Antarctic digital elevation models (DEMs) are essential
for fieldwork, ice motion tracking and the numerical modelling of the ice sheet.
In the past 30 years, several Antarctic DEMs derived from satellite data
have been published. However, these DEMs either have coarse spatial
resolution or aggregate observations spanning several years, which
limit their further scientific applications. In this study, the
new generation satellite laser altimeter Ice, Cloud, and Land Elevation
Satellite-2 (ICESat-2) is used to generate a new Antarctic DEM for both the
ice sheet and ice shelves. Approximately 4.69
Knowledge of the detailed surface topography in Antarctica is essential for fieldwork, ice motion tracking (Bamber et al., 2000) and the numerical modelling of the ice sheet (Cornford et al., 2015). Digital elevation models (DEMs) of Antarctica, for example, can be used for presenting the topography of ice sheets and ice shelves and thus provide a crucial reference for ice dynamics and glacier velocities (Wesche et al., 2007; Slater et al., 2018), which is necessary for Antarctic mass balance monitoring and potential sea level rise contribution estimation (Ritz et al., 2015; Mengel et al., 2018).
Due to the remoteness of Antarctica, most of the previously published Antarctic DEMs were derived from satellite or airborne data, e.g. elevation measurements from radar altimeters (Fricker et al., 2000; Helm et al., 2014; Slater et al., 2018), laser altimeters (DiMarzio et al., 2007), a combination of radar and laser altimeters (Bamber et al., 2009), stereophotogrammetry (Korona et al., 2009; Cook et al., 2012; Howat et al., 2019) and interferometry (Wessel et al., 2021). The currently available continent-scale Antarctic DEMs include one DEM derived from ICESat (hereafter called the ICESat DEM, DiMarzio et al., 2007), one based on the combination of ICESat and ERS-1 elevation measurements (hereafter called the ICESat/ERS-1 DEM, Bamber et al., 2009), two DEMs derived from CryoSat-2 (hereafter called the Helm CryoSat-2 DEM, Helm et al., 2014 and Slater CryoSat-2 DEM, Slater et al., 2018), one DEM derived from stereophotogrammetry using GeoEye-1 and WorldView-1/2/3 imageries (hereafter called the Reference Elevation Model of Antarctica, REMA, DEM, Howat et al., 2019), and one DEM derived from Interferometric Synthetic Aperture Radar (InSAR) using TerraSAR-X and TanDEM-X data (hereafter called the TanDEM-X PolarDEM, Wessel et al., 2021).
All these DEMs provide reasonable elevation estimates for Antarctica; however, some flaws still cannot be totally avoided. The coverage of ICESat is limited in ice sheet margins due to its coarse across-track resolution, hence for ICESat DEM most of the elevations in ice sheet margins were interpolated based on the neighbouring data. Although the ICESat/ERS-1 DEM improves the data coverage by combining the measurements from ICESat and ERS-1 elevations, this DEM aggregates observations spanning several years due to the different time spans (1994–1995 for ERS-1 and 2003–2008 for ICESat) of these two satellite altimeter datasets. This issue also exists with the REMA DEM and TanDEM-X PolarDEM, where multiyear satellite imageries were used. Different from the abovementioned DEMs, the Slater CryoSat-2 DEM was derived based on a model fitting method by using 7-year CryoSat-2 data (from July 2010 to July 2016). This method can quantify the measured elevation fluctuations due to interannual variations, and can provide a DEM for each month during the time span of applied data; however, although the radar penetration depth of the CryoSat-2 Ku-band into snowpack can be corrected either empirically or theoretically using a waveform fitting approach (Davis, 1996, 1997), the spatial and temporal variations of radar penetration depth are still difficult to account. As multitemporal and large-scale satellite radar altimeter data are usually used, the accuracy of estimated elevations is reduced. A similar problem also exists with the Helm CryoSat-2 DEM and TanDEM-X PolarDEM (the penetration depth of the X-band into snow may be several metres, Fischer et al., 2020; Dehecq et al., 2016).
