The study area is too large to be imaged by Pléiades in tri-stereo mode with a single scene; therefore the area was imaged in two strips, which overlapped by 3 km in winter and 1.5 km in summer in the along-track direction. The snow-on triplets were both acquired on 1 May 2017, while the snow-off triplets were acquired on 8 and 13 August 2017 (Fig. 1, Table 1, image ID in Table S1). The imaged area covers 280 km2 in total. Images were acquired in panchromatic and multispectral mode, with incidence angles along-track between -7 and 9∘. The base to height (B / H) ratio of successive pairs is around 0.1 (Table 1). The panchromatic images have a resolution of 0.5 m at nadir and are used to calculate the DEMs. For the snow-on acquisition, we requested to reduce the number of time domain integration (TDI) lines used to image the scene. This is recommended to curb image saturation over sun-exposed snow surfaces (Berthier et al., 2014). As a result, there are no saturated pixels in the images of this study. Pléiades multispectral images have a resolution of 2 m. We only use the multispectral image that was acquired closest to the nadir view angle to compute the multispectral orthoimage. Pléiades images were obtained at no cost through the DINAMIS program (only for French scientists; see https://dinamis.teledetection.fr/, last access: 1 May 2020). It is also open to European scientists working at public research institutions. Otherwise Pléiades images can be ordered from Airbus Defense and Space.

A snow-off DEM on 13 October 2015 and a snow depth map on 2 May 2017 from the ASO are used for comparison with the Pléiades products (Fig. 1, Table 1). The ASO program, operating since 2012, provides snow depth, snow water equivalent (SWE) and snow albedo maps over full mountain watersheds to support scientific campaigns and operational water management (Painter et al., 2016). The ASO laser-scanning system measures the distance between the target and aircraft and is combined with aircraft position and orientation measurements to generate a collection of reflection points: a “point cloud”. Ground points are aggregated to a 3 m grid to derive a gridded DEM (Painter et al., 2015). Snow depth maps are obtained from the difference of a snow-on and snow-off DEM in unforested areas. The values in the snow-free areas are used to bias-correct the snow-on elevations and are set to zero. From comparison with 80 in situ manual measurements, no bias is observed on the HS maps, and the root mean square error (RMSE) per pixel at a 3 m resolution is 0.08 m (Painter et al., 2016). For the evaluation of the Pléiades HS maps, we excluded 25 km2 near the catchment divide in the northeast part of the study area because we observed artifacts in the ASO HS map probably due to issues with the aircraft position and orientation data.

Figure 2 presents the workflow we developed to produce HS maps from Pléiades images using the ASP version 2.6.2 (Shean et al., 2016; Beyer et al., 2018) and the Orfeo ToolBox (Grizonnet et al., 2017). We detail below the calculation of the DEMs, the HS maps and the land cover classifications.

We evaluated the Pléiades HS maps for the area where both the Pléiades (snow mask) and ASO (HS greater than zero) HS maps had snow. The term HS residuals in the rest of the article refers to the difference between the Pléiades and the ASO HS (Pléiades HS minus ASO HS). We also evaluated the Pléiades dDEM over stable terrain, where we expect no elevation difference over time. The stable-terrain residuals are the Pléiades dDEM values as ASO products are set to zero over snow-off terrain. The distribution of the residuals was characterized with the bias (both the mean and the median), the root mean square error (RMSE) and the normalized median absolute deviation (NMAD) of the residuals. The NMAD is a measure of the dispersion suited for populations with outliers (Höhle and Höhle, 2009).

