The amount of reflected energy by snow and ice plays a fundamental role in their melting processes. Different non-ice materials (carbonaceous particles, mineral dust (MD), microorganisms, algae, etc.) can decrease the reflectance of snow and ice promoting the melt. The object of this paper is to assess the capability of field and satellite (EO-1 Hyperion) hyperspectral data to characterize the impact of light-absorbing impurities (LAIs) on the surface reflectance of ice and snow of the Vadret da Morteratsch, a large valley glacier in the Swiss Alps. The spatial distribution of both narrow-band and broad-band indices derived from Hyperion was analyzed in relation to ice and snow impurities. In situ and laboratory reflectance spectra were acquired to characterize the optical properties of ice and cryoconite samples. The concentrations of elemental carbon (EC), organic carbon (OC) and levoglucosan were also determined to characterize the impurities found in cryoconite. Multi-wavelength absorbance spectra were measured to compare the optical properties of cryoconite samples and local moraine sediments. In situ reflectance spectra showed that the presence of impurities reduced ice reflectance in visible wavelengths by 80–90 %. Satellite data also showed the outcropping of dust during the melting season in the upper parts of the glacier, revealing that seasonal input of atmospheric dust can decrease the reflectance also in the accumulation zone of the glacier. The presence of EC and OC in cryoconite samples suggests a relevant role of carbonaceous and organic material in the darkening of the ablation zone. This darkening effect is added to that caused by fine debris from lateral moraines, which is assumed to represent a large fraction of cryoconite. Possible input of anthropogenic activity cannot be excluded and further research is needed to assess the role of human activities in the darkening process of glaciers observed in recent years.
Mountain glaciers represent an important source of fresh water across the planet. These resources are seriously threatened by global climate change (Immerzeel et al., 2010), and a widespread reduction of glacier extension has been observed in recent years (Oerlemans, 2005; Paul et al., 2004). Surface processes that promote glacier melting are driven by both climate (i.e., temperature and precipitation) and changes in albedo. The latter is mainly influenced by the metamorphism of snow in the infrared part of the reflectance spectrum of snow and by the impurity content (such as dust, soot, ash, algae, etc.) in the visible domain (Painter et al., 2007). The studies on the impact of light-absorbing impurities (LAIs) on the cryosphere on a global (Flanner et al., 2007, 2009), regional (He et al., 2014; Lee et al., 2017; Painter et al., 2010; Sterle et al., 2013) and local scale (Oerlemans et al., 2009) suggest that LAI accumulation on snow and ice decreases the albedo, with consequences on the radiative and mass balances, both in mountain glaciers and ice sheets (Dumont et al., 2014; Gabbi et al., 2015; Wittmann et al., 2017).
Carbonaceous particles such as black carbon (BC), elemental carbon (EC) and organic carbon (OC) represent an important class of LAIs because, in accordance to their chemical structure, they absorb visible light with extreme efficiency (Andreae and Gelencsér, 2006; Hartmann et al., 2013). Their role in the climatic system has been largely acknowledged in the scientific literature (e.g., Bond et al., 2013). Given their capability to absorb visible light, carbonaceous particles are responsible of a positive radiative forcing when deposited on snow and ice (Doherty et al., 2010; Hadley and Kirchstetter, 2012; Meinander et al., 2014). Recently, Painter et al. (2013) combined ice core data and climate simulations, suggesting that carbonaceous LAI depositions may have played a crucial role in fostering the end of the Little Ice Age in the European Alps.
Recent studies showed that Alpine glaciers are undergoing a darkening process, and this was ascribed to the impact of regional and/or global warming and to the deposition and accumulation of LAIs on the glacial surface (Azzoni et al., 2016; Gautam et al., 2013; Ming et al., 2012, 2015; Qu et al., 2014). Impurities such as BC are commonly present in the atmosphere in Alpine regions (Lavanchy et al., 1999; Ménégoz et al., 2014; Nair et al., 2013; Sandrini et al., 2014). Their sources can be both natural (e.g., biomass burning, wildfires) and anthropogenic (e.g., fossil fuel and biofuel combustion). Instead, mineral dust (MD) originates from natural sources including areas surrounding the glaciers or distant arid regions (e.g., deserts). The determination of the composition of non-ice material is fundamental to attribute the provenance of these particles and to assess the actual role of anthropogenic and natural activity in decreasing the albedo of glaciers and accelerating the melt.