The new generation satellite laser altimeter Ice, Cloud, And Land Elevation
Satellite-2 (ICESat-2) of the National Aeronautics and Space Administration
(NASA), which was launched on 15 September 2018, provides near-global (up to
88
Here, we use a 1-year time series (from November 2018 to November 2019) of ICESat-2 elevation measurements to generate a new Antarctic DEM that covers both the ice sheet and ice shelves (hereafter called the ICESat-2 DEM). The applied data, DEM generation method and quality control criteria are presented in Sect. 2. Furthermore, we present the map of the ICESat-2 DEM and construct an accuracy evaluation by comparing it to the elevation measurements from the NASA Operation IceBridge (OIB) airborne mission and kinematic GPS and the Russian Global Navigation Satellite System (GLONASS) (GNSS) transects in Sect. 3. The performances of the ICESat-2 DEM and six currently available Antarctic DEMs are compared in Sect. 4, Sect. 5 provides the data availability and Sect. 6 summarizes the conclusions of this study.
The ICESat-2 ATL06 land ice elevation product (version 3, Smith et al., 2019) from November 2018 to November 2019 is used. This product provides land ice elevation measurements at a spatial resolution of 20 m after correcting instrument-specific biases (i.e. corrections for transmit-pulse shape and first-photon bias, Neumann et al., 2019); here, only ATL06 data with good quality (according to the surface signal confidence metric from ATL06 data, i.e. those for which atl06_quality_summary equals zero) are used to generate the DEM. For the data collected over Antarctic ice shelves, corrections for ocean tide and inverse barometer effects are also applied (Egbert et al., 1994; Egbert and Erofeeva, 2002; Padman et al., 2002). Elevation measurements from all six beams are used to produce the densest surface height coverage. Although the signal energies of strong and weak beams are different, all six beams provide centimetre-scale elevation measurements, and the biases of two beams in one pair are less than 2 cm for flat regions (Brunt et al., 2019) and 5 cm for steep surfaces (Shen et al., 2021b). Thus, the effect of elevations estimated from weak beams is negligible.
Elevation measurements from the OIB airborne mission in Antarctica are used
here to evaluate the accuracy of the ICESat-2 DEM on a continental scale,
including in the stable ice sheet interior and active marginal ice shelves.
Surface heights from OIB airborne missions are measured by the Airborne
Topographic Mapper (ATM), a conically scanning laser altimeter (at 532 nm)
with a swath width of 140 m and footprint size of 1–3 m. The elevation
measurement accuracy of the ATM is approximately 10 cm or better (Kurtz et al.,
2013). Here, the IceBridge ATM L2 Icessn elevation, slope and roughness
(V002) product (Studinger, 2014) is used, and a data filter (Young et
al., 2008; Kwok et al., 2012; Studinger, 2014) is applied to remove
abnormal values due to geolocation errors or cloud cover. To reduce the
effect of interannual changes of surface elevations on DEM evaluation, the
time difference between applied OIB airborne data and ICESat-2 DEM should be
less than 1 year. Thus, OIB airborne data in October and November 2018 and
October and November 2019 in Antarctica (Fig. 1) are chosen to evaluate the
accuracy of the ICESat-2 DEM. In order to provide a comprehensive and more
robust evaluation of the ICESat-2 DEM, OIB data in areas of low elevation
change (i.e. ice sheet interior) from 2009 to 2017 are also used
additionally (Fig. 1). The CryoSay-2 low rate mode (LRM) mask in Antarctica
(which was designed for flat ice sheet interior measurements) is used to
extract the regions of low elevation change. The CryoSat Geographical Mode Mask
(V4.0, updated in 19–26 August 2019) at
In addition, elevation records from kinematic GNSS observations
(Schröder et al., 2017) in areas of low elevation change are also used
for an additional DEM elevation comparison (Fig. 1). These GNSS profiles
were measured in the region from the Vostok Station (106.8
Although OIB and GNSS data in the low elevation change areas and OIB data
with a small time difference (
Maps of the OIB airborne data in October and November 2018 and October and November 2019 (red), from 2009 to 2019 in ice sheet interior (blue). Map of the GNSS transects from 2001 to 2015 in Antarctica (yellow). The dashed lines show the boundary of the region where we assume there is low elevation change, which is the mode mask boundary of CryoSat-2 LRM data in Antarctica.
Detailed introduction to six previously published Antarctic DEMs, including the source data, time span of the source data, spatial posting or resolution.
