The Aneto glacier is the largest glacier in the Pyrenees (48.1 ha (0.48 km2) in 2022), a mountain range where only four glaciers are larger than 10 ha. It is located in the Maladeta massif (Fig. 1), on the northeast (NE) side, between the Maldito (3354 m a.s.l.) and Aneto (3404 m a.s.l.) peaks. The high elevation of this massif, with more than 40 peaks above 3000 m a.s.l., has allowed the preservation of other smaller glaciers (Eastern Maladeta and Tempestades) and ice patches (Western Maladeta, Coronas and Barrancs) in the area. In 2022, the Aneto glacier consisted of two bodies whose glacier front was at 3026 m a.s.l. in the case of the main body and at 3170 m a.s.l. in the case of the secondary body.

In this area, the 0 ∘C mean annual isotherm ranges from 2700 to 3000 m a.s.l. (Jomelli et al., 2020), and the mean annual precipitation is about 2000 mm, with winter and spring being the wettest seasons (Buisan et al., 2015). The mean annual temperature for the period 2007–2022 was 4.6 ∘C at the weather station of the Renclusa hut (2140 m a.s.l.); meanwhile the mean temperature for the same period in the ablation season (June–September) was 11.6 ∘C. The year 2022 was an especially warm year, in which the annual mean temperature was 5.2 ∘C and the summer mean temperature was 12.1 ∘C (data from the AEMET database).

GPR uses the transmission and rebound of electromagnetic pulses at different frequencies to determine ice thickness and glacier interfaces (rocks, bedrock basin, snow, etc.) (del Rio et al., 2014). Different works have studied the variation in ice thickness, surface area or volume on glaciers using different techniques, which highlight the importance of the methodology to be applied in each case, considering its scope and limitations (Procházková, 2019; Bohleber et al., 2017; Marcer et al., 2017; Fischer, 2009).

GPR fieldwork was conducted on 25–26 July 2020, using a Malå Geoscience radar system consisting of a ProEx control unit and a 100 MHz rough terrain antenna (RTA). Occasionally several transects were also carried out with the 100 MHz shielded antenna (see Supplement). Georeferenced radargrams were created using the AtlasLink GNSS smart GPS antenna connected to the GPR, which were obtained in “time” tuning. A total of 32 georeferenced radargrams were recorded in the main glacier body in a common offset mode, corresponding to a length of 6.8 km and covering almost the entire glacier surface (more detailed information can be found in Fig. S1 in the Supplement). The campaign was conducted during a period when the glacier surface was covered with snow, in order to allow safe displacement of the instrument and operators, thus hampering the observation of deeper ice layers. This required differentiation of the snow layer in post-processing to accurately quantify glacier thickness.

Radargrams were processed using Reflexw version 9.1.3 (Sandmeier scientific software), with the following workflow: (1) the adjustment of the time origin (t=0) to coincide with the arrival of the first surface signal on the glacier; (2) the homogenisation of the trace increment, since the acquisition of the radargrams with the RTA antenna was done in time mode and varied in each radargram depending on the speed of the movement of the antenna on the ground (0.1 m ns-1 was fixed, since this was the smallest value obtained in the radargrams); (3) the removal of the background; (4) the correction of the energy loss of the signal when penetrating the terrain by applying a gain factor of 0.2 (energy decay); and (5) the application of a frequency bandpass filter so that only signals with frequencies between 50 and 200 MHz remain (the nominal frequency of the antenna is 100 MHz).

As a first approximation, 0.17 m ns-1 was set as the propagation velocity of the waves in the glacier to get a first idea of the thickness of the snow and ice layers in the radargram representation. Snow and ice layers must be defined from the radargrams to create a thickness model of both. To do this, the wave propagation velocities (RWVs) in both media must be available beforehand. In a similar study on the Monte Perdido glacier, RWVs of 0.200±0.005 m ns-1 for snow and 0.163±0.007 m ns-1 for ice were obtained for the 500 and 200 MHz antennas, respectively (López Moreno et al., 2019). The coherence of these velocities was checked in the 1054 radargram at the points where diffraction hyperbolas occurred (plot of diffraction hyperbolas is shown in Fig. S2).