When impurities are deposited on snow and ice, they can aggregate on the surface of the glacier, forming a characteristic sediment defined as cryoconite (Nordenskiöld, 1883). Cryoconite is constituted of dust and organic matter, so it decreases the albedo of the ice and promotes its melting (Takeuchi et al., 2001). Cryoconite can accumulate in typical holes in the ablation areas of mountain and polar glaciers (i.e., “cryoconite holes”) or can be dispersed on their surface (Stibal et al., 2012). Its geochemical and microbiological composition has been studied in polar and nonpolar glaciers (Aoki et al., 2014; Bøggild et al., 2010; Hodson et al., 2007; Nagatsuka et al., 2014; Takeuchi et al., 2014; Wientjes et al., 2011) to evaluate its provenance and to determine the role of microorganisms (i.e., extremophiles) in sustaining life in such harsh environments. Nevertheless, a lot of uncertainty still exists regarding its formation and evolution. Moreover, most of the literature focused on Arctic, Antarctic and Asian glaciers, while little information is available for the European Alps (Edwards et al., 2013; Franzetti et al., 2016), despite their high sensitivity to environmental and climatic changes (Beniston, 2005). The Alps rise on the top of the Po river plain, one of the most industrialized and densely populated areas of Europe. As a consequence, the presence of anthropogenic activities is expected to strongly influence the geochemical and radiative characteristics of cryoconite (Baccolo et al., 2017; Cook et al., 2015; Hodson, 2014; Tieber et al., 2009).
Within this study, we demonstrate the potential of field and satellite hyperspectral data to characterize the spatial distribution of impurities and cryoconite on the surface of the Vadret da Morteratsch Glacier (Swiss Alps) and to evaluate their effect on snow and ice reflectance. Specific objectives consist in (i) the analysis of the variability of the Morteratsch Glacier surface reflectance combining hyperspectral satellite data and field spectroscopy data, (ii) the exploitation of different narrow-band and broad-band indices to determine the impact of impurities and cryoconite on the glacier, and (iii) the characterization of the optical and chemical properties of cryoconite in the study area.
Location of the Morteratsch Glacier in the Bernina Range. Glacier polygons are extracted from the Randolph Glacier Inventory version 5 (Pfeffer et al., 2014). Oblique lines represent the extension of the Hyperion scene acquired on August 2015. The black star represents the area where field spectroscopy measurements and sampling were conducted. The Vadret Pers and Fellaria glaciers are also displayed and considered in the satellite data analysis.
Vadret da Morteratsch (46
A field campaign was conducted on 30 July 2015 on the ablation zone of the
Morteratsch Glacier (see Fig. 1). Field spectral measurements were collected
to characterize the spectral reflectance of the glacier surface characterized
by different physical conditions and impurity content. Four main classes were
identified by visual inspection: clean ice, dirty ice, melt pond and
cryoconite. A minimum of three different points were measured for each class.
Spectral reflectance was measured with an Analytical Spectral Devices (ASD)
field spectrometer (HandHeld), which collects reflected radiance from 325 to
1075 nm with a spectral sampling interval of 1 nm and a spectral resolution
(full width at half maximum, FWHM) of 3.5 nm at the band centered at
700 nm. The hemispherical conical reflectance factor (HCRF)
Samples of ice, cryoconite and debris from the lateral moraine were collected
(three samples for each class) in the areas measured with the field
spectrometer. Ice sampling was conducted by crushing the surface ice, cryoconite
samples were picked using a small spoon and lateral moraine samples were
collected near the glacier terminus. All samples were stored in sterilized
Corning tubes (50 mL). They were kept frozen and taken back to the
university campus (Milano-Bicocca), where they were stored at
Samples of cryoconite collected from the glacier surface were stored in
frozen conditions until the preparation of the samples for the analyses.