The six previously published Antarctic DEM products are compared to the ICESat-2 DEM, i.e. ICESat DEM (DiMarzio et al., 2007), ICESat/ERS-1 DEM (Bamber et al., 2009), Helm CryoSat-2 DEM (Helm et al., 2014), Slater CryoSat-2 DEM (Slater et al., 2018), REMA DEM (Howat et al., 2019) and TanDEM-X PolarDEM (Wessel et al., 2021), as shown in Sect. 4. Detailed information concerning these DEMs is provided in Table 1, and all DEMs have been referenced to the WGS84 ellipsoid.
To separate the various contributions (i.e. local surface terrain and
elevation change), following Slater et al. (2018) a model fitting method is
applied here. The elevation is estimated using a quadratic function based on
the local surface terrain and a time term (Eq. 1). This function is fitted
in each grid (at resolutions of 500 m and 1 km, see following
subsection) by using an iterative least-squares fit to all the included
elevation measurements. By considering the surface elevation fluctuations
and sub-annual changes, this method tends to obtain more accurate elevation
estimates (Flament and Rémy, 2012; McMillan et al., 2014).
This method suits the ICESat-2 orbit cycle, which samples dense ground tracks
comparing to previous satellite altimeters, more measurement points are
included in the grid cell and the estimated elevations are more robust. It
is possible for a quadratic form to model the topography at the resolutions
of 500 m and 1 km and smaller elevation residuals can be found than using a
simple linear fit (Flament and Remy, 2012). In addition, the model fitting
method can provide the estimation of elevation change rate (
To reduce the effect of any poor fit the quality control criteria listed in Table 2 are performed, which includes the number of ICESat-2 measurement points used, the time span of the data used, the root-mean-square deviation (RMSD) of the residuals of fitted elevations, the elevation rate of change and the uncertainty. These criteria are constructed for all grid cells, and thus, there are some elevation gaps in the initial DEM. The remaining gaps are filled by using ordinary kriging interpolation (semi-variogram model: spherical, nugget: 0, sill: 1 652 285.953, radius: 10 km), which is widely used for generating previous DEMs (Helm et al., 2014; Slater et al., 2018). During the interpolation process, a search radius of 10 km is applied to obtain neighbouring measurement points. Similar estimation models have also been used in previous studies (Moholdt et al., 2010; Flament and Remy, 2012; McMillan et al., 2014; Konrad et al., 2017; Slater et al., 2018), and the evaluation in Sect. 3.2 also demonstrates its validity.
Quality control criteria applied to remove the unrealistic elevations due to the poor fitting performances in each grid cell.
The performance of this surface fit method is also affected by the spatial
distribution and number of ICESat-2 measurement points. After quality
control, 4.69
Map of the numbers of valid ICESat-2 measurement points in each 500 m grid cell. The numbers of ICESat-2 measurement points in 1 km grid cells are resampled to the resolution of 500 m.
Uncertainties in DEM are calculated based on the approach described in Helm et al. (2014). The OIB elevation data are used as the reference and the elevation
differences due to the time difference between OIB data and DEM are
corrected based on the elevation change rates from Smith et al. (2020). The
DEM uncertainty is then calculated from surface slope, roughness, number of
data points used(
Fitting coefficients and weights used for the DEM uncertainty estimation.
Maps of the observed grid cells of DEMs at the spatial
resolution of 250 m
The selection criterion for DEM resolution is to present the detailed pattern of elevations and ensure enough spatial coverage of observed elevations (a smaller resolution tends to cause more observed elevation gaps). Although a much finer scale (e.g. 250 m) can reveal a more detailed elevation pattern, this contributes to more gaps among observed elevations. The overall spatial coverage of observed elevations when applying 250 m, 500 m and 1 km resolutions (which are usually applied in the Antarctic DEM) are 26 %, 46 % and 72 %, respectively. High latitude areas always have higher observed elevation coverages; in lower latitudes there are still some 250 m grid cells with estimated elevations (Fig. 3); however, 250 m DEM only has 26 % coverage. The detailed variations in the spatial coverages of observed grid cells at different latitudes at variable spatial resolutions (250 m, 500 m and 1 km, which are usually applied in the Antarctic DEM) are shown in Fig. 4a. A reliable grid size is 500 m, which provides denser spatial coverage of the observed elevations but a single resolution cannot obtain ideal spatial coverage, especially in low latitude areas. To increase the coverages of observed elevations as much as possible, referring to Slater et al. (2018), two spatial resolutions are used to estimate the surface elevations from ICESat-2, i.e. elevations are estimated at resolutions of 500 m and 1 km. The observation gaps in the 500 m DEM are filled by the resampled 1 km DEMs (resampled to the 500 m DEM). The addition of DEMs at 1 km greatly increases the observation coverage, the overall spatial coverage is approximately 74 %, and the remaining gaps are filled using ordinary kriging interpolation. Although two resolutions are applied, 1 km and interpolated elevations are both resampled to the posting of 500 m to provide a consistent DEM dataset; hence, the final ICESat-2 DEM is posted at a resolution of 500 m.