The distribution of the GPR data does not follow a homogeneous pattern; the GPR record tracks were distributed along parallel and perpendicular lines, forming an irregular grid (Fig. S1). Therefore, to determine the thickness of the glacier over its entire extent, an interpolation method is required. For this type of data, the interpolation method used was the radial basis function (RBF), as Otero-García (2008) recommended. Given the poor distribution of the data, after several tests, the best method is to work with 16 neighbours, two per octant (the closest points in each direction), in a circular area with a radius of 457.62 m, in the same way, again, as Otero-García (2008). The thickness for glacier limits in 2020 was established as 0 m. To validate this interpolation method, the data were divided into two groups: training with 70 % of the sample and test with the other 30 %.

The delineation of the Aneto glacier surface was done manually (Table S4 in the Supplement) in a GIS software (ArcGIS), considering: (1) the orthomosaic of the historical aerial imagery from 1981; (2) a RapidEye satellite image from 2011 and improved outlines from RGI (RGI Consortium, 2017); and (3) the orthomosaics derived from UAV flights in 2020, 2021, and 2022. Due to the small extent of these very small glaciers, the slope was considered in the calculation of glacier surface to obtain the true glacier area (3D surface) rather than the 2D projection of glacier extent. This calculation is justified because the glaciers are strongly bound to wall cirques, and these had a steepness of 24.3∘ in 2020. When the slope is not taken into account, the glacier surface is underestimated (Vidaller et al., 2021). Otherwise the 2D area computation would also be affected by the changes in slope during the study period.

Data from DEMs available for this work varied in accuracy. The most accurate geolocation is that of the UAV, which was used as a reference for the point cloud due to the post-processed kinematic (PPK) GPS geolocation technique (geolocation RMSE <0.05 m). This geolocation error is equivalent for the 2020, 2021 and 2022 point clouds (0.019 for 2020, 0.025 for 2021 and 0.021 for 2022; the differences were due to weather conditions). Based on the low magnitude of these geolocation errors, we assume that the error introduced in ice thickness differences is nearly negligible. 3D point cloud differences in ice-free areas had RMSEs below 0.02 m, (error computed following Vidaller et al.'s (2021) accuracy method). To coregister the lidar point cloud (2011) and the point cloud from the historical aerial imagery (1981), several areas of stable terrain such as ridges, peaks, polished surfaces, etc. were selected in these later point clouds and in the 2020 UAV-derived point cloud. These areas were evenly distributed around the glacier. A rotation and translation matrix was calculated for these areas to align (separately) the 1981 and 2011 point clouds with that of 2020 using an ICP algorithm (Rajendra et al., 2014), from CloudCompare software (Girardeau-Montaut, 2016), in the same way as Vidaller et al. (2021). Subsequently, these matrices were applied to the entire point clouds to derive point clouds that were finally coregistered. Glacier thickness loss (normal surface differences; see the Supplement for more information) between these point clouds were computed using the CloudCompare tool M3C2 (James, 2017) to determine the differences (surface perpendicular) between the glacier surfaces observed in different years. Glacier change statistics were derived from this later comparison, calculated over the most recent (and smallest) glacier surface.

Glacier thickness loss was determined by considering only data within the smallest (or more recent) surface of the glacier. When considering the oldest surface, there are zones of the glacier that are not present in the most recent acquisitions, so the ice thickness loss would be underestimated (Vidaller et al., 2021). The mass balance was calculated assuming a density conversion factor of 850±60 kg m-3 (Huss, 2013). Thus, the specific mass balance presented in this study was determined considering the recent surface of the glacier.

