SPOT5 is an optical satellite that was operated between May 2002 and March 2015. It was equipped with the HRS sensor that acquired image pairs in panchromatic mode (0.48–0.71 µm). Stereoscopic pairs were acquired in a single pass of the satellite, 90 s apart. The sensor was also characterized by a large 120 km swath and rectangular 5 m × 10 m pixels (along-track x cross-track, ). As part of the SPIRIT program (SPOT5 stereoscopic survey of Polar Ice: Reference Images and Topographies, ), the Antarctic Peninsula was surveyed during the austral summers of 2006, 2007 and 2008 (Fig. ). The data recorded by HRS used 8-bit encoding, which only allows for 256 different gray-levels. This is one of the main limitations of this sensor in glaciological studies where ice and snow surfaces, that are highly reflective, can produce saturated pixels . To help mitigate this effect, during the SPIRIT program, the sensor gains were adapted every week to ensure good contrast over ice sheets surfaces relative to the time of year and latitude. In 2021, all the SPOT5-HRS data archive was made publicly accessible through the CNES Spot World Heritage Program.

The REMA project was created to produce the first gapless DEM of Antarctica at high spatial resolution . Sub-meter stereoimages were acquired by four commercially operated satellites: GeoEye-1 (from 2008, 15.2 km swath), WorldView-1 (from 2007, 17.6 km swath), WorldView-2 (from 2009, 16.4 km swath) and WorldView-3 (from 2014, 13 km swath). For each available stereopair, a DEM was generated at 2 m pixel resolution. Due to the high spatial and radiometric resolution of the stereoscopic images, even areas with low contrast terrains, like snow, ice and shadows, could be measured. Time-stamped DEMs and their associated binary masks (pixel validity) are publicly available at a 2 m pixel resolution . We used REMA DEMs acquired during the austral summers of 2020, 2021 and 2022. These years were selected because they had the most acquisitions over the AP compared to later years.

ICESat was launched by NASA in 2003, primarily to monitor the mass change of the ice sheets . Until 2009, elevation data was collected from the GLAS (Geoscience Laser Altimeter System) instrument over Greenland and Antarctica, with a polar orbit up to 88° north and south. Laser pulses of 1064 nm wavelength (infrared) were used to measure surface heights with an along-track sampling interval of 172 m and a footprint diameter of 60 m. Due to unexpected laser malfunctions, the initially continuous operation mode was converted into 9 campaigns of 8 to 37 d every three to six months . To match the SPOT5-HRS temporal coverage (2006–2008), we selected the high energy campaigns 3D, 3E, 3G, 3H, 3I and 3J (Fig. ). We used the GLA12 Records Release 34 ice sheets elevation product .

Building on the legacy of ICESat, the ICESat-2 mission was designed to extend and improve the elevation measurements from space-borne laser altimetry . The satellite began operations in September 2018 on the same orbit as its predecessor (88° north and south) and is still operational in 2026, with a repeat cycle of 91 d. The Advanced Topographic Laser Altimeter System (ATLAS) carries three pairs of laser beams that measure the surface height, enabling slope effect correction and increasing spatial coverage. The photon counting detectors also improved the along-track spatial sampling to 0.7 m with a footprint diameter of 17 m. For glaciological purposes, we used the ATL06 product Land Ice Elevation at 40 m along track resolution. In line with the epoch of REMA DEMs, we retrieved ICESat-2 data for the austral summers of 2020, 2021 and 2022 (Fig. ).

We used the Ames Stereo Pipeline (ASP, version 3.3.0, ) software developed by NASA to produce DEMs from the optical stereoscopic image pairs. As outlines of the AP vary depending on the dataset considered (up to a few tens of kilometers at the southern limit with west Antarctica), we used the largest dataset extent, SIL20, to identify the image pairs included in the AP. RPC models were attached to the imagery using add_spot_rpc pre-processing routine. We applied the Semi Global Matching algorithm to correlate the image pairs. We changed the following parameters from the default settings, inspired by : –stereo-algorithm 1 –xcorr-threshold 1 –corr-kernel 7 7 –corr-tile-size 4096 –ip-per-tile 800 –subpixel-mode 9. We generated DEMs with a pixel size of 20 × 20 m using point2dem. Before any further processing, we manually removed the DEMs presenting too few data (less than 20 km²) or too much noise (from visual inspection). The austral summers of 2006, 2007 and 2008 were selected as they contain the majority of the HRS acquisitions of the AP. We obtained a total of 481 HRS DEMs acquired on 110 different dates between 21 December 2005 and 22 March 2008, which we grouped per austral summer. The temporal coverage is shown in Fig.  and spatial coverage in Fig. , which contains all geographic information hereafter mentioned. The coverage of each austral summer is available separately in Fig. S1 in the Supplement.