The application of two resolutions may include additional effects, i.e.
different grid cell resolutions tend to present different elevation
estimates. Here, we compare the elevation difference at the overlapped areas
in Antarctica at different spatial resolutions (Fig. 4b). The elevation
values become lower when a larger spatial resolution is applied, which acts
as a running mean. Although applying different spatial resolutions affects
the elevation values, an averaged elevation difference of 0.04
Finally, in order to remove additional elevation outliers in the generated
DEM, a 3 standard deviation filter (3
The ICESat-2 DEM and previously published DEMs are resampled to the OIB and GNSS
data locations and the differences for evaluation are calculated to reduce the
effect of resolution differences between various DEMs. Four indexes are used
to evaluate the DEM performance, including median deviation (MeD), median
absolute deviation (MeAD), standard deviation (SD) and RMSD. The
corresponding calculation equations are listed as follows:
The effective time stamp of the ICESat-2 DEM is May 2019, which is halfway between November 2018 and November 2019. The ICESat-2 DEM provides a complete surface elevation reference for Antarctica, which illustrates higher elevations in the ice sheet interior and lower values in marginal ice shelves (Fig. 5). Negative elevations can be found in the ice shelves, especially in the Ross Ice Shelf. The local slope shows a pattern similar to the DEM, and undulated slopes are found in areas with rugged terrain, such as the Antarctic Peninsula and Transantarctic Mountains (Fig. 6). Both elevation and slope uncertainties show topography-dependent patterns, and larger values tend to be found at rugged areas, which may be related to the local surface conditions (i.e. slope and roughness). Larger elevation uncertainties can be found for interpolated grid cells than for observed ones, which is due to the difference in methods when deriving the surface elevations.
According to the shaded relief map of Antarctica derived from the ICESat-2 DEM (Fig. 7), obvious topographical patterns and flat terrain can be found in the mountain environments and ice sheet interior, respectively. On the Antarctic Peninsula, the ice shelf limit is visually identified from the shaded relief map (Fig. 7b). Other large-scale terrain features, e.g. subglacial lakes and floating ice shelves, can also be visually detected (Fig. 7c and d).
Two spatial resolutions are used in the ICESat-2 DEM, and the distributions of three kinds of grid cells (interpolated and observed at resolutions of 500 m and 1 km) show obvious latitude-dependent patterns. Regardless of whether at the basin scale or regional scale, more elevations at higher resolutions tend to be located in high-altitude areas, while elevations at lower or interpolated resolutions are mostly located in low-altitude regions (Fig. 8).
Comparisons between the ICESat-2 DEM and OIB airborne elevation measurements (including data in areas of low elevation change from 2009 to 2017 and data in the Antarctica from 2018 to 2019) in observed and interpolated areas for individual regions (i.e. the ice sheet and ice shelves). MeD: median deviation, MeAD: median absolute deviation, SD: standard deviation, RMSD: root-mean-square deviation.
In total, approximately 5.2
We also evaluate the elevation performance for observed and interpolated grid cells (Table 4). Generally, the bias of observed elevations is smaller than that of interpolated elevations in both ice sheets and ice shelves, which indicates that the observed elevations tend to be more accurate than those estimated from interpolation. Larger biases will be included in the ICESat-2 DEM if the coverage of interpolated elevations is high, hence the elevation gaps in the 500 m DEM are firstly filled by the resampled 1 km DEM to reduce the coverage of interpolated elevations. The accuracy of the ICESat-2 DEM has an obvious relationship with local terrain conditions, and the bias increases when the slope or roughness becomes larger, which is visible for three surface types (Table 5) and different surface slope conditions (Fig. 9). The bias in rocks is obviously larger than those for snow or firn and blue ice areas (BIAs), which is mainly due to the local terrain condition, as they are mostly located in the Transantarctic Mountains and the Antarctic Peninsula, while snow, firn and BIAs tend to have flat surface terrain; hence, they have a smaller bias. While in the low-slope regions, the ICESat-2 DEM shows good agreement with both the OIB and GNSS data, in the large-slope areas, larger biases occur (Fig. 9).