With the aim of determining areas of future glacier lake formation, the mountain basal topography was derived from the GPR interpolation and the 2020 UAV acquisition (subtraction of the 2020 glacier surface from the ice thickness interpolation from the GPR). The topographic position index (TPI) is capable of identifying terrain depressions at various search distances (Weiss et al., 2001). From this basal topography, the TPI (de Reu et al., 2013) was derived for 70, 100, 150 and 200 m search distances to describe depression areas that potentially favour future lake formation. This index has previously been used in studies of debris-covered glaciers (Westoby et al., 2020) to determine areas of potential debris accumulation, but as far as the authors are aware, this is the first time this index has been used to determine areas of potential lake formation following the retreat of mountain glaciers. In addition, overdeepenings detected by the TPI were corroborated using the longitudinal GPR radargrams.

GPR ice thickness measurements with a 100 MHz RTA antenna are subject to intrinsic error. Assuming a RWV velocity for ice of 0.163 m ns-1, the λ value is 1.63 m, so the minimum spatial resolution is λ/2=0.815 m. Summing this uncertainty for snow and ice gives a thickness resolution of 1 m for this delineation. Thus, the uncertainty in the determination of the ice layer thickness is 1.8 m.

To check the coherence of the determined thicknesses, a test was performed at all intersections between transects to detect any inconsistencies in the values. At these 28 intersections, the average difference is 1.6±1.6 (σ) m, with some outliers of 3–5 m (Table S2). This value is consistent with the uncertainty associated with RWV velocity and ice layers' delineation (1.8 m). The lengths of the radargrams were determined using Reflexw from the GPS coordinates coupled to the GPR (see Supplement for more details). General GPR uncertainty in ice thickness was determined considering different velocities for temperate ice in the transects (1043, 1062 and 1073). Based on existing literature (Jiménez-Vaquero, 2016; López-Moreno et al., 2019), we assumed 0.2 m ns-1 in the snow and between 0.157 and 0.186 m ns-1 in the ice. With these velocities, mean and maximum ice thickness was determined for each transect (Table S3). As a result, mean ice thickness variation that could be derived from different velocities into the temperate ice would fit in the range of the estimated margin of error band (<1.6 m) and would be smaller than the uncertainties obtained from the differences in thickness at transect crossings (<1.8 m).

To validate the interpolation of glacier thicknesses, 30 % of the GPR data were randomly selected, and the remaining 70 % of the GPR dataset was used for the interpolation (Otero-García, 2008). The mean error between the interpolated thickness and the thickness observed with the GPR was 0.0018 m, and the RMSE was 0.3021 m.

The delineation of glacier boundaries also has some uncertainty due to pixel size, geometric correction, visual identification, and the presence of residual snow or debris cover at the glacier boundaries. The surface uncertainty is 0.048 ha (0.00048 km2) for the Aneto glacier (Vidaller et al., 2021) in the case of the glacier surface of 2011, 2020, 2021 and 2022; the uncertainty error of the 1981 glacier outline is 0.58 ha (0.0058 km2).

The coregistration of point clouds from historical aerial imagery and lidar survey with UAV surveys was tested in a buffer zone around the glacier, always using snow- and ice-free zones in both years of comparison. This means that the comparison of the 1981 and 2020 point clouds was performed in a buffer zone with a 300 % larger extent than the 1981 glacier boundaries (over stable terrain); the coregistration error between the 2011 and 2020 point clouds was determined in the same way. In the first case for the Aneto glacier the RMSE is 0.06 m and in the second case 0.4 m (Vidaller et al., 2021).