The accuracy of HS maps is often discussed at (or close to) the highest resolution that is allowed by the sensor (e.g., Nolan et al., 2015; Marti et al., 2016). In practice however, HS maps may be subject to spatial averaging to assimilate in a snowpack model, to estimate catchment-scale HS, or to compare with coarser satellite products and model output (Painter et al., 2016; Margulis et al., 2019; Shaw et al., 2019). The accuracy of the mean HS of a set of contiguous pixels is expected to be higher than that of a single pixel but depends on the spatial correlation of the errors (Rolstad et al., 2009; Anderson, 2019). Hence, we performed an empirical assessment of the evolution of the accuracy of Pléiades HS as a function of resolution by aggregating the HS residual map to resolutions ranging between 3 and 180 m (Berthier et al., 2016; Brun et al., 2017; Miles et al., 2018). An average resampling scheme was used, which calculates the average value of all valid contributing pixels. For each resolution, we compared the distribution of the HS residual or measured error to the standard error that is obtained from the error model of Rolstad et al. (2009). Using a spherical semivariogram model to measure the spatial correlation, Rolstad et al. (2009) estimate the random error of the spatially averaged residual, σA, as 1σA2=σ21-Llcor+15L3lcorifL<lcor2σA2=σ2lcor2L2ifL>lcor, where σ is the standard deviation of the elevation difference residual, lcor is the semivariogram range or length of autocorrelation, and L is the length of aggregation (half of the pixel spacing). A is the area of aggregation (pixel area) and is related to L (A=4×L2). This formula assumes no spatial trend in the HS residual map. We estimated lcor from a semivariogram analysis of the HS residuals at the highest resolution (3 m). The value of σ was taken as the NMAD of the HS residuals at the highest resolution (3 m). We also tested this equation using the stable-terrain residuals to set the value of lcor and σ. The measured error was taken as the NMAD of the aggregated HS residual. By comparing the measured error and the modeled error, we aim to verify whether (i) the error model from Eqs. (1) and (2) is valid and (ii) its parameters (σlcor) can be estimated from stable-terrain residuals only. It is important to evaluate whether the stable-terrain residuals can be used to parameterize the error model because that is the only available information in regions without HS reference data (Deschamps-Berger et al., 2019).

We first investigate the impact of different image geometries on the HS maps while keeping the stereo SGM-binary option fixed. The NMAD of the HS residuals with respect to ASO data (Table 2) is larger for maps from pairs of images with B–H around 0.12 (1.13 m for front–nadir, 1.07 m for nadir–back) than from pairs of images with B / H around 0.20 (0.68 m for front–back) or triplets of images. The NMAD of the snow depth residuals from the front–nadir–back triplets (0.69 m) is slightly better than from the nadir–front–back triplets (0.78 m) and very similar to the NMAD from the front–back pair. The NMAD over stable terrain is lower, but relative values between two geometries are similar (Table 2). For the different image geometries, the RMSE evolves similarly to the NMAD over snow-covered areas but very differently over stable terrain. The largest RMSE over stable terrain is 1.35 m for front–back, and the smallest is 1.06 m for nadir–front–back. The mean differences over snow-covered areas range from +0.01 m (front–nadir) to +0.16 m (front–back). The absolute means and medians over stable terrain are all less than 0.06 m. The relative results for the different geometries are similar to the SGM-ternary and local-search options except for the mean error (not shown here). In the following sections, the HS map from the front–nadir–back geometry is used as it yields the lowest bias, RMSE and NMAD of all the geometries, although it is similar to the front–back geometry.

We compare the stereo options on the HS maps from the front–nadir–back geometry (Table 3 and Fig. 3). The SGM sets of options provide DEMs without data gaps. The local-search option produces snow-on DEMs with gaps, which results in ∼ 2 km2 missing in the HS maps compared to the SGM options (Table 3). Visual examination of the winter DEMs shows large differences in snow fields and forest. Linear artifacts are observed over snow in the DEM produced with the SGM-ternary option (Fig. S1). The same regions are noisy in SGM-binary. Patches of typically 20 m × 20 m with abnormally large HS (> 10 m) compared to ASO (∼ 3 m) are also observed with the local-search options around isolated trees. These artifacts are not visible with the SGM-binary or ternary options (Fig. S1).