All DEMs were coregistered horizontally and vertically to a common reference DEM following using the Python package xdem (10.5281/zenodo.19636034, ). We created the reference DEM by mosaicking the improved TanDEM-X DEM in the northern AP with the Copernicus GLO-30 DEM edited version of the TanDEM-X DEM in the southern AP. Usually, ice-free terrain of intermediate slope (from 2 to 20°), considered stable, is used to apply the coregistration. However, in the AP, glaciers occupy the vast majority of the surface (>99% of the area) and the remaining areas not covered by ice are often small and very steep . Rock outcrops also become even scarcer in the south of the AP. Therefore, we considered all available elevation data with intermediate slope, including the ice and snow, to perform the coregistration. We identified 20 DEMs with unrealistic horizontal shifts (x- or y-shift larger than median ±3 ⋅ NMAD; normalized median absolute deviation). Three-quarters of these DEMs were successfully re-coregistered with an overlapping HRS DEM, and the remaining ones were kept despite the extreme horizontal shifts, as they provided unique data on flat areas less impacted by a potential coregistration error. We obtained the following mean (μ) and standard deviation (σ) for the x-, y- and z-shifts: μx = -0.67 and σx = 12.20 m, μy = -24.13 m and σy = 12.33 and μz = -3.60 and σz = 5.40 m (Fig. S2).

We removed pixels with steep slopes (≥ 50°) that were likely to have large errors . We then computed the elevation difference to the TanDEM-X reference DEM as a first coarse estimate to filter out outliers (threshold of 200 m). Then, an asymmetric filter was added to remove very high positive elevation change values (threshold of +20 m). To remove remaining aberrant pixels, we applied a custom percentile-based radial outlier filter. For each pixel, we calculated the 20th (h20) and 80th (h80) percentile within a disk of a radius (r). Defining a threshold (α), pixels with an elevation h such that |h-h20|>α or |h-h80|>α were removed. We applied three filters with the following parameters: r = 2 pixels and α = 20 m, r = 5 pixels and α = 50 m, r = 10 pixels and α = 100 m. Details on the effects of each filter can be found on Fig. S4. In total, 16 % of the initial data were removed.

The high vertical accuracy and precision of laser altimetry measurements make them a useful tool for correcting the vertical biases of less accurate DEMs. ICESat elevation measurements have already been used to estimate the accuracy of DEMs (e.g. ). Thus, we harnessed the temporal overlap of ICESat and HRS DEMs (Fig. ) to vertically adjust the latter. The spatial sampling was too sparse to use ICESat elevations for horizontal coregistration, but sufficient for a vertical adjustment.

We preprocessed the GLA12 ICESat data identically to . Then, for each HRS DEM, we selected the closest ICESat campaign in time. This resulted in a 52 d maximum time difference between HRS acquisition dates and the central date of ICESat campaigns. We extracted all intersecting points from the chosen campaign. To compute the elevation difference, we bilinearly interpolated the HRS DEM at the coordinates of the ICESat points. We also extracted the slope from the TanDEM-X reference DEM and removed points with a value ≥ 20° to mitigate the effect of slope-induced errors. We imposed a minimum of 20 valid points to compute the mean elevation difference and vertically shift the HRS DEMs by the obtained value. 388 of the 481 initial DEMs were corrected. The mean vertical adjustment, added to the z-shift obtained from the coregistration, was 1.30 m with a standard deviation of 3.05 m (Fig. S2).

Due to the wide spacing of the ICESat tracks (∼ 20 to 30 km) and the brief duration of the acquisition campaigns, 93 DEMs had no simultaneous intersecting ICESat points. Two alternative corrections were considered to vertically adjust the DEMs. First, DEMs that overlapped with an ICESat-adjusted neighbor were coregistered to that neighbor (44 DEMs). Then, DEMs with no ICESat-corrected neighbors were coregistered to the REMA DEM mosaic (see Sect. ), but only using only ice-free terrain with slopes from 2 to 20° (37 DEMs). Any remaining DEMs were removed from the dataset (12 DEMs) with the exception of two DEMs covering the Clarence and Elephant islands (South Shetland Islands archipelago), which were kept uncorrected as they provide unique information on these isolated areas.