After melting, they were decanted for several hours to separate solid and
liquid fractions. The liquid part was removed, and solid cryoconite was
successively dried at 60
For the elemental carbon and organic carbon determination, a TOT instrument
(Sunset Lab Inc.) was used, applying the NIOSH 5040 protocol (Birch and Cary,
1996). According to the instrument manual, an uncertainty of 8–10 % is
associated with the retrieved EC and OC concentrations. The original
solutions have been subjected to a chemical determination of the levoglucosan
(1,6-anhydro-
Hyperspectral reflectance spectra were acquired also in laboratory samples on dried cryoconite and moraine sediment samples. Each sample was lighted with a halogen stable light source (1000 W, LOT Quantum Design). The lamp produces a uniform collimated output beam (with 50 mm diameter) that provides a homogenous stable illumination of the sampled area. Reflectance was measured with an ASD field spectrometer (FieldSpec Pro) that operates from 350 to 2500 nm with an FWHM of 5–10 nm and a spectral sampling of 1 nm, which is different from the one used in the field.
Remotely sensed data used in this study were collected with the hyperspectral Hyperion sensor onboard on the NASA Earth Observing 1 (EO-1) satellite mission, launched in November 2000 (Middleton et al., 2013). The Hyperion sensor features a swath of 7.7 km and a spatial resolution of 30 m. It collects spectral radiance from 400 to 2400 nm with a spectral resolution of 10 nm (242 bands). The signal-to-noise ratio of Hyperion data varies from 150 : 1 (for 400–1000 nm) to 60 : 1 (for 1000–2000 nm). Hyperion reflectance retrievals have been validated several times with independent measurements, as well as with airborne sensor such as AVIRIS (Kruse et al., 2003).
The Hyperion image was acquired on 7 August 2015 (close to the field
campaign) with a footprint covering the sites visited in the field campaign
and comprising the entire Morteratsch, Pers, and Fellaria glaciers and also other
minor glaciers (Fig. 1). The look angle of Hyperion was 23
The goodness of the Hyperion atmospheric correction was evaluated with
concurrent Landsat 8 Operational Land Imager (OLI) data acquired a few
minutes before Hyperion on 7 August. The OLI surface reflectance Climate Data
Record product, already
corrected for the influence of the atmosphere (Vermote et al., 2016), was
compared with atmospherically corrected Hyperion data. The Hyperion scene was
classified with respect to land cover using a support vector machine (SVM)
algorithm (Wu et al., 2004). SVM is a supervised classification method
derived from statistical learning theory. The algorithm separates the classes
with a decision surface (i.e., optimal hyperplane) that maximizes the margin
between the classes. The full Hyperion spectrum was used as input for the
spectral discrimination of the classes. SVM has already been successfully
applied to complex and noisy data such as the Hyperion ones (Petropoulos et
al., 2012). We manually selected the training set for eight predefined land
cover classes: snow, bare ice, debris cover, lakes, rocks, and sparse
vegetation; clouds and shadows were also included. The goodness of the
classification was evaluated by considering 100 randomized points (weighted
for each class) and building a confusion matrix to estimate the producer's
and user's accuracies for each class, as well as the global accuracy of the
SVM classification. For the two main classes of interest (snow and bare ice),
the ratio between training and test set pixels was
Mean and standard deviation of the reflectance spectra acquired on
the glacier ablation zone with the FieldSpec ASD:
From Hyperion surface reflectance data, the snow darkening index (SDI) (Di
Mauro et al., 2015) was calculated as a normalized difference between red
and green bands (Eq. 1).
The impurity index (
Finally, Hyperion data were used to estimate the glacier albedo (
In this study, we did not characterize the anisotropy of snow on the glacier,
so we were not able to perform a proper estimate of hemispherical albedo (see
Naegeli et al., 2015), but we used
The narrow-band (SDI,
Nonlinear ordinary least squares (nOLS) regression (
In order to assess the sensitivity of the SDI,
Figure 2 shows the spectral behavior of the four classes (clean ice, dirty
ice, melt pond and wet cryoconite) measured in the field and the two classes
(dry cryoconite and moraine sediment) measured in the laboratory. Mean and
standard deviation are obtained by averaging spectra of the same class. Clean
ice shows an average reflectance in the visible and near-infrared wavelengths
which is higher with respect to the other classes (Fig. 2a). Dirty ice shows
very low reflectance values of circa 0.3 in the visible wavelengths and
lower than 0.2 in the near infrared. Melt pond shows an enhanced absorption
at 1000 nm, and wet cryoconite shows the lowest average reflectance values
(i.e.,
Comparison between ASD field reflectance spectra (red line) for the class dirty ice and Hyperion reflectance spectra (blue line) for the class bare ice.