Annual surface loss has decreased uniformly over time (-2.2 ha yr-1). However, it must be noted that the relative changes are larger in the latter years, since the losses occurring in the most recent period are measured with respect to a progressively smaller surface. Thus, there has been no recent acceleration in surface loss per year, but the relative surface loss has increased. Oppositely, the rates of glacier thickness loss have increased during the study period (-0.6 m yr-1 from 1981 to 2011 and -1.1 m yr-1 from 2011 to 2022), indicating an acceleration of glacier ice thickness loss, especially in the last decade, and more pronounced in the last 3 years. In terms of specific mass balance (considering only changes at the smallest surface glacier, the most recent year of comparison), the losses are 0.6 m w.e. yr-1 for the period 1981–2022, 0.5 m w.e. yr-1 for the period 1981–2011, 1 m w.e. yr-1 for the period 2011–2022, 1.2 m w.e. yr-1 for the period 2020–2021 and 2.7 m w.e. yr-1 for the period 2021–2022. Based on these results, two inflexion points can be identified, one after 2011 and another after 2020 – in both cases the thickness loss has accelerated sharply. This ice thickness loss is mainly accelerated (among other factors) by the fact that the accumulation zone over the glacier in summer is negligible, especially during very hot summers as in 2022, and the ablation zone covers the entire glacier, as no ELA is observed for some years. Unfortunately, due to the small extent of this glacier no reliable satellite observations of sufficient resolution are available for the Aneto glacier ELA in late summer, and this absence of accumulation area is based on field work observations of UAV operators.

Various studies of other glaciers in the Pyrenees have also shown a continuous increase in glacier thickness and area losses, with a high interannual variability but a clear negative trend over longer time periods. These works focused on the Monte Perdido glacier (López-Moreno et al., 2019), Ossoue glacier (Gascoin and René, 2018), Maladeta glacier (Pastor Argüello, 2013) and La Paul glacier (Rico et al., 2015). Hugonnet et al. (2021) also determined a mean ice thinning of -0.96 m yr-1 for the Pyrenean glaciers for the period 2000–2019. Although this work focuses on the period 1981–2022, the glaciers of the Aneto–Maladeta massif had about 610 ha at the end of the LIA, so they lost about 338 ha from 1850 to 1984 (Rico et al., 2017).

The mean annual specific mass balance values of -0.6 m w.e. yr-1 on the Aneto glacier determined for the period 1981–2022 are similar to those in other studies in the Alps, such as Davaze et al. (2020), who estimated an annual mass balance of -0.7 m w.e. yr-1 from 2000 to 2016 for 239 Alpine glaciers. Similarly, Carturan et al. (2016) determined the mean annual mass balance of nine Italian glaciers from 2004 to 2013, which ranged from -1.8 to -0.8 m w.e. yr-1. This is also supported by other climatic data showing an increase in air temperatures over the past century (Bolch et al., 2012; Rabatel et al., 2013), particularly a sharp rise in temperatures at high elevations and low latitudes (Vuille et al., 2008; Pepin and Mountain Research Initiative EDW Working Group, 2015), accompanied by a shorter duration of seasonal snow cover (Brown and Mote, 2009).

Vidaller et al. (2021) describe the changes in ice thickness of the Aneto glacier (among other glaciers of the Pyrenees) based on an ice thickness decrease of 8.5 m during the period 2011–2020. Based on the ice thickness reconstruction data of the Aneto glacier presented in this study, the mean ice thickness in 2020 was 15.0 m, while in 2011 it was 19.2 m, so the loss is 4.2 m. This difference is due to the fact that the mean ice thickness of 2011 was calculated based on the extent of 2011, and the ice thickness of 2020 was calculated based on the area of 2020, while in the case of Vidaller et al. (2021) the ice thickness loss was calculated considering only the ice thickness loss within the glacier area of 2020. A similar problem exists when comparing the remaining ice thickness in 1981 (32.9 m) and in 2022 (11.9 m) with the ice thickness losses for the period 1981–2022 (30.5 m). The remaining ice thickness in 1981 is similar to those losses calculated for the period 1981–2022; meanwhile the remaining average ice thickness was 11.9 m. The mean ice thickness for a particular year was calculated based on the extent observed for that year.