Eventually, for each acquisition date, all intersecting HRS DEMs were mosaicked into a single continuous DEM, taking the mean value of overlapping areas. We thus produced 162 mosaics, called HRS segments, covering 76 % of the AP surface. The distribution of time differences between the HRS segments acquisition date and the ICESat campaign center date weighted by area has a mean of -4.8 d and a standard deviation of 37.8 d. Given the negligible mean time difference, we infer that potential seasonal errors introduced by this correction are mitigated when analyzing the whole dataset.

For each HRS segment, we identified the overlapping REMA DEMs, acquired in the austral summers 2020, 2021 and 2022 (Fig. ). We limited the day-of-year difference to ±62 d (2 months) to avoid seasonal errors. For example, a 12 October 2007 SPOT5-HRS DEM would not be compared to a 12 January 2021 REMA DEM as the seasonal difference (3 months) is out of the 2-month maximum interval. We made an exception for the particular case of Drygalski and Hektoria-Green-Evans (HGE) glacier system (Fig. c and e). Acknowledging that the extreme elevation changes observed in those areas exceed seasonal effects, we extended the day difference to ±150 and ±125 d respectively in order to get a more complete coverage.

We used the binary mask associated to each REMA DEM to keep only valid pixels i.e. pixels with a 0-flag (no water, nor cloud, nor edge). Then, we coregistered REMA DEMs to the same TanDEM-X reference DEM as the HRS segments (Sect. ). REMA DEMs with a x- or y-shift outside of the interval median ± 3 ⋅ NMAD were removed. In total, 3941 DEMs were downloaded, of which 2796 were successfully coregistered. The x-, y- and z-shifts means (μ) and standard deviations (σ) are the following respectively: μx = 25.64 m and σx = 15.66 m, μy = 6.31 m and σy = 15.44 m, μz = -0.43 m and σz = 5.04 m (Fig. S3).

Similarly to the SPOT5-HRS DEMs processing, we vertically adjusted the REMA DEMs using near-coincident laser altimetry data, this time from ICESat-2. For 2298 REMA DEMs, we extracted intersecting ICESat-2 data acquired ±31 d apart. We extended the time window to ±62 d for 227 additional DEMs. All ICESat-2 points with a quality flag equal to 0 and a slope value ≤ 20° extracted from the TanDEM-X reference DEM were selected. We set the same threshold of 20 points at minimum as for the SPOT5-HRS DEMs adjustment on ICESat. Then, we bilinearly interpolated the DEMs at each ICESat-2 point and computed the mean elevation difference that was used as a vertical correction. We removed REMA DEMs with a vertical shift larger than 3 NMAD from the median. The mean vertical shift is 0.69 m and the standard deviation is 13.24 m (Fig. S3). We combined all the processed REMA DEMs into a mean mosaic that was used as a stable terrain reference DEM (Sect. ), covering 90 % of the AP surface.

We derived all pairwise elevation difference maps (dh maps) possible between the two elevation datasets SPOT5-HRS and REMA, respecting the maximum 62 day-of-year difference, and obtained 8381 dh maps. We converted the elevation differences to rates of elevation change (dh / dt maps) by dividing the dh map by the time difference between the two austral summers considered (between 12 to 16 years). To remove remaining aberrant values, we first applied a spatial filter to all the dh / dt maps, inspired by . For each pixel, the 20th (dh / dt20) and 80th (dh / dt80) percentile were calculated within a disk of a radius (r). We set the radius to 5 pixels and the threshold (β) to 0.5 m a⁻¹. Pixels with a dh / dt value larger than dh / dt80+β or lower than dh / dt20-β were excluded. Then, we implemented a filter based on the binning of dh / dt values against elevations extracted from the TanDEM-X reference DEM, independently for each catchment from the SIL20 outlines. The latter is called hypsometric signal and represents the relationship between elevation change rates and elevations at catchment scale. To extract a clear hypsometric signal, we duplicated the dh / dt maps dataset on which we applied a more restrictive spatial filter (radius of 20 pixels and threshold of 1 m a⁻¹). For each catchment, we collected all this strictly spatially filtered dh / dt data and distributed it according to 50 m elevation bins. We computed the mean and standard deviation of each bin and retained values within mean ± 7 ⋅ std. Non-robust statistics were used to account for the effect of valid but outlying values on the metrics, which is essential for further volume calculations. Then, we applied this hypsometric filter on the less strictly spatially filtered dh / dt maps cropped to the catchment outlines. Figure S5 illustrates the effects of the spatial and hypsometric filters on an example catchment. Eventually, we merged all the dh / dt data available for each catchment. In areas with different dh / dt estimations, we used the median value, less sensitive to potentially remaining outliers. The obtained dataset covers 66,7 % of the total area of the AP and is shown on Fig. .