The mean differences from the ASO snow depth data range from +0.08 m (SGM-binary) to +0.49 m (local-search option). The mean differences are larger for SGM-ternary (+0.24 m) than SGM-binary. The NMAD of the residuals is smaller for SGM-binary (0.68 m) than the local-search (0.80 m) and SGM-ternary options (0.85 m). Over stable terrain, the absolute mean and median of the elevation differences are less than 0.03 m except for the mean of the local-search option, which is -0.32 m. The mean of the elevation differences for local search decreases to -0.03 m when the elevation differences are excluded if they exceed 3 times the NMAD value. This is expected as the same filtering is used during the coregistration process to remove outliers. In the following, the SGM-binary was selected since it gives the lowest bias and NMAD with respect to ASO data and the lowest NMAD over stable areas (Table 3).

The Pléiades HS map calculated with the selected image geometry and ASP configuration (front–nadir–back images, SGM-binary) compares well with the ASO HS map (Fig. 4). Typical mountain snowpack features (e.g., avalanche deposits and snow drift accumulation) can be identified on the Pléiades HS map (Figs. 4d, e, 5). Pléiades HS data are available over 215 km2 of open terrain but not for the 23 km2 of forest. No HS was higher than 30 m, but 0.25 km2 of HS were excluded because HS was less than -1 m. This occurred in areas covered with low-density deciduous vegetation, which was classified as snow. The intersection area of Pléiades and ASO snow-covered areas is 138 km2 after erosion of the Pléiades snow mask. The Pléiades mean (median) HS is 4.05 m (4.13 m) against 3.96 m (4.02 m) for ASO over the common snow-covered area. ASO and Pléiades HS exhibit a similar relationship between HS and elevation (Fig. 6) except between 1900 and 2100 m and between 3500 and 3700 m, where the mean residual over snow-covered areas is greater than 0.25 m (Fig. 7). This corresponds, however, to elevation ranges that cover less than 0.05 km2 each.

The NMAD of the Pléiades dDEMs over the 4.07 km2 of stable terrain is 0.40 m against 0.69 m for the HS residual. The distribution of residuals on stable terrain is similar for most aspect classes with the exception of the north-facing slopes (0.26 km2 , aspect classes 315–360∘ and 0–45∘; Fig. 7). Based on a visual analysis of the residual map, we attribute these errors to shaded slopes of steep summits. The distribution of HS shows a similar spread for all aspects but a larger positive bias (∼ 0.20 m) for south-facing slopes (90–270∘, Fig. 7). The distribution of HS residuals against the terrain slope is similar between 0 and 50∘ but has a greater spread in steeper terrain, which covers 2.13 km2 . The same trend is observed over stable terrain but only above 70∘.

The semivariogram of the residual increases from 0.2 to 0.8 linearly for lag distances between 3 and 20 m (Fig. 9a). Low-amplitude undulation for lag distances between 2000 and 8000 m (Fig. 9b) is related to a low-frequency undulation in the HS residual map, which has an amplitude of approximately 0.30 m and a wavelength of about 4 km (Fig. 8). The crests of the undulation are oriented in the east–west direction (Fig. 8). Such an undulation pattern was observed in other Pléiades products, ASTER images (Girod et al., 2017) and WorldView DEMs (Fig. 10 in Shean et al., 2016; Fig. 6 in Bessette-Kirton et al., 2018). It is attributed to nonmodeled satellite attitude oscillations along-track (jitter). A similar semivariogram shape is obtained over stable terrain. From this semivariogram analysis we estimate that the correlation length of the residuals (see Sect. 4.2) is about 20 m for both snow and stable areas.

The measured error of the HS map decreases with increasing resampling resolution (Fig. 10). The NMAD of the HS residuals is reduced by a factor of almost 2 by resampling from the original resolution of 3 m (NMAD = 0.69 m) to 36 m (NMAD = 0.38 m). As explained in Sect. 4.2, we computed two error models using either the HS residuals (lcor=20 m, σ=0.69 m) or the stable-terrain residuals (lcor=20 m, σ=0.40 m) to parameterize Eqs. (1) and (2). We find that the NMAD of the HS residual well matches the error modeled for averaging areas smaller than 103 m2 when lcor and σ are calculated with the HS residuals (Fig. 10). However it does not match with the modeled error for averaging areas larger than 103 m2 (Fig. 10). This is due to the lower decrease in the residual dispersion with spatial resolution. The measured NMAD decreased by 0.07 m between 36 m resolution and 180 m resolution, while the modeled error decreased by 0.22 m between the same resolutions. We attribute this mismatch to the undulation pattern identified in Sect. 5.3 (see Sect. 6.5 in Discussion).