Confusion matrix obtained from the validation scheme of the support vector machine (SVM) classification of Hyperion data. Rows represent the true classes, and columns represent the predicted classes. User and producer accuracies are reported for each class. Global accuracy is also displayed.
The presence of impurities in glacier ice strongly alters its optical
properties, reducing its reflectance. Results from the linear ordinary least
squares (OLS) regressions between SDI,
While field spectroscopy enables characterizing the heterogeneous surface of
the ablation zone with high precision, remote sensing data allow
investigating the optical properties of the glacier as a whole. A comparison
of field and satellite data is shown in Fig. 4, where ASD reflectance of
dirty ice is compared with the average spectra of the bare ice class from
Hyperion data (RMSE
Figure 5 shows the result of the SVM classification applied to the whole Hyperion scene. The SVM algorithm properly separates different surfaces in the Bernina Range. Table 1 presents the confusion matrix obtained from the validation scheme. Global accuracy of the classification is 78 %. The user's accuracy for the snow and bare ice classes is 90 and 70 %, while the producer's accuracy is 85.7 and 73.7 % respectively. The rocks and debris cover classes are sometimes misclassified; their user accuracy is 60 and 70 % respectively. The class with the lower accuracy (40 %) is lakes.
Examples of spectra extracted from Hyperion data for the classes snow, bare ice
and debris cover classes for the Morteratsch and Pers glaciers are shown in Fig. 6a.
The accumulation zone of the Morteratsch Glacier shows an overall higher reflectance
in the visible and near-infrared wavelengths, with enhanced absorption
features at 1030 and 1250 nm. Beyond 1400 nm almost all the radiation is
absorbed by snow. The accumulation zone of the Pers Glacier shows a lower
reflectance in the visible and near-infrared wavelength, with stronger
absorption before 600 nm. Ice reflectance from both Morteratsch and Pers
glaciers is lower than snow reflectance; in particular, the absorption at
1030 nm is more pronounced and shifted to lower wavelengths. The reflectance
spectrum of debris cover shows different features and an overall higher
reflectance in the shortwave infrared. Figure 6b shows a comparison between
the atmospherically corrected reflectance from Hyperion and Landsat 8 OLI
(mean and standard deviation) (RMSE
Figure 7 shows the
Summary of elemental carbon (EC), organic carbon (OC), total carbon (TC) and levoglucosan concentrations in cryoconite samples (CR3–CR6).
Comparison between SDI,
In Fig. 8, we present the contour plots obtained from the SNICAR simulations.
Plots refer to the sensitivity of narrow- and broad-band indices to MD
variations (upper panels) and to BC variations (lower panels).
Table 2 shows the concentration of elemental carbon, organic carbon, total carbon (TC, calculated as the sum of EC and OC) and levoglucosan in the cryoconite samples. Different samples show similar values of EC and OC concentrations respectively ranging from 0.3 to 0.4 % for EC and from 4.2 to 5.4 % for OC (Fig. 9). Traces of levoglucosan were found only in cryoconite samples CR5 and CR6.
Concentration of organic carbon (blue bars) and elemental carbon (red bars) for different cryoconite samples (CR3–CR6).
The comparison between the MWAA of a cryoconite (CR4) and a
moraine sediment (MS1) sample is presented as an example in Fig. 10. The
MWAA measures the absorbance (Abs) of samples at five different
wavelengths. This parameter quantifies the fraction of light absorbed by the
sample and it is defined as Abs
The results presented here confirm the effect of impurities and cryoconite in reducing the spectral reflectance of snow and ice in the European Alps (Naegeli et al., 2015). The combined use of ground and satellite hyperspectral observations and the exploitation of narrow-band and broad-band indices provide original data to evaluate the effect of impurities and cryoconite in the Morteratsch Glacier. Furthermore, the laboratory determination of the composition of non-ice material suggests an important role of carbonaceous material in the darkening of the Morteratsch ablation zone. This effect is superimposed on the one caused by fine debris from the lateral moraine.