Based on the observation of the elevation difference rates maps here produced and on the BedMap3 grounding line vector product, we manually edited the grounding line (marine terminating glaciers) or terminus (land-terminating glaciers) of 61 catchments from SIL20 outlines. This enabled to more accurately delineate grounded ice before further calculations. We also separated the SIL20 catchments into the peripheral glaciers and the APIS. The two domains defined are very similar to the RGI 7 (peripheral glaciers) and to the IMBIE and ZWA12 outlines (APIS). This, and the fact that the outlines of SIL20 were manually corrected, consolidated our choice of using them as primary outlines.

The HRS segments and REMA DEMs, even if close to providing complete coverage when aggregated (76 % and 90 % coverage, respectively), do not support a dh / dt dataset covering the entire AP. As complete dh / dt maps are required for further volume and mass changes computations, we included a spatial interpolation step. To add elevation change information where dh / dt map were missing, we extracted elevation data from ICESat-2 matching the same time constraints as for the REMA DEMs extraction (±62 d apart from the HRS DEM acquisition date) for each HRS segment. We computed the elevation difference rate dh/dtIS2 between HRS and ICESat-2. We grouped all the data per catchment and removed pixels outside of mean ± 7 ⋅ std.

We then focused on hypsometric methods , using first a local hypsometric method to fill the uncovered areas. For each catchment, we calculated again the hypsometric signal per 50 m elevation bins. To do so, we averaged the hypsometric signals coming from: (1) the dh / dt maps between REMA and HRS DEMs, (2) the dh / dtIS2 points between HRS DEMs and ICESat-2 data. We extracted the elevation of empty pixels from the TanDEM-X reference DEM used previously. Eventually, we considered the mean dh / dt value of the corresponding elevation bin to fill the missing data. For 30 catchments representing 3 % of the total study area, no dh / dt data was available. We applied a regional hypsometric method to estimate the elevation change on those catchments: we extracted an AP-scale hypsometric signal considering all dh / dt data available that was used to fill the last voids (see Fig. a).

We compared our dh/dt maps to those from previous studies, on the APIS (Fig. ) and on the peripheral glaciers (Fig.  ), selecting the closest time period available when possible. These studies are based on different acquisition techniques, thus presenting different coverages, noise and spatial resolutions. They also cover different time periods, which durations range between 4 to 15 years. This spatial comparison highlights the differences and similarities between the studies but also the advantages and drawbacks of each technique.

Focusing on the APIS (Fig. ), we compared our results to the three following studies: (Synthetic Aperture Radar or SAR), (laser altimetry) and (optical stereoscopic imagery). Over the APIS, the glaciers presenting extreme thinning rates, such as Drygalski Glacier and HGE glacier system, are clearly visible in all studies (Fig. a–d). Similarly, the spatial pattern of intense thinning of a large part of the Fleming Glacier and slight thickening in the upper part of the catchment is observed by all three studies that cover it (Fig. e–g). However, many differences in the observed patterns arise, underlying the difficulties inherent in each method. First, the level of noise is higher in the two studies covering a short period (≤ 5 years), and . Our study benefits from the long time period of study (14 years), which enables us to largely reduce the noise, especially in the close-to-balance areas. Then, regarding the laser altimetry study , the 5 km resolution added to the interpolation required to obtain gridded data leads to very smooth maps. A consequence is that the mass loss is not well localized, with negative dh/dt values leaking outside the high-loss glacier catchments. Eventually, the dh/dt dataset based on Synthetic Aperture Radar measurements presents many unrealistic patterns such as the intense thinning at the highest elevations of the western outlet glaciers at the latitude of Boydell and Sjoegren Glaciers, not observed by the other studies (Fig. b). This is likely linked to the complex correction of firn penetration biases, exacerbated by the short period of study. dt/dt maps seem to include a systematic bias in high elevation areas (Fig. d), which is hard to correct because of the scarce stable areas in the region.