Comparison between the ICESat-2 DEM and OIB airborne elevation measurements (including data in areas of low elevation change from 2009 to 2017 and data in the Antarctica from 2018 to 2019) with respect to three surface types, i.e. snow or firn, blue ice areas (BIAs) and rocks. The surface type data are obtained from Hui et al. (2017).
Scatter plots of ICESat-2 DEM elevation and OIB airborne
elevation
Although OIB airborne data provide an independent evaluation of the
generated DEM, they still cannot present a comprehensive comparison. Most of
the OIB airborne data were obtained in ice sheet margins or mountain
environments, with high slopes and low elevations. Approximately 78 % of
the OIB elevations used were less than 1500 m, and 76 % of the observed surface
slopes from the OIB mission were less than 1
In order to evaluate the DEM performance in more detail, the elevations
along two OIB tracks in the flat ice sheet interior and rough ice sheet margins
are shown in Fig. 10. In the ice sheet interior where surface slopes are
small (Fig. 10a), elevation differences of approximately 5 m can be found
(the median elevation difference for ICESat-2 DEM is
Differences between the ICESat-2 DEM and OIB elevations
along two OIB flight paths in
Elevation differences between the ICESat-2 DEM and six
previously published DEMs, i.e. ICESat DEM
Additionally, by comparing to the OIB or GNSS elevation data (see Sect. 4), we can estimate the actual ICESat-2 DEM uncertainty as the SD of the
differences to OIB or GNSS elevation data. In the estimated uncertainty map
(Fig. 5b), a median value of 5.84
Comparisons between the ICESat-2 DEM, ICESat DEM, ICESat/ERS-1 DEM, Helm CryoSat-2 DEM, Slater CryoSat-2 DEM, REMA DEM, TanDEM PolarDEM and OIB airborne elevation measurements in areas of low elevation change from 2009 to 2019.
Median differences between seven DEMs and OIB airborne elevation
measurements in areas of low elevation change from 2009 to 2019 with respect
to surface slope
When compared to the altimeter-derived DEMs, the elevation difference rises when the surface slope becomes larger, especially in mountainous environments (e.g. Transantarctic Mountains and Antarctic Peninsula, Fig. 11). This may be due to their differences in spatial resolution and measurement accuracy; this effect is considerably reduced when the local terrain is flatter (e.g. ice sheet interior). Compared to the REMA DEM and TanDEM PolarDEM, smaller elevation differences can be found in both the flat ice sheet interior and steep mountains and marginal ice sheets. REMA DEM and TanDEM PolarDEM have much smaller spatial resolutions and better performances and similar elevations indicate the reliability of ICESat-2 DEM in mountain environments. In particular, the ICESat-2 DEM shows a generally higher surface height than the TanDEM PolarDEM, which is assumed to be caused by the penetration depth of the X-band (TerraSAR-X and TanDEM-X) into snowpack (Dehecq et al., 2016; Fischer et al., 2020). Here it should be noted that the time differences between these DEMs can still cause the uncertainties in the DEM comparison.
To indicate a general comparison between the ICESat-2 DEM and other DEMs, OIB airborne data in areas of low elevation change from 2009 to 2019 are used to evaluate individual DEMs and the same evaluation method applied for the ICESat-2 DEM is used (as described in Sect. 2.2). The evaluation result shows that the ICESat-2 DEM has a reliable performance and is comparable to other DEMs derived from altimetry, stereo-photogrammetry and interferometry (Table 6).
