This study uses brightness temperature from SSMIS on the DMSP-F17 satellite over the hydrological years 2011–2021. DMSP-F17 was selected for its sun-synchronous, dawn–dusk orbit stability, which provides two consistent Antarctic overpasses per day at approximately 06:00 LT (hereafter “M”, morning observation) and 18:00 LT (hereafter “E”, evening observation) local time (https://www.remss.com/missions/ssmi/, last access: 1 December 2025). All brightness temperatures were obtained from the National Snow and Ice Data Center (NSIDC) (https://nsidc.org/data/nsidc-0630/versions/2, last access: 6 June 2025) and preprocessed in Google Earth Engine . Our analysis concentrates on the H polarized 19 GHz channel, offering a 6.25km×6.25km enhanced footprint – the finest available at this frequency . This channel is widely used for melt detection because 19 GHz is sensitive to small amounts of liquid water while still penetrating into dry firn, yielding low brightness temperatures under dry-snow conditions and a marked increase when liquid water is present . We also evaluated the 37 and 91 GHz channels at their enhanced resolutions (3.125 km), but these higher-frequency channels, characterised by much shallower penetration depths did not provide a consistent improvement in our methodology for melt detection.

AWS observations are the foundation for the method in this paper. For the melt volume to be calculated, only AWSs that measure sufficient variables to close the surface energy balance qualify. This grossly reduces the number of available AWS locations, since the full radiation budget is only measured at a handful of stations in Antarctica. A major provider of data for this study is the Institute for Marine and Atmospheric Research Utrecht (IMAU) AWS dataset, which is described in . Only the IMAU AWSs with at least one entire hydrological year of data within the June 2011 – May 2021 window were used for calibration and evaluation. These comprise AWS11 (Halvfarryggen Ice Rise), AWS14 (northern Larsen C ice shelf), AWS15 (central Larsen C ice shelf), AWS16 (Princess Elisabeth station), AWS17 (Scar Inlet as a remnant of Larsen B ice shelf) and AWS18 (Cabinet Inlet on western Larsen C ice shelf) (Fig. S1 in the Supplement). All six sites record the standard meteorological variables and the four components of net surface radiation, with measurements corrected for common errors as detailed in . Melt volumes are subsequently computed at each station using the SEB model of . In this framework, turbulent fluxes are calculated using similarity theory, surface temperature is determined via iterative closure of the SEB, and excess energy at 0°C is converted into meltwater. Meltwater percolates through the firn using a bucket scheme until refreezing occurs. Shortwave radiation penetration into the subsurface layers of the snowpack is neglected. Of the six IMAU AWS stations meeting our requirements, four (AWS14, AWS15, AWS17, and AWS18) are situated on or immediately adjacent to Larsen C ice shelf, whereas the remaining two (AWS11 and AWS16) provide a few years of measurements in locations with lower melt. We augment the dataset with a decade (2011–2021) of measurements from the German Neumayer station. Although Neumayer also exhibits generally low melt rates, its continuous and long-term record substantially strengthens the calibration dataset and introduces a well-sampled coastal East Antarctic climate distinct from the high-melt conditions of Larsen C. At Neumayer, we use the surface radiation observations from the Baseline Surface Radiation Network (BSRN) station , meteorological observations , and surface height observations .

For additional analysis, we also use observations of near-surface air temperature scaled to a nominal height of 2 m above the surface.