Over the peripheral glaciers (Fig. ), this study was compared to three other dh/dt datasets: (laser altimetry), (swath interferometric radar altimetry) and (optical stereoscopic imagery). These four studies agree on the spatial changes of several areas, such as the negative signal over the Snow hill Island (Fig. a–d), the thinning of Hampton Glacier on Alexander Island and the thickening over the western coast of the Rothschild Island (Fig. i–l). The thinning outlet glaciers from James Ross Island are identified by all studies except the laser altimetry one , likely a result of its too coarse resolution. and are affected by a high level of noise, mostly on the high elevation areas and at the coast, respectively. The central part of the Alexander Island tip is problematic for all four studies, in terms of interpolation artifacts, missing data and noise. Both and provide nearly complete maps, whereas our study and present large missing areas, in particular in the central parts of the islands. The peripheral glaciers, through their reduced size in comparison to the APIS, challenge the capabilities of each measurement technique.

The spatial comparison detailed previously also emphasizes the complementarity between the methods, in particular between the laser altimetry and the optical stereoscopic imagery, which both capture surface elevations without firn penetration biases. Satellite laser altimetry measurements, following orbital tracks of variable spacing, require interpolation methods to obtain smooth dh/dt maps at a spatial resolution of 1–5 km. Over wide homogeneous areas, e.g., the ice sheet areas outside of the northern APIS, this is very suitable and the high vertical accuracy allows one to precisely measure small elevation changes. This is exactly where optical stereoscopic methods can struggle, producing many artifacts, outliers or simply data gaps. However, intense small-scale dh/dt signals can be missed by sparse laser altimetry measurements. On Fig. , the strongly negative dh/dt signal from Sjoegren and Boydell Glaciers is present in the three high-resolution datasets (this study, 30 m; , 30 m, , 50 m), but not in the 5 km one . Likewise, the localized thinning of Crane, Flask and Leppard Glaciers is completely absent from this dataset. Additionally, when the signal is sufficiently wide to be observed by laser altimetry, the elevation change measurements are still lower (in absolute value) than those observed by high-resolution methods. For example, the HGE glacier system elevation change signal is less negative and more spread out on Fig. b than on the other panels a, c and d. Furthermore, the laser altimetry can be very sensitive to outliers that sometimes propagate, as around the Getman Ice Piedmont Glacier (Fig. g) or on Alexander Island (Fig.  j). On the other side, optical stereoscopy allows the observation of very local changes at high spatial resolution. Still, it requires costly data processing and storage, and remains less efficient than laser altimetry on homogeneous areas, in terms of data gaps and vertical accuracy. This illustrates the potential of combining optical stereoscopic dh/dt maps and laser altimetry to observe ice sheet and glacier elevation change at small scale as well as over wide homogeneous areas.

The Larsen B embayment is located on the northeast coast of the AP. It hosted the Larsen B Ice Shelf that was fed by several tributary glaciers until the ice shelf broke up in 2002. This event triggered an intense thinning and acceleration of all tributary glaciers (e.g. Hektoria, Green, Evans, Crane) and has been widely analyzed (e.g. ; ). Only a portion of the ice shelf remains, known as the Larsen B remnant or SCAR Inlet Ice Shelf, providing a temporary buttress to Leppard and Flask Glaciers in particular. Other studies have worked on understanding the complex dynamics of the Larsen B Embayment glaciers, showing phases of retreat and re-advance (e.g. , , ). The Larsen B Embayment has been repeatedly identified as the most intensely thinning area of the APIS over the past two decades (e.g. , ), and we confirm this. On Hektoria Glacier, the most negative dhice/dt rates are found between 6 and 16 km upstream from the 2011 glacier's front (before the 2011–2022 readvancement, ), reaching -17.1 m_ice a⁻¹. The same pattern is observed on Green, Evans and Crane Glaciers, where elevation change rates peak a few kilometers away from the glacier front at -13.8 m_ice a⁻¹, -6.7 m_ice a⁻¹ and -5.2 m_ice a⁻¹ respectively. Leppard and Flask are still buttressed by the SCAR Inlet Ice Shelf, and show less negative dhice/dt signals (-3.2 m_ice a⁻¹ and -2.7 m_ice a⁻¹ at minimum). However, the gradual decline in stress caused by the increasing instability of the ice shelf over the period could explain the observed thinning pattern of these two glaciers . Even though the Larsen B tributary glaciers represent only 2 % of the AP area, they contribute significantly to the total AP mass losses. Over 2007–2021, Hektoria Glacier (-2.3 ± 0.2 Gt a⁻¹) and Green Glacier (-1.80 ± 0.1 Gt a⁻¹) are listed among the 30 highest contributors. More generally, the Larsen B tributary glaciers present a total mass change rate of -6.92 ± 0.7 Gt a⁻¹, corresponding to 17 % of AP mass loss. While those mass changes are in continuity with the 2002–2011 estimate of -8.9 Gt a⁻¹ from , the region has by then undergone several major changes as a result of short-term climate trends (e.g. ) that cannot be disentangled in our analysis.