Remote sensing techniques have developed rapidly in recent years, allowing observation of the Earth's surface with a spatial resolution that was previously impossible. This work exploits historical aerial photographs to reconstruct a digital surface model for the year 1981 and provides a comparison to observe changes in landscapes and surfaces in detail.

Campos et al. (2021) calculated changes in the Aneto glacier from the LIA to 2017 using data from 1957, 1983, 2000, 2006, 2015 and 2017. In 1983, they reported an area of 103.2 ha (1.03 km2), in contrast to the 135.7 ha (1.36 km2) for 1981 described in this work. The large difference may be due in part to the fact that they did not consider the slope angle of the terrain in their calculations (2D vs. 3D surface). Nonetheless, considering our delineation, but ignoring the effect of slope angle on the area estimate, we would have reported a value of 115.5 ha (1.16 km2) for 1981, which underestimates our value by 20 %. This study also uses the National Fly photograms to convert to point clouds, accounting for stable GCP during the study period. This is a more accurate method because it avoids distortion of the Plan Nacional de Ortofotografía Aérea (PNOA) orthophotos used by Campos et al. (2021), who acknowledge a source of uncertainty: “The extension for the 1983 stage should be considered with caution. Due to the lower quality of the 1983 aerial image (especially in the southeast part of the glacier)”. The area determined in our study is closer to that reported by Arenillas-Parra et al. (2008), who reported an extent of 136 ha (1.36 km2) for the Aneto glacier in 1982 based on aerial photographs of a specific flight in the glaciated areas of the Pyrenees.

The values of ice thickness from the GPR reported in Campos et al. (2021) also show significant differences not consistent with our results. In 1994, the ERHIN programme estimated a maximum ice thickness of 52 m using 17 transects spaced 100 m apart (Arenillas-Parra et al., 2008; Jiménez-Vaquero, 2016). In 2008, those authors determined a maximum ice thickness of 30 m calculated from 31 GPR transects (Jiménez-Vaquero, 2016). Considering these data, Campos et al. (2021) reconstructed the subglacial topography of the Aneto glacier, and based on this topography they determined a maximum ice thickness of 55 m for 1983, 37 m for 2006 and 29 m for 2015. These values are in stark contrast to our estimates (maximum of 96.5 m in 1981, 52.5 m in 2011, 44.7 m in 2020, 43.5 m in 2021 and 41.8 m in 2022). Comparing the values of remaining thickness reported in 2008 (maximum ice thickness of 30 m; Jiménez-Vaquero, 2016) and the rate of ice thickness loss (-1.0 m yr-1) established by Vidaller et al. (2021) for the period 2011–2020, the expected maximum thickness in 2020 would be 18 m instead of the 44.7 m we observed in 2020. Additionally, large areas currently covered by the glacier would be ice free according to the previous ice thickness loss estimates. Considering that we used comparable values for wave propagation velocity of GPR signal to those used in the above-cited work, the differences between previous literature studies and the glacier thicknesses reported here are likely related to the more modern and accurate antennas used in our survey and the much denser net of transects conducted in the 2020 campaign. This methodology significantly reduces the uncertainties associated with the interpolation process, making the results obtained here more robust, and also permits a better understanding of the glacier's dynamic and its future behaviour.

The rate of surface and ice thickness losses calculated in this study and the reconstruction of ice thickness for the year 2022 indicate the critical situation of this glacier. There are no signs of slowdown in glacier surface and thickness loss rates; on the contrary, we have observed the high vulnerability of the Aneto glacier to the occurrence of extremely hot summers in recent years, as in 2022, when summer temperatures were 0.5 ∘C above the mean for the period 2007–2022, according to the Renclusa station (2140 m a.s.l.), and almost 2 ∘C in general in the Iberian Peninsula (AEMET). Thus, the continued loss of surface area and thickness could be due to an increase in temperature.