By comparing the Pléiades HS with the ASO data, we find an NMAD of 0.69 m in the best case (i.e., best acquisition geometry and ASP options), which is close to or higher than most previous evaluations (Table 4). Only Marti et al. (2016) measured a larger NMAD (0.78 m), with a reference HS map of 3.15 km2 that was obtained by UAV photogrammetry. The spread in accuracy between studies in Table 4 could be due to differences in (i) the satellite data (i.e., acquisition geometry, image resolution), (ii) the characteristics of the study site and (iii) the representativeness of the validation data. The comparison with snow probe measurements showed NMAD about a third lower than this study, at 0.45 m (n=442; Marti et al., 2016) and 0.47 m (n=36; Eberhard et al., 2020), but covered limited portions of the studied sites. The B / H for the images of Marti et al. (2016) range between 0.21 and 0.25 for all consecutive stereo pairs, while our B / H ranges between 0.08 and 0.12. This is consistent with photogrammetry theory, which states that the accuracy of the DEM increases with the B / H up to a certain limit (Delvit and Michel, 2016). We find a similar NMAD to Eberhard et al. (2020), who calculated an HS map from a Pléiades snow-on DEM and an airplane structure from motion (SfM) snow-off DEM and compared it to HS from the airplane SfM over 75 km2 (NMAD = 0.65 m). Finally, McGrath et al. (2019) found an NMAD of 0.24 m for HS from WorldView-3 stereo DEMs using 2107 point observations from ground-penetrating radar surveys over a flat area of roughly 50 km2 . This lower NMAD value might result from the higher resolution of the WorldView-3 images (0.3 m) together with the flatter terrain used for evaluation. As the ASO provides a much larger reference dataset over a complex terrain, we argue that our study provides a more robust evaluation of the HS accuracy that can be expected from Pléiades in high mountain regions. While the ASO data itself may add some error, the published accuracy of the ASO HS data is significantly better than Pléiades. In all these studies, the absolute mean biases range between 0.01 m (McGrath et al., 2019) and 0.35 m (Eberhard et al., 2020).

We find that the HS map accuracy is sensitive to the B / H ratio of the input images and to the configuration details of the photogrammetric processing. We do not find a large added value of the tri-stereo images for the map accuracy compared to an optimal bi-stereo configuration. Tri-stereo might provide greater benefits in case of image occlusion in steep slopes, which is more prone to occur with higher B / H.

The NMAD of the Pléiades HS is improved by 36 % when using images with a B / H of 0.22 instead of 0.11 (Table 3). Marti et al. (2016) used pairs of front–nadir and nadir–back images (B / H = 0.2) as they observed that the front–back pair (B / H = 0.4) led to too many no-data pixels. From these two studies and for similar terrain, a triplet of images with a B / H for consecutive images around 0.2 seems to be a good compromise. It should ensure high coverage and good DEM precision. Further work is needed to confirm this statement by testing varying B / H values.

Using tri-stereo instead of bi-stereo images did not significantly improve the Pléiades HS map accuracy. It seems like the processing of a triplet of stereo images (front, nadir, back) with the ASP stereo function is equivalent to the processing of the best stereo pair of the triplet, the front–back pair in our case. There were no data gaps due to view obstruction by steep relief in this study area. Should this be the case, the tri-stereo may offer better coverage. Several studies have evaluated the benefits of tri-stereo imagery against bi-stereo (Berthier et al., 2014; Zhou et al., 2015; Bagnardi et al., 2016; Marti et al., 2016). However, these studies used different photogrammetric software that does not handle the combination of three images in the same way. For example, either multiple disparity maps, point clouds or DEMs can be calculated and merged to produce a final single DEM. The use of tri-stereo results in increasing the density of the point cloud (Zhou et al., 2015; Bagnardi et al., 2016) and decreasing the area with missing data in the final DEM (Berthier et al., 2014; Zhou et al., 2015). The accuracy of elevation products from tri-stereo was slightly modified in Berthier et al. (2014) and Marti et al. (2016) compared to bi-stereo, with an increase or decrease in the NMAD by a few percent.