Fleming is a large glacier located on the southwest coast of the AP, historically feeding an ice shelf in Wordie Bay along with other smaller tributary glaciers (Rotz, Airy, Seller and Prospect). The dynamic evolution of Fleming and its neighbor glaciers after the slow disintegration of the ice shelf between 1960 and the late 1990s has been documented by several studies, in terms of velocity, elevation or grounding line changes. reported a decrease of dh/dt comparing the periods 1966–2008 and 2008–2015 over Fleming Glacier. They found elevation change rates of -1.5 m a⁻¹ at the front of the glacier between 1966 and 2008 and minimum elevation change rates of -1.9 m a⁻¹, reached in areas more upstream. Looking at the period 2007–2021, we also obtained elevation change rates much more negative than the Zhao et al.'s first period (1966–2008), -2.7 m a⁻¹ at the glacier front and -6.5 m a⁻¹ at minimum, corroborating the intensified thinning of Fleming glacier over the past decades. The most negative elevation change rates are located a few kilometers upstream from the glaciers' front, closer to the front than reported by over 2011–2014. Our study also shows that Airy glacier presented less negative elevation change rates in comparison to the other glaciers, likely related to a more favorable bedrock geometry. The Fleming catchment, encompassing Fleming, Rotz, Airy, Seller and Prospect glaciers in SIL20 outlines, presented a mass change rate of -5.3 ± 0.2 Gt a⁻¹, second largest contribution in ice mass loss over the period 2007–2021. These significant figures suggest that the tributary glaciers of the former Wordie Bay ice shelf are still responding dynamically to its breakup, and that they have not yet found a new equilibrium state by 2021.

Tidewater glaciers of the northwest AP have presented different dynamic behaviors over the past few decades, which have been predominantly linked to an oceanic forcing. identified two different oceanic circulations controlling the position of glacier's front over multidecadal time scales. In the northern part of the AP's west coast, glaciers are exposed to the cool Bransfield Strait Water (BSW). Further south, from Anvers Island until the north of Alexander Island, the oceanic circulation is dominated by the warm Circumpolar Deep Water (CDW). The CDW presents a particular depth pattern: a cooling trend is observed up to 100 m depth, followed by a warming trend below 200 m depth and beyond. According to , this warming is a primary control of the retreat observed for many glacier's front exposed to the CDW.

Investigating the consequences of the subsurface ocean warming, documented the evolution of Cadman Glacier over the past three decades. Following a period of front stability until the early 2000s, the glacier started to retreat at a moderate rate. It underwent an intense acceleration event in 2018–2019, which was related to the long-term thinning of its ice shelf. In March 2021, the latter collapsed entirely, leading to a 5.2 km-front retreat and contributing to the glacier's acceleration and thinning. report extreme elevation change rates of -20.1 ± 2.6 m a⁻¹ after the 2019 acceleration, which is indeed more negative than our rate of -10 m a⁻¹ over the period 2007–2021. They highlighted that the dynamic thinning affected inland areas up to 8 km away from the 2015 grounding line. We also identify this propagation of negative elevation rates, but the reported value of -3.0 ± 0.5 m a⁻¹ is reached 6 km closer to the grounding line. Their work suggests that thinning intensified towards the end of our study period, which would explain why we observed less negative elevation changes over the entire 2007–2021 period.