Taking into account the average current glacier thickness of 11.9 m, we can affirm that the Aneto glacier is indeed in its terminal stage, with evident fragmentation into smaller ice bodies, the absence of a significant accumulation zone and obvious signs of ice stagnation. In this context, glacier retreatment is exposing new areas of unconsolidated bedrock material (granite boulders and debris) and destabilising cirque walls in many areas. This process is also accompanied by a degradation of surrounding wall permafrost (Rico et al., 2021). Under this situation the occurrence of unusual warm periods, such as those observed in the 2021–2022 period, triggered hazardous rockfalls, as were also noticed in other mountain areas (Huggel et al., 2010; Kellerer-Pirklbauer et al., 2012). This behaviour could also anticipate the behaviour of other temperate mountain glaciers in their final deglaciation phases.

Another aspect that determines the evolution of the Aneto glacier is the darkening of the glacier surface. However, a detailed quantification of the darkening of the glacier surface and its effect on the energy and mass balance has not been carried out yet. Early spring (summer) snowmelt and glacier thickness loss result in a grey (dark) appearance of the glacier surface, which reduces the albedo effect and increases the absorption of thermal energy, leading to an acceleration of glacier surface and thickness losses (Shaw et al., 2021). The obvious similarities with the remaining glaciers of the range suggest that the Pyrenees may become an ice-free mountain range in the next few decades.

The rise in temperature in recent decades, combined with a slight decrease in precipitation, has resulted in less snow accumulation during the winter months. This results in longer exposure of the glacier during the ablation season, which increases the melting of the glacier from year to year. Compared to Pyrenean glaciers that have a minimal contribution to water resources in downstream areas (López-Moreno et al., 2020; Milner et al., 2017), changes in snowpack can lead to severe changes in the downstream water regime (García-Ruiz et al., 2011).

Also of note is the presence and development of new proglacial lakes, as in the case of Ibón Innominato. This small lake is in constant change due to the surface and thickness of the glacier, where the retreat of the glacier front has opened new outlets beneath the glacier, and consequently the water level of the lake decreases. Similarly, as the Aneto glacier shrinks, other lakes would be formed in the depression areas derived from the subglacial topography. The presence of proglacial lakes negatively affects the glacier's equilibrium by acting as an energy collector and accelerating the rate of thawing at the front of the glacier (Otto, 2019). In addition, the dark appearance of the glacier surface caused a decrease in albedo and therefore an increase in the surface and thickness losses of the glacier (Yue et al., 2020).

On the other hand, the maximum ice thickness (>44 m) is located under the Maldito pass, a protected area fed by avalanche channels and protected by the shadow of the Maldito peak. In these areas, longer persistence of the ice body is expected.

The fast surface loss of the Aneto glacier in the last few decades and the relatively low ice thickness observed together with the potential development of new lakes clearly show the consequences of climate change in mountain areas. Those changes happening nowadays in most mountain glaciers (Kääb et al., 2021; Barrand et al., 2017; DeBeer and Sharp, 2009) will have a major impact on mountain landscapes and ecosystems (Huss et al., 2017), showing the necessity of monitoring and understanding the recent fast evolution of these environments.

How long the glacier will maintain the ice movement and a surface greater than 2 ha to still be considered a glacier is a very uncertain issue to be estimated. The duration of the glacier depends on several factors, such as the temperature evolution in the next few years, the evolution of precipitation (mainly snowfall in winter), the ability of the glacier to transport the debris fallen from the headwalls (and avoid the darkening of the surface), possible events of dust deposition (which may be frequent in winter and spring) and many other factors. In addition, according to the study by Vidaller et al. (2021), it is possible that these very small glaciers, once they become smaller than 10 ha, will have a greater topoclimatic control, so their preservation could be prolonged if there are no more very hot summers, as in 2022. Otherwise, glacier extinction could be imminent if there are a few summers like 2022 in the next decade. However, more detailed studies are needed to answer such a simple question to reduce the uncertainty in observations and simulations and also to provide a deeper understanding of those processes that govern small and very small glaciers.