To our knowledge, surface elevation changes were never computed from a large number of very-high-resolution satellite stereo images (> 10), but a study suggests that the combination of multiview images can improve the DEM quality. The fusion of 16 WorldView-3 images improved the NMAD of the residual by 20 % compared to a set of 6 images over an industrial zone (Rupnik et al., 2018). Therefore, the most important use of tri-stereo may not be to improve the accuracy of HS maps but rather to obtain complete coverage of complex terrain and have a less distorted nadir orthoimage for land surface classification. We did not evaluate the extent to which the front and back images would provide a different land surface classification from the one obtained with the nadir image.

The choice of the photogrammetric options has an impact on the elevation difference accuracy over stable terrain and snow-covered areas. The NMAD over snow-covered areas is improved by 0.16 m by modifying the cost function (binary census transform instead of ternary census transform). However, such a decrease in the dispersion will hardly impact the HS averaged over a region of interest since the random error decreases rapidly with increases in averaging area (see Sect. 6.5). More important is the larger mean bias over snow-covered areas introduced with the SGM-ternary option (0.24 m) and local-search option (0.49 m) compared to the SGM-binary option (0.08 m). This bias is particularly important for south-facing slopes. It seems to result from difficulties in image matching in bright areas for the three options and from the impact of isolated trees for local search. The impact of the trees is likely due to the larger kernel size (25 pixels) used in the local-search option. The exact origin of the bias on south-facing slopes remains unknown.

We found a mean difference of +0.08 m between Pléiades (SGM-binary, front–nadir–back) and ASO HS despite the correction of the vertical offset between the snow-on and snow-off DEM using stable terrain after coregistration. This bias is low given the differences in the characteristics of the ASO and the Pléiades products. This can be due to many factors, including the effect of vegetation. First, the ASO snow-off DEM is a digital terrain model, while the Pléiades snow-off DEM is a digital surface model. Tall vegetation (i.e., trees) is identified during the classification of the MS images and does not impact the HS evaluation. But short vegetation completely covered with snow in winter is not identified in the classification. For ASO products, filtering based on the multiple lidar returns produced by vegetation should provide the ground elevation, but short vegetation often does not produce multiple returns (Painter et al., 2015). Furthermore, there is a large known error in vegetation height measured with Pléiades DEMs (Piermattei et al., 2018). Thus, it is still unclear which surface is sensed by each method between the top of the vegetation and the underlying ground.

We found that the random error is larger on snow-covered terrain (NMAD = 0.69 m) than on stable terrain (NMAD = 0.40 m). This is true for all slopes and most aspect classes (Fig. 7). Although mountainous snow surfaces tend to have smoother topography, thereby increasing the accuracy of the photogrammetric processing, bright snow surfaces also tend to have less texture than snow-free surfaces, which decreases the accuracy of the photogrammetric processing. The lower accuracy of snow areas is not due to saturation since no pixel was saturated in the panchromatic images. In addition, the residuals over stable terrain were computed from Pléiades data only, while residuals over snow-covered areas were computed from Pléiades and ASO data. Finally, the coregistration of the snow-on DEM was optimized using the stable terrain (Sect. 4.1.4); therefore a lower NMAD on stable terrain may be due to the coregistration step and not the photogrammetric processing in itself. Based on the above, we cannot conclude whether the larger dispersion over snow-covered areas results from the properties of the surface.