From field spectroscopy, we were able to characterize different glacier
components in the ablation zone only, while satellite data allowed having an
overview on the reflectance spatial variability on a catchment scale. We
tried to select flat areas for the reflectance measurements. However, the
surface of the glacier was quite rugged, so possible uncertainties related to
the forward scattering of snow may be present in the data. Reflectance higher
than 1 in the visible wavelengths is often found in the literature (Painter
and Dozier, 2004; Schaepman-Strub et al., 2006) and can be related to forward
scattering, glacier surface slope, sensor tilt, inappropriate bidirectional
reflectance distribution function
correction, etc. Over the ablation zone of the Morteratsch Glacier, the
optical properties of ice are largely variable probably because of the
patterns of melting, refreezing and surface runoff that shape the local
glacier morphology. The presence of impurities and cryoconite causes a
dramatic reduction of the spectral reflectance of ice. In the visible
wavelengths, the ice reflectance decreases from circa 1 to 0.3 due to
impurity presence and to less than 0.1 due to cryoconite presence (see
Fig. 2). The decrease in reflected energy has an important impact on the
radiative balance of a retreating glacier such as the Morteratsch (Klok et
al., 2004; Oerlemans et al., 2009). We showed that the concentration of
impurities and
In this study, we evaluated the reliability of the Hyperion reflectance through a comparison with Landsat and ASD reflectance, and we obtained respectively an RMSE of 0.015 and 0.03. The comparison between field and Hyperion reflectance spectra shows that ASD field spectra of the class dirty ice are comparable with those measured from the Hyperion sensor. Hyperion spectra show a decrease in reflectance before 500 nm; this could be due to the presence of contaminated (non-pure) pixels of snow and ice, as previously observed by Negi et al. (2013). Otherwise, it could be related to the presence of meltwater increasing the absorption of solar radiation during the melting season, as observed from airborne hyperspectral reflectance data in other glaciers of the European Alps (Naegeli et al., 2015). Spectra of bare ice exhibit low reflectance, showing an enhanced absorption around 1030 nm, slightly shifted to lower wavelengths (Dumont et al., 2017; Green et al., 2002).
The classification of Hyperion data using the SVM algorithm provides
satisfying results for the classes snow and bare ice (user's accuracies of 90
and 70 %), confirming the successful use of Hyperion for snow and ice
monitoring (Bindschadler and Choi, 2003; Casey et al., 2012; Doggett et al.,
2006; Negi et al., 2013; Zhao et al., 2013). For other land cover classes
(e.g., rocks and debris cover), a lower user's accuracy was obtained
(60–70 %), probably due to both the coarse spatial resolution (30 m) of
Hyperion for mapping these classes and the spectral similarity between the
two classes. Hyperion data provide important information regarding the
optical properties of snow and ice in relation to the impact of
light-absorbing impurities and grain size. In particular,
In agreement with previous satellite investigations on Alpine glaciers, low
The spatial resolution of the optical satellites used in this study (i.e., 30 m) hampers the discrimination of the impact of different impurities in Alpine glaciers. New opportunities are now offered by the new ESA Sentinel-2 mission featuring a higher spatial resolution in visible and near-infrared wavelengths (Drusch et al., 2012). With the decommission of the EO-1 mission (22 February 2017), there will be no hyperspectral sensors orbiting the Earth. Hyperspectral data are very useful for monitoring the optical properties of snow and ice (Painter et al., 2016), and future satellite missions such as EnMAP (Environmental Mapping and Analysis Program) (Stuffler et al., 2007) and PRISMA (Hyperspectral Precursor and Application Mission) (Labate et al., 2009) will provide new important hyperspectral data for cryosphere monitoring both in Alpine and polar regions.
A previous work on the Morteratsch Glacier (Oerlemans et al., 2009) showed that the ablation zone is undergoing a darkening process. The authors attributed the albedo decrease to the accumulation of dust from the moraines surrounding the glacier. In this study, we show how spectral reflectance is distributed over this glacier and in particular over the ablation zone. The presence of cryoconite and impurities over the bare ice in Alpine glaciers plays an important role in albedo decrease and melting enhancement (Cook et al., 2015; Takeuchi et al., 2001). The spectral reflectance of ice gathered from field measurements showed a decrease from 0.9 to 0.1 in visible wavelengths due to the presence of surface non-ice material. Furthermore, the spectral reflectance of moraine sediments showed overall higher values than both wet and dry cryoconite samples. These results suggest considering the impact of organic and inorganic material on albedo decrease for glacier mass balance modeling.