Modelling surface melt in an SEB model carries uncertainties because of model settings, model assumptions, and errors in the input. This uncertainty is estimated using a number of sensitivity tests. First, the uncertainty from the IMAU AWS forcing is estimated by separately including or removing one of four measurement corrections: the window heating of the pyrgeometer, the shortwave heating of the passively ventilated temperature sensor, the correction for relative humidity for ice and sensor sensitivity at very low temperatures, and the correction for tilt and bias of the pyranometer, which are all described by . To constrain the uncertainty associated with the SEB model, five different model settings were individually adjusted: (i) the sensor height fixed to 2 m above the surface instead of varying in time, (ii) the roughness length for momentum increased from 0.1 to 1 mm, (iii) the surface longwave emissivity decreased from 1 to 0.97, (iv) the snow thermal conductivity parameterised after instead of , and (v) allowing the snow height to freely evolve in the model instead of being prescribed by surface height observations. These choices result in one reference and nine perturbed time series of SEB components and surface melt per IMAU station, where each perturbed timeseries results from just one omitted measurement correction or one different model parameter at the time. This sensitivity analysis was only conducted for the AWS where the uncertainty of the observations and of the SEB model are both expected to impact the melt volume computations. These are the AWS that are left unattended for a year or more and located in areas with substantial melt, namely AWS14, AWS15, AWS17, and AWS18.

To assess whether the 6.25km×6.25km resolution of an SSMIS pixel is sufficient to represent melt conditions at each AWS site, we compared it against the higher-resolution U Melt binary melt product , available at 500 m spatial resolution. For each station, a 13×13 grid of U Melt pixels was centered over the AWS location, and the melt/no-melt state of all surrounding pixels was compared to that of the central pixel for all days, including both melt and non-melt days.

Two metrics were computed: (i) homogeneity rate, defined as the fraction of surrounding pixels with melt flags matching the central pixel, which exceeded 98 % at all stations; and (ii) local variability, defined as the standard deviation of binary melt values within the window, which remained below 0.02.

These results indicate that, around each AWS, the nature of melt conditions is highly homogeneous at a scale similar to that of the SSMIS pixel footprint. Therefore, we conclude that the 6.25km×6.25km resolution of the SSMIS pixel is sufficient to represent local melt conditions and is appropriate for calibration purposes.

To translate SSMIS brightness temperatures (Tb) into surface melt-day detections, we assembled a suite of candidate indicators drawn from established microwave-based methods and calibrated each against in situ AWS melt observations (≥0.5 mmw.e.d-1). This threshold was applied to avoid labelling negligible melt amounts, often within the numerical noise of SEB-derived melt estimates, as true melt events, since very small daily values may reflect model uncertainty rather than physically meaningful surface melt.

All metrics were computed at 19, 37, and 91 GHz, using both horizontal (H) and vertical (V) polarizations. The indicators were grouped as follows (see Table S1 in the Supplement for a detailed description of all candidate variables):

Pure Brightness-Temperature: We tested absolute Tb at each frequency and polarization, for both morning and evening observations.

The second major step in the melt volume estimate is to relate the annual number of melt days (m) to total annual melt (M). To that end, we fitted the AWS-derived decadal record (2011–2021) to an exponential model: 1M=a(ebm-1), where parameters a and b were estimated using a least-squares approach to minimize the residuals between the model and the observed melt values (Fig. a). The above functional form follows an empirically demonstrated non-linear relationship between melt days and meltwater production . This non-linear behaviour likely reflects melt-albedo feedbacks, and the longer time required for refreezing of larger melt volumes, such that warmer summers produce disproportionately more runoff . By fitting a and b from Eq. () to AWS stations, which compute melt via a full SEB model, our approach embeds these physical feedbacks into the SSMIS-derived, AWS-calibrated framework.

When the exponential model is applied to the satellite-derived melt-day count, a pixel-level estimate of total annual melt is obtained. A Monte Carlo-based confidence interval for the m–M relationship is derived by propagating measurement and model uncertainties (detailed in Sect. , see Fig. a): for each AWS-year combination, ten m–M pairs corresponding to distinct setups are available, and in each of 1000 Monte Carlo iterations one setup is randomly selected for each AWS-year, yielding n data points. The exponential model () is then fitted to each sample, producing 1000 realizations of M(m) which are evaluated over m∈[0,100] to characterize the variability of melt estimates. The light pink band in Fig. a represents the 3σ confidence envelope, the blue line denotes the median-fit relationship, and the red line corresponds to the fit obtained using the reference setup alone.