We further compared Pléiades snow-off DEM with the ASO snow-off DEM and Pléiades snow-on DEM with the ASO snow-on DEM. The latter was calculated by adding the ASO snow-off DEM and the ASO HS. Both Pléiades DEMs are coregistered as described in Sect. 4.1.4. We find a mean bias over snow-covered terrain of +0.13 m for snow-off conditions and +0.21 m for snow-on conditions (Table S3). These biases are of the same order of magnitude and suggest that a bias in the Pléiades snow-on DEM is partially compensated by the difference of the surface observed in the snow-off DEM (see above). In addition, the ASO snow-off DEM was acquired in October 2015 and the Pléiades snow-off DEM in August 2017. Growth or decay of the vegetation can occur over almost 2 years, leading to elevation differences between the snow-off DEMs. The NMAD is larger for snow-off DEMs (0.80 m) and snow-on DEMs (0.93 m) compared to the HS residual (0.69 m). This shows that some errors are consistently present in the snow-off and snow-on DEMs of each type (airplane lidar or satellite photogrammetry). Pléiades DEMs indeed overestimate the surface elevation as the terrain slope increases (Fig. S2). This suggests that combining satellite photogrammetry and airplane lidar DEMs may lead to larger errors than comparing DEMs from the same platform.

The error predicted with Eqs. (1) and (2) does not agree with the NMAD of measured HS error for averaging areas larger than 103 m2 (Fig. 10). This is likely because Eqs. (1) assumes a randomly distributed error beyond the short-distance correlation length (here 20 m), while the undulation pattern identified in Fig. 8 introduces an additional spatial correlation at larger scales in the HS residual map. To verify this explanation, we applied an empirical correction to remove the undulation pattern from the residual map. We averaged the HS residuals by pixel rows in the across-track direction and used a Fourier transform to identify the undulation frequencies (adapted from Girod et al., 2017). Then, we modeled this error by selecting the frequencies lower than 4×10-4 m-1 (i.e., wavelength longer than 2.5 km) and removed it from the HS map. As expected, this correction makes the semivariogram of the HS residual flatter for lag distances between 2000 and 8000 m (Fig. 9b). As a result, there is a better agreement between the HS residuals' NMAD and the modeled error, with σ and lcor estimated from the HS residuals (lcor=20 m, σ=0.69 m; Fig. 10). The improvement is more marked at lower resampling resolution. For instance, the HS NMAD is reduced after correction by 50 % at a resolution of 180 m. The improvement is under 10 % at 20 m resolution as expected since the correction only dampers a low-frequency signal. When the stable-terrain residuals are used to compute Eqs. (1) and (2) (lcor=20 m, σ=0.40 m), the modeled error is lower than the measured error. This is expected since the NMAD of the stable-terrain residuals is lower than the NMAD of the HS residual. However, the discrepancy between both models decreased at coarser resolution.

This analysis shows that the model proposed by Rolstad et al. (2009) provides a good first-order estimation of the random error after spatial aggregation under the assumption that there is no spatial drift in the error at scales beyond the correlation length. In most cases, the statistics of the HS residuals are not available and might only be measured on stable terrain. Interestingly, in this study the correlation length of the error is similar over stable terrain and snow terrain. However, the dispersion (NMAD, standard deviation) is 2 times larger over snow-covered terrain than stable terrain, which leads to a proportional underestimation of the error. Finally, although the bias or systematic error is corrected on stable terrain, there remains a bias on HS of the order of ∼ 0.20 m (Table 4) that should be taken into account in the error calculation. According to the literature, this bias can be estimated by comparing the mean and median of elevation differences over stable terrain (Gardelle et al., 2013) or by calculating the residual of the coregistration vector when more than two elevation datasets are available (Nuth and Kääb, 2011).