Regarding the characterization of cryoconite, the chemical and thermo-optical analyses allow us to produce a first overview on the composition and the light-absorbing properties of the material contained in cryoconite. The concentrations of OC (5 % of the total weight) found in cryoconite are comparable with independent measurements on other glaciers (Stibal et al., 2008; Takeuchi et al., 2001). We found no concentration of EC presented in the literature on cryoconite. The presence of EC (as a proxy of BC) in the cryoconite suggests a possible influence of anthropogenic activities (e.g., fossil fuel combustion) in the formation and evolution of cryoconite. Furthermore, carbonaceous particles are stronger absorbers in visible and near-infrared wavelengths than mineral particles, and they can heavily enhance the radiation absorbed by bare ice during summer season (Li et al., 2017). In Fig. 10, we show the absorbance and MAC of cryoconite and moraine sediments. A considerable difference was found between the two samples; in particular, cryoconite samples show absorbance values two times higher than those of the moraine. This suggests that the input of moraine sediment alone could not explain the darkening observed on the glacier surface, as previously proposed (Oerlemans et al., 2009). The interaction between mineral material and microbiological activity, with the consequent accumulation of organic matter in cryoconite, further increases the absorption of solar radiation, promoting the melt. Additionally, MAC values also reflect this behavior, since they are calculated by normalizing the absorbance with the dry weight of the samples. The presence of levoglucosan in two cryoconite samples also suggests a possible input from biomass burning aerosols, which may represent an important source of carbon in Alpine ranges.
The interaction dynamic between carbonaceous particles and microorganisms living in the cryoconite is still largely unknown (Cook et al., 2015), and the lack of EC and BC measurements in cryoconite in the scientific literature did not allow us to perform quantitative comparisons with other glaciers. We can assess that the source of carbon can be classified on the basis of the optical properties of particles (Massabò et al., 2015), but new methodologies have to be developed to achieve this comparison for cryoconite samples. The main sources of carbonaceous particles in the atmosphere at mid-latitude are the combustion of fossil fuels and biofuels, and biomass burning. The source partitioning of BC and EC is still missing for Alpine glaciers, and further research is needed to assess the impact of natural and anthropogenic carbon emissions on the albedo of glaciers in different mountain ranges and on the margin of ice sheets (Musilova et al., 2016; Tedesco et al., 2016).
In this paper, we show how non-ice materials influence the optical properties
of a valley glacier (Morteratsch) in the European Alps. Results from field
campaigns and satellite hyperspectral data show that impurities and
cryoconite decrease ice spectral reflectance in visible wavelengths from 0.9
to 0.1 and 0.06 respectively, adding a consistent input of energy to the
glacier radiative balance. Hyperion satellite data show low
Beside climatic drivers, glacier darkening due to multiple processes can promote further glacier shrinkage and therefore a considerable loss of fresh water reservoir in the European Alps. Further analyses are needed to assess the possible impact of anthropogenic activities in glacier darkening processes. Field observations represent an important tool to validate satellite retrieval of surface reflectance, and the characterization of non-ice material that decreases the reflectance of ice is fundamental for evaluating their provenance. The impact of organic and inorganic material on snow and ice optical properties is receiving much attention from the scientific community, and the inversion of newly developed radiative transfer models (Cook et al., 2017; Libois et al., 2013) will allow estimating the concentration of different impurities from future multi- and hyperspectral remote sensing observations.
Data used in this paper will be made available upon request to the first author (biagio.dimauro@unimib.it).
BDM conceived the idea of the paper, organized the field campaign and wrote the manuscript with discussions and contributions from all the other authors. BDM, GB and RG collected the samples and acquired spectral reflectance on the Morteratsch Glacier. MR and CG helped in data interpretation. DM and AP performed the chemical and optical analyses of cryoconite and helped in their interpretation. RC and MR supervised the research.
The authors declare that they have no conflict of interest.
We acknowledge Lawrence Ong, Elizabeth M. Middleton and Petya K. E. Campbell (NASA-GSFC) for scheduling the EO-1 satellite data acquisition. The samples from Morteratsch were stored at the EuroCold facility of the Earth and Environmental Sciences Department of the University of Milano-Bicocca. We thank the PIs of the AERONET stations of Davos (Natalia Kouremeti) and Ispra (Giuseppe Zibordi). We also thank the editor and the two anonymous reviewers for the comments on a previous version of the manuscript. Edited by: Marie Dumont Reviewed by: two anonymous referees