Fig. S4a in the Supplement shows the site-specific exponential fits at each of the four AWS locations (AWS14, AWS15, AWS17, AWS18) where the sensitivity analysis was conducted (see Sect. ); Fig. S4b in the Supplement presents the combined fit across the selected four AWS stations, illustrating how the ten SEB-model permutations produce a modest spread in the resulting m–M curves.

For an independent assessment of the m–M relation, it was also derived for fully independent, model-only, RACMO2.4 melt-day and melt-volume output for 2011–2021, both across the entire Antarctic domain (Fig. b) and separately at four selected AWS locations (see Fig. S4c in the Supplement). In both cases, the resulting exponential parameters and curve shape closely matched those derived from the AWS-SEB calibration, demonstrating the robustness and spatial generality of the m–M relationship. This also demonstrates that the collection of AWS observations used for this study sufficiently captures the variability in surface melt conditions across the Antarctic Ice Sheet as represented by a physically-based model. The agreement in functional shape, despite the melt days underestimation by SSMIS, supports the application of the AWS-derived fit to satellite-derived melt-day counts across the full Antarctic dataset.

Applying the fit described in Sect. , we produced estimates of total annual melt across Antarctica (Fig. a). We assessed our results by comparing annual SSMIS-derived melt days and melt fluxes with coincident AWS-based observations, yielding a strong linear correlation (R2=0.91 and R2=0.83, respectively; Fig. b). However, given the limited number of in situ AWS sites – which were also employed during calibration – this evaluation is inherently circular. Dividing the AWS record into independent calibration and validation subsets was considered not feasible due to the small sample size and the constrained spatial variability of the available stations.

We compare our ten-year decadal-mean melt-flux estimates from SSMIS with two independent products:

QuikSCAT (1999–2009). decadal-mean annual melt flux derived from Ku-band backscatter at 4.45 km resolution , see Fig. b.

Across the Antarctic Peninsula, all three datasets show consistently high decadal-mean surface melt rates. On Larsen C Ice Shelf, SSMIS, QuikSCAT, and RACMO2.4 all exceed 350 mmw.e.yr-1. SSMIS and QuikSCAT place their highest decadal-mean melt values along the western inlets (e.g. Mill Inlet), whereas RACMO2.4 shifts its maximum eastward toward Scar Inlet, a spatial offset also noted in earlier satellite-based analyses . Farther south, on Wilkins and George VI ice shelves, decadal-mean melt rates exceed 200–250 mmw.e.yr-1 in all datasets.

Along coastal West Antarctica, including the Amundsen and Ross Sea sectors, decadal-mean melt rates remain low, around 20–30 mmw.e.yr-1 in all products. These areas represent some of the lowest-melt regions outside the high-elevation interior. Over the Ross Ice Shelf, SSMIS and QuikSCAT show the strongest decadal-mean melt along the western flank, whereas RACMO2.4 simulates higher melt along the eastern margin.

In East Antarctica, the three products again show broadly consistent patterns. Decadal-mean surface melt rates of around 200 mmw.e.yr-1 occur on Roi Baudouin and the inner Fimbul ice shelves, while northeast Amery Ice Shelf shows rates near 150 mmw.e.yr-1 across all products.

To allow a direct comparison with QuikSCAT, we also extracted RACMO2.4 outputs for 1999–2009 (Fig. d). Over this shared period, the two datasets show comparable decadal-mean melt magnitudes (within ∼10 %) and similar spatial patterns, including RACMO2.4's modest eastward displacement of melt maxima and QuikSCAT's tendency to underestimate melt in low-intensity coastal regions. These agreements and discrepancies are consistent with those observed between SSMIS and RACMO2.4 for 2011–2021.

Interannual melt volumes from SSMIS and RACMO2.4 over 2011–2021 exhibit similar temporal patterns. The mean annual Antarctic melt volume is approximately 85 Gtyr-1 for SSMIS and 100 Gtyr-1 for RACMO2.4 (Fig. S5 in the Supplement). The corresponding annual mean melt-flux maps (Fig. S6 in the Supplement) further demonstrate the close spatial and temporal agreement between the two datasets, and additionally provide a regional view of the Antarctic Peninsula, where most Antarctic surface melt occurs.