ALS provides HS maps with a better accuracy (RMSE < 0.10 m) than Pléiades and potentially a finer horizontal resolution too (Painter et al., 2016). One significant advantage of ALS is that it can measure HS under the tree canopy and in shaded areas. It is also able to acquire data in overcast conditions provided that the clouds are above the aircraft. However, from this study and that of Marti et al. (2016), it appears that the accuracy of Pléiades HS maps is sufficient to provide valuable information in regions where there is no ALS monitoring capability (the vast majority of mountain regions with snow cover). A limitation of current very-high-resolution sensors such as Pléiades is their narrow swath (20 km for Pléiades), which impedes the acquisition of large areas with a frequent revisit. In particular, there are areas of high tasking competition at lower latitudes, where it can be challenging to obtain a stereo pair at the right time of the snow season. More frequent acquisitions should, however, become easier as new stereo satellite fleets are to be launched in the coming years (Pléiades Neo, WorldView Legion). The acquisition of visible images will always be limited by the presence of clouds, making some regions hard to study at least during some seasons.

HS maps from UAV SfM typically exhibit a centimeter-scale bias (0.05 to 0.11 m) and an RMSE between 0.05 and 0.30 m based on comparison with snow probe and Global Navigation Satellite System (GNSS) measurements. This is more accurate than what is currently achieved with satellite photogrammetry. However, UAV campaigns are currently limited to areas of a few square kilometers due to battery limitation and often rely on numerous ground control points. This greatly limits the possibility to cover large and remote areas. Airplane SfM exhibits accuracy close to UAV SfM, with an NMAD typically of 0.30 m (Bühler et al., 2015), and presents the same potential and logistic limitations as airplane laser-scanning campaigns. The reader is referred to the study of Eberhard et al. (2020) for a detailed discussion on the different approaches to map snow depth with photogrammetry.

Several snow applications could benefit from HS maps from satellite photogrammetry. First, this study could be reproduced in any place of the globe provided that (i) there is a window to acquire snow-off images and (ii) there is a way to coregister the series of DEMs, for example with stable terrain. This method is particularly suited for snow volume evaluation at a basin scale in alpine areas (this study site; Marti et al., 2016; McGrath et al., 2019; Shaw et al., 2019). Observing shallow snowpack (HS roughly below 0.5 m, e.g., polar environments) might not be as straightforward as the typical spatial variability lies within our range of uncertainty (roughly 0.5 m). However, even landscapes with shallow snowpack often feature local accumulation of snow, which would be measurable with satellite photogrammetry. Therefore it is hard to qualify this method as unfit to any region, but future studies are required to confirm its usefulness in these challenging contexts. Study of shallow snowpack would clearly benefit from higher-accuracy DEMs through correction of the satellite jitter or increases in image resolution.

A lack of well-distributed stable terrain in snow-on and snow-off DEMs can complicate the coregistration process in some regions. The horizontal component of the coregistration vector can be measured without differencing stable terrain and snow-covered terrain (Marti et al., 2016), but the vertical component requires some stable terrain or an elevation reference. Ground control points (GCPs) could be used but would limit the applicability of the method in remote mountains. Besides, it remains to be tested how many GCPs would be required and how precisely their position should be measured.

There are already a number of efficient free and open-access photogrammetric software tools that are under continuous development. These tools enable a high level of automation and are compatible with high-performance computing environments (Howat et al., 2019). In our workflow, the last step to automate is the collection of training samples for image classification. This could be done by using an unsupervised classification algorithm or by using an external land cover classification. Preliminary work with a time series of Pléiades images in the Pyrénées (not shown here) suggests that it is not possible to simply use the classification model from a previous year to generate the classification of the current year. A possibility may be to use a Sentinel-2 snow cover map to extract training samples in the Pléiades multispectral images since Sentinel-2 images have a shortwave infrared band, which enables a robust and unsupervised detection of snow cover (Gascoin et al., 2019). Differentiating terrain covered with vegetation from stable terrain would remain challenging.

We find that the selection of the image configuration and the processing options can lead to changes in the NMAD up to ∼ 0.3 m. Figure 10 suggests that this variation is likely to become insignificant if the HS map is aggregated at a larger spatial scale (grid spacing larger than 100 m × 100 m). Such optimization is therefore more important for the study of small-scale features (wind drift, avalanches, typically about a few tens of meters) or to decrease bias on specific terrain (south slopes, fields with isolated trees). The optimization of the photogrammetric processing can also be important when little stable terrain is available for the coregistration step.