 6 and 8o, w, Tables S1 and S2), potentially reflecting the higher importance of long-term precipitation effects. In this context, correlation analysis with previous-year precipitation revealed significant negative relationships over Riiser-Larsen and particularly Amery Ice Shelf (Table 1, Fig. 9d and f). Most notably, we found a strong link between 2016–2017 and 2018–2019 high meltwater ponding and previous-year low precipitation over Amery Ice Shelf. In contrast, correlation with previous-year precipitation over Nivlisen Ice Shelf showed a mostly positive correlation pattern, implying the higher importance of other drivers (Fig. 9e). In this context, the high recurrence of lakes on EAIS ice shelves (Fig. 5d–f) reflects lakes reforming in the same location each year and thus the high importance of the local topography further controlling supraglacial lake distribution (Arthur et al., 2020b; Echelmeyer et al., 1991; Langley et al., 2016). Over Riiser-Larsen and particularly Nivlisen Ice Shelf, elongated meltwater bodies were found to reoccupy the same ice shelf surface depressions (Fig. 5e) at low elevations and low surface slopes, while lake ponding over Amery Ice Shelf seems to be dictated by both local surface depressions on grounded ice and longitudinal flow stripes migrating with ice flow on floating ice (Figs. 2f and 5f) (Glasser and Gudmundsson, 2012; Spergel et al., 2021). Furthermore, the refreezing of meltwater in firn pore space could have influenced lake ponding through a modification of the ice surface topography and firn air depletion (Kuipers Munneke et al., 2014; Spergel et al., 2021).

Furthermore, cross-correlation analysis with air temperature indicates that supraglacial lake formation is driven by temperature at lags 0–2 (Figs. 6 and 8a, b, i, j, q, r). Similarly to what is found for API ice shelves, the spatial pattern of temperature time lags (Fig. 8b, j and r) revealed larger lags for regions where lakes are spreading onto the ice shelves during years of high meltwater ponding and lower lags close to and upstream of the grounding lines. This likely reflects the temporal evolution of drainage systems through lateral meltwater transport across the ice shelves (Dell et al., 2020; Spergel et al., 2021; Kingslake et al., 2017) (Figs. 2d–f, 5d–f and S3) or a longer response time to air temperature, particularly in regions where the local glaciological setting does not promote supraglacial lake formation through melt–albedo feedbacks.

Cross-correlation with solar radiation returned positive relationships at lags 0–2 over all three ice shelves (Figs. 6 and 8c, d, k, l, s, t). As above, time lags were highest where lakes expanded farthest onto the ice shelves during years of high lake coverage, likely reflecting similar effects as described for air temperature also given the widespread distribution of blue ice or rock near all three EAIS ice shelves (Figs. 1d–f, 2d, and 5e, f). Similarly, the availability of incoming solar radiation was shown to contribute to enhanced melting over Roi Baudouin Ice Shelf, EAIS, where katabatic-wind-driven blue ice and firn exposure initiated albedo lowering and enhanced absorption of solar radiation (Lenaerts et al., 2017).

Moreover, we note that supraglacial lake formation on Riiser-Larsen Ice Shelf is dictated by prevailing low wind speeds from a (south-)easterly direction following periods of anomalously high wind speeds from a north-easterly direction (Figs. 6, 8e, S5a, and S7a, b). Considering the detected pattern, we suggest that meltwater ponding on Riiser-Larsen is driven by katabatic wind forcing. A similar pattern can be observed over Nivlisen Ice Shelf (Fig. 8m). Here, shifts in wind direction from low-speed south-easterly winds to strong (south-)easterly katabatic winds preceding high supraglacial lake coverage can be attributed to the observed correlations (Fig. S7c and d, Table S2) and meltwater ponding, e.g. by preserving the nearby blue ice region near the Wohlthat Mountains (Fig. 1) (Horwath et al., 2006). The spatial correlation pattern over Amery Ice Shelf is twofold (Figs. 8u, S5i and S7e, f). Along the north-east grounding line, the correlation pattern revealed that supraglacial lakes peak shortly after occurrence of high-speed (south-)easterly winds, likely reflecting the influence of katabatic winds flowing onto the ice shelf via the adjacent eastern plateau (Stokes et al., 2019). In contrast, supraglacial lakes in the south of Amery Ice Shelf peak during low-speed winds from the south(-east) or south-(west) following high-speed southerly katabatic winds. This indicates that katabatic winds drive melting in agreement with the local topography, determining the direction of wind inflow. Easterly katabatic winds warm adiabatically through vertical mixing and can lead to the exposure of albedo-lowering blue ice and firn through wind scouring, which further enhances surface melt, as observed over other EAIS ice shelves (Arthur et al., 2020a; Lenaerts et al., 2017). In fact, lag 1 correlation with modelled snowmelt events over Riiser-Larsen and Amery ice shelves (Figs. 6 and S5c, d, k, l) could reflect localized katabatic wind-induced peaks in melting.

In addition, the negative correlation with SAM over Riiser-Larsen Ice Shelf (Table 1) could point towards Southern Hemisphere atmospheric modes similarly affecting western East Antarctica, e.g. due to the spatial proximity to the Weddell Sea. In fact, January 2020 lake extents were above average with respect to the overall period 2016–2021, in agreement with a positive DMI and negative SAM (Fig. 10). Correlations with SAM over Nivlisen and Amery ice shelves were weakly positive, potentially reflecting the opposite climate effects SAM can have on the API and EAIS (Kwok and Comiso, 2002; Wachter et al., 2020). As mentioned, also the influence of an atypical ASL driving below-average temperature in 2020–2021 (NSIDC, 2021) has to be considered, most likely explaining the observed low lake extents over EAIS ice shelves in 2020–2021.

To summarize, EAIS supraglacial lake formation in 2016–2021 was mainly controlled by the regional near-surface climate, with air temperature, solar radiation and (katabatic) wind conditions showing the strongest influence. In this context, some of the observed climate variability can be explained by large-scale climatic drivers including SAM, IOD or ASL. Further, local controls including melt–albedo feedbacks, the firn air content and the ice surface topography showed strong influence on EAIS supraglacial lake formation and distribution.

Table 2 provides a summary of all detected drivers for the API and EAIS. Even though each ice shelf showed slightly different driving factors for supraglacial lake formation, local controls as well as regional climate variability with respect to air temperature, solar radiation and wind conditions were found to influence all six investigated regions in a similar manner. This points towards their transferability to other ice sheet regions as well as their role as primary driving factors. Furthermore, the potential influence of large-scale atmospheric modes at continent-wide scale is highlighted through the influence of the intense ASL in 2020–2021 on air temperature and lake formation across all six Antarctic ice shelves. Considering this, similar trends in lake evolution can be expected, e.g. in terms of low lake extents over other Antarctic ice shelves during the 2020–2021 summer melting season. Likewise, also the influence of SAM and IOD was shown to influence adjacent ice sheet sections (Table 2) over the API and EAIS, making it likely that similar lake evolution can be observed in surrounding regions, e.g. in the form of high lake extents during the 2019–2020 melting season over the API. At the same time, it has to be noted that the complex interplay of climate variables might have led to non-linear relations not captured in our correlations, and the comparatively low number of observations might have led to correlations being flagged as insignificant despite the existence of causal links. Further, also the coarse resolution of climate reanalysis data might have affected correlation values particularly at a small scale. Finally, more detailed investigation of teleconnections would require additional atmospheric indices.