Unlike Cadman Glacier, the Funk and Lever neighboring glaciers have shown constant velocities and stable terminus over the past decades . We confirm that between 2007 and 2021, a much more negative elevation change was obtained for Cadman Glacier (-0.89 ± 0.12 m_ice a⁻¹) than for Funk and Lever Glaciers (-0.06 ± 0.08 m_ice a⁻¹ and -0.13 ± 0.10 m_ice a⁻¹, respectively). Moreover, several nearby glaciers presented thickening glacier tongues, including Trooz and Funk Glaciers, which participate in the slightly positive or close-to-zero mean elevation change obtained for these glaciers. postulated that the disparities between the recent dynamic evolution of the Cadman Glacier and the neighboring glaciers are related to the different bathymetry configurations. Following this, the seabed dataset provided by , based on swath multibeam data sets from five national programs, provides some insight into the elevation changes of several glaciers of the northern AP. Shallow trough bathymetry is often associated with low glacier flux in this area . For example, the basin in front of Iliad Glacier, presenting a mildly positive elevation change (+0.22 ± 0.12 m_ice a⁻¹), is notably shallow (< 300 m). Similarly, the positive elevation changes observed on other glaciers of Anvers Island could be related to the shallow bathymetry near the island's eastern coast. The area just offshore of the Jason Peninsula is also very shallow, but glaciers at the very tip of the Peninsula tend to present slightly negative elevation changes. The positive elevation changes of the other glaciers of the Jason Peninsula are likely explained by buttressing from SCAR Inlet Ice Shelf.

Along the Trinity Peninsula and the Davis and Danco Coasts, exposed to the BSW, we observe that glaciers have remained stable or even thickened over the period 2007–2021 (Fig. ), in accordance with the results from and . Further south along the western coast, between Anvers and Adelaide Island, elevation changes are mostly negative, except for a few glaciers close to equilibrium. Adelaide, Renaud and Lavoisier Islands also host some of the most thinning glaciers, which is consistent with the CDW predominance in the southwestern part of the AP. However, we observe no clear relationship between elevation changes and Bedmap3 bathymetry at glacier's front for catchments exposed to either CDW or BSW (not shown). As the widespread ice discharge increase reported by started in ∼ 2019, at the very end of our study period (2007–2021), it might require more time to initiate significant negative elevation changes. This increase could also be the result of the thickening of glacier tongues (Fig. ). Furthermore, the difficulty to constrain bedrock models in many places hinders accurate bathymetry estimates. Large discrepancies appear between the different models (e.g. ), complicating the identification of significant correlations. More broadly, the relationship between elevation changes and bathymetry at the glacier's front is likely to be more complex than simply depth-dependent. Taking other parameters into account, such as the bed geometry and inclination, could be beneficial to better understand the spatial variability of elevation changes over the northwestern AP.

The choice of a GIA or elastic rebound model has minor implications for the final volume changes. The associated corrections systematically represent less than 4 % of the signal, and fall within the observational uncertainty. However, the crustal deformation corrections applied in the study (GIA and ELR) do not account for recent viscoelastic response, which might be required to fit with observed bedrock uplift rates (; ). The misfit induced could exceed the GIA and ELR signals in the northern Peninsula. Even if the signal amplitude remains lower than the firn-related uncertainty in this region (Fig. S6), future studies should consider applying a viscoelastic correction at the AP scale in order to account more accurately for crustal deformation.

On the other hand, the four FDMs used in this study lead to significantly different estimates, both in terms of spatial patterns and numerical values. To get a sense of the FAC correction applied, we calculated the mean density corresponding to the volume changes observed. We obtained a density of 878 ± 62 kg m⁻³ (1σ spread between the models) for the APIS and 843 ± 50 kg m⁻³ for the peripheral glaciers. This number falls into the uncertainty bar of the density conversion factor commonly used by the glacier and ice cap community, 850 ± 60 kg m⁻³ from . However, at the AP scale, the FAC correction induces the largest uncertainty (Fig. S13). It accounts for up to 20 % of the signal on some glaciers, with a large spread between the models over the AP, reaching 0.08 m a⁻¹ at maximum. Given the 14-year study period, which allows for limited densification and therefore limits the impact of the model densification choices, the likely sources of differences in FAC change are the forcing (as described in ), surface parameterizations (e.g. surface snow density), and potentially the treatment of liquid water from melt and rain. The spatial resolution, ranging from 10 km (GEMB-FDM) to 27.5 km (MAR-FDM) may also affect the ability of the models to resolve small-scale spatial patterns, such as the East-West gradient across the AP. This is particularly visible in the northern AP, where the standard deviation between the models reaches its maximum (Fig. S6). The AP is a difficult place to validate climate and FDM models due to the scarcity of in situ measurements. The period selected to compute the FAC correction (start, end or entire summer) also affects the results at orders of magnitude similar to the spread between the models. Thus, the conversion from volume to mass, applied through the FAC correction, constitutes one of the major limitations of this method. A better understanding of firn processes and firn modeling is therefore a priority to improve the DEM-differencing method for glaciers and ice sheets mass changes studies .