The median differences in surface slope and roughness for these seven DEMs
illustrate that all the elevation biases become more uncertain with
increasing slope and roughness (Fig. 12). The REMA DEM always has more
stable performances than the ICESat-2 DEM, as stereo-photogrammetry can
generate more consistent elevation estimations at the regional scale than
altimetry. A similar situation occurs for the TanDEM PolarDEM when slopes are
Here, kinematic GNSS data from 2001 to 2015 in the ice sheet interior are used to construct an additional DEM comparison. The comparison method and adjustment for time difference are the same as for OIB data. As surfaces in the interior of East Antarctica are flat, better performances for all DEM except TanDEM PolarDEM are found than those based on OIB airborne data. Similarly, ICESat-2 DEM has a reliable performance and is comparable to other DEMs (Table 7). Additionally, the accuracy of the ICESat-2 DEM may be related to the surface slope (Fig. 9b); however, as the terrain conditions of GNSS measurement points are relatively flat, this relationship is not obvious. It should be noted that the spatiotemporal coverages of the OIB and GNSS data used are limited here and they cannot provide an unbiased evaluation for ICESat-2 DEM and other DEMs. Hence the comparisons above only give a general reference for the performances and cannot be used as the quantitative accuracy evaluation.
Comparisons between the ICESat-2 DEM, ICESat DEM, ICESat/ERS-1 DEM, Helm CryoSat-2 DEM, Slater CryoSat-2 DEM, REMA DEM, TanDEM PolarDEM and GNSS elevation data in areas of low elevation change from 2001 to 2015.
Compared to other DEMs, the elevation change rate can be obtained when deriving the ICESat-2 DEM, which provides an additional reference for ice topography and mass balance estimation. Additionally, in previous studies several years of altimeter data are needed to derive the DEM in Antarctica. Due to the high-density measurements of ICESat-2, 13 months of ICESat-2 data can be used to generate a DEM for Antarctica and the performance is comparable to other DEMs, indicating that the ICESat-2 DEM can be updated annually. This study demonstrates the feasibility and reliability of using 1-year ICESat-2 data to derive the Antarctic DEM and provides a reference for the processing scheme of DEM (e.g. in higher resolution, regularly updated) based on ICESat-2 in future.
The generated ICESat-2 DEM (including the map of uncertainty) can be
downloaded from the National Tibetan Plateau Data Center, Institute of Tibetan
Plateau Research, Chinese Academy of Sciences at
A new DEM for Antarctica with a posting of 500 m is presented based on the
surface height measurements from ICESat-2 by using a model fitting method.
This DEM has an elevation measurement that accounts for 74 % of
Antarctica, and the remaining 26 % is estimated based on the ordinary
kriging method. The accuracy of the ICESat-2 DEM is evaluated by comparing
it to the independent airborne data from the OIB mission. Overall, the
ICESat-2 DEM shows a median bias of
Compared to DEMs derived from satellite altimeters (i.e. the ICESat DEM, ICESat/ERS-1 DEM, Helm CryoSat-2 DEM, and Slater CryoSat-2 DEM), larger differences are found in regions with high slopes, which is due to their resolution difference, while smaller elevation differences compared to the REMA DEM and TanDEM PolarDEM support the reliability of the ICESat-2 DEM. Based on the OIB airborne data and kinematic GNSS transects, the ICESat-2 DEM shows reliable performance and is comparable to other DEMs, which further demonstrates the reliability of the ICESat-2 DEM.
Here 13 months of ICESat-2 data are used to generate the Antarctic DEM and the evaluation results show that the corresponding DEM is reasonable and valid. This means that the ICESat-2 DEM can be provided in a sustainable way, i.e. this DEM can be updated annually and thus accumulated on an annual basis. Additionally, reasonable elevation change rates can also be obtained when deriving the DEM. The combination of the derived DEMs and elevation change rates can be further used for the references of fieldwork planning, ice motion tracking, numerical modelling of ice sheets and the mass balance estimation. More importantly, these data can be provided on an annual basis, which has a large application potential for Antarctic research especially under climate warming.
XS and YF developed the related algorithm, generated and evaluated the ICESat-2 DEM; LD constructed the comparison to previously published DEM products; CQK supervised this work.
The contact author has declared that neither they nor their co-authors have any competing interests.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
We thank all the data providers for their data. We thank the editors and reviewers for their constructive comments which helped to improve this manuscript.
This research has been supported by the National Natural Science Foundation of China (grant nos. 41976212, 41830105) and the Natural Science Foundation of Jiangsu Province (grant no. BK20210193).
This paper was edited by Baptiste Vandecrux and reviewed by Veit Helm and two anonymous referees.