As a third of AP area is not covered by observation data, interpolated data represent a non-negligible part of the final estimates. Over large regions of the APIS, the interpolation represents the second source of uncertainty (Fig. S13). For example, the mass changes obtained for Region 24 using the ZWA12 outlines (-8.0 ± 3.4 Gt a⁻¹, Fig. a) includes a large uncertainty that is essentially caused by the partial coverage of dh/dt maps (57 %). The local hypsometric interpolation performs well in smooth high altitude areas but struggles more reproducing local patterns of intense thinning at lower elevations (e.g. Drygalski or Fleming Glaciers). If some of the contrasting dh / dh spatial patterns do come from observations (e.g. Stefan Ice Piedmont and Peter Glaciers, Cadman Glacier, Fig. 7), the interpolation can artificially generate contrasted values, as in the south of Anvers Island (Fig. 7). Glaciers' mean ice elevation changes must therefore be interpreted with caution, especially when dh/dt observations are scarce. We suggest the interpolation could be improved by taking into account more input data to explain the spatial variability of the dh/dt signal (e.g. velocity acceleration). As mentioned in Sect. , another alternative to fill data gaps could be to combine laser altimetry maps, which may allow to better reproduce the spatial patterns.

The grounding line is a physical delimitation difficult to observe, both with in situ and satellite measurements . Its location depends on many poorly constrained factors, including bedrock geometry and ice thickness. AP-scale grounding line datasets are scarce and not available at an annual resolution. Nevertheless, rapid fluctuations, including frequent retreat of the grounding line, are induced by glacier dynamics at sub-annual scales. In many areas, the position of the grounding line has not yet been agreed upon (e.g. Hektoria Glacier, ). Here, considering a 14-year period complicates the choice of grounding line dataset. The grounding line from the first epoch may encompass areas that subsequently became ungrounded, while the one from the second epoch could miss former grounded ice changes. Cadman Glacier illustrates this dilemma, exhibiting pronounced differences between the grounding line datasets used in this study. However, we emphasize that a limited number of AP's glaciers are prone to those grounding line disparities. With the SIL20 outline, Cadman Glacier (Fig. ) presented mass changes of -0.27 ± 0.02 Gt a⁻¹ over a total area of 332 km². The Bedmap 3 grounding line , located ∼ 5 km more inland, identifies the lowest portion of the glacier as the ice shelf. This area, covering 15 km², lost 0.03 Gt a⁻¹, which represents ∼ 10 % of the SIL20 outlines estimate. At the scale of the AP, the mass changes caused by grounding line differences are negligible. Nevertheless, catchment-scale studies might require a more accurate grounding line, including knowledge of its migration.

As each epoch extends over three years, we cannot exclude potential influence of inter annual variations that affect the mass change results, such as the strongly positive surface mass balance in 2022 . Still, we tried to reduce the seasonal influence by establishing strict constraints on the seasonal time difference between the acquisitions. Our elevation change maps are based on the comparison of DEMs over two epochs and hence does not allow us to observe and analyze elevation changes throughout the period. For example, potential variations of the elevation difference rates trends or specific events happening during the time period cannot be isolated in this study. This is quite clear when comparing the results from and , that cover two distinct time periods and show contrasted elevation changes over the fast-changing glaciers. The extreme changes that Hektoria Glacier underwent over the two past decades (slow retreat between 2002 and 2011, followed by a readvancement until the 1.5-year extreme retreat starting in 2022, and ) cannot be identified either. This is also the case for many other physical processes presenting a strong inter annual variability (e.g. summer melt, ). However, a key advantage of our large study period is that it greatly reduces the noise, especially at high elevations.