ArticlePDF Available

Theia Snow collection: High-resolution operational snow cover maps from Sentinel-2 and Landsat-8 data

  • Centre National d’Etudes Spatiales | CNES

Abstract and Figures

The Theia Snow collection routinely provides high-resolution maps of the snow-covered area from Sentinel-2 and Landsat-8 observations. The collection covers selected areas worldwide, including the main mountain regions in western Europe (e.g. Alps, Pyrenees) and the High Atlas in Morocco. Each product of the Theia Snow collection contains four classes: snow, no snow, cloud and no data. We present the algorithm to generate the snow products and provide an evaluation of the accuracy of Sentinel-2 snow products using in situ snow depth measurements, higher-resolution snow maps and visual control. The results suggest that the snow is accurately detected in the Theia snow collection and that the snow detection is more accurate than the Sen2Cor outputs (ESA level 2 product). An issue that should be addressed in a future release is the occurrence of false snow detection in some large clouds. The snow maps are currently produced and freely distributed on average 5 d after the image acquisition as raster and vector files via the Theia portal (
Content may be subject to copyright.
Earth Syst. Sci. Data, 11, 493–514, 2019
© Author(s) 2019. This work is distributed under
the Creative Commons Attribution 4.0 License.
Theia Snow collection: high-resolution operational snow
cover maps from Sentinel-2 and Landsat-8 data
Simon Gascoin1, Manuel Grizonnet2, Marine Bouchet1, Germain Salgues2,3, and Olivier Hagolle1,2
1CESBIO, Université de Toulouse, CNRS/CNES/IRD/INRA/UPS, Toulouse, France
2CNES, Toulouse, France
3Magellium, Toulouse, France
Correspondence: Simon Gascoin (
Received: 22 November 2018 – Discussion started: 29 November 2018
Revised: 16 March 2019 – Accepted: 20 March 2019 – Published: 16 April 2019
Abstract. The Theia Snow collection routinely provides high-resolution maps of the snow-covered area from
Sentinel-2 and Landsat-8 observations. The collection covers selected areas worldwide, including the main
mountain regions in western Europe (e.g. Alps, Pyrenees) and the High Atlas in Morocco. Each product of
the Theia Snow collection contains four classes: snow, no snow, cloud and no data. We present the algo-
rithm to generate the snow products and provide an evaluation of the accuracy of Sentinel-2 snow products
using in situ snow depth measurements, higher-resolution snow maps and visual control. The results suggest
that the snow is accurately detected in the Theia snow collection and that the snow detection is more accu-
rate than the Sen2Cor outputs (ESA level 2 product). An issue that should be addressed in a future release
is the occurrence of false snow detection in some large clouds. The snow maps are currently produced and
freely distributed on average 5 d after the image acquisition as raster and vector files via the Theia portal
1 Introduction
The snow cover is an important driver of many ecological,
climatic and hydrological processes in mountain regions and
in high latitude areas. In these regions, in situ observations
are generally insufficient to characterize the high spatial vari-
ability of the snowpack properties. The snow cover was in-
cluded as one of the 50 essential climate variables (ECVs)
to be monitored by satellite remote sensing by the Global
Observing System for Climate (GCOS) in accordance with
the Committee on Earth Observation Satellites (CEOS) agen-
cies. ECVs are intended to support the work of the United
Nations Framework Convention on Climate Change and the
Intergovernmental Panel on Climate Change.
The snow cover is not a variable sensu stricto, but an ob-
ject which can be characterized through many variables, in-
cluding snow-covered area (SCA), fractional area (fSCA),
albedo, liquid water content, snow depth and snow water
equivalent (Frei et al., 2012). An international survey by the
Cryoland consortium (,
last access: 12 April 2019) shed light on the user require-
ments for snow services based on satellite remote sensing.
SCA and fSCA products were ranked as the most impor-
tant by the respondents (Malnes et al., 2015). The survey
also indicated the need for operational products provided on
a daily basis, with latency times shorter than 12 h. High-
resolution data down to 50 m resolution were sought by road
and avalanches authorities (Malnes et al., 2015). The respon-
dents requested regional products, e.g. on the scale of entire
mountain ranges like the Alps or even the whole of Europe,
and preferred the Universal Transverse Mercator (UTM) pro-
jection (Malnes et al., 2015). At the national level in France
there is also a need for operational high-resolution snow-
covered area maps as revealed by the recent roadmap for
satellite applications issued by the French Government (Plan
d’applications satellitaires 2018, Commissariat général au
développement durable, 2018). Based on a wide panel of end
users, this plan selected the monitoring of the snow-covered
area in French national parks as a one of the key actions
Published by Copernicus Publications.
494 S. Gascoin et al.: Theia Snow collection
which should be achieved using Earth observation satellites
in the near future (2018–2022).
Operational SCA maps have been generated from satellite
observations since the 1960s (Ramsay, 1998). Current SCA
and fSCA products are mostly derived from MODIS data
(Hall et al., 2002; Sirguey et al., 2008; Painter et al., 2009;
Metsämäki et al., 2015), but their spatial resolution (1 km to
250 m) can be too coarse for various applications, especially
in mountain regions where the snow cover properties vary on
scales of 10 to 100 m (Blöschl, 1999). High-resolution (30 m)
snow cover maps can be generated from Landsat images but
the low temporal revisit of the Landsat mission (16 d) is an
important limitation for snow cover monitoring, especially
considering that the cloud probability can exceed 50 % in
temperate mountains (Parajka and Blöschl, 2008; Gascoin
et al., 2015). Since 2017, with the launch of Sentinel-2B, the
Copernicus Sentinel-2 mission offers the unique opportunity
to map the snow cover extent at 20m resolution with a re-
visit time of 5 d (cloud permitting) (Drusch et al., 2012). The
combination of Sentinel-2 and Landsat-8 data provides the
opportunity for even more frequent observations of the snow
cover with a global median average revisit interval of 2.9 d
(Li and Roy, 2017).
The principles of snow detection from multispectral op-
tical imagery is well established since the pioneering work
of Dozier (1989) with the Landsat Thematic Mapper. Today,
the challenge for scientists and end users is rather to cope
with the large amount of data that is generated by a mission
like Sentinel-2. The generation of a single snow map from
a Sentinel-2 level-1C tiled product (monodate orthorectified
image expressed in top-of-atmosphere reflectance) involves
downloading a 700+Mb zip file (once uncompressed, the
product contains 12 folders and 108 files, including 15 raster
files and 13 XML metadata files). Since March 2018, ESA
began the operational processing of level 2A (L2A) prod-
ucts (monodate orthorectified images expressed in surface
reflectance, including a cloud mask). Each ESA L2A prod-
uct also includes a snow cover mask. However, the size of
a single L2A product before unzipping can exceed 1 Gb (15
folders, 137 files). In addition, the quality of the ESA L2A
snow mask can be improved for two reasons: (i) ESAs L2A
algorithm is a general-purpose algorithm, which was not op-
timized for snow detection, and (ii) because ESA’s L2A pro-
cessor treats each image independently (i.e. mono-date ap-
proach, Louis et al., 2016), the output cloud mask has a lower
accuracy than a cloud mask generated by a multi-temporal
algorithm (Hagolle et al., 2010).
In this article we introduce the Theia Snow collection, a
new collection of snow cover maps, which are derived from
Sentinel-2 at 20 m resolution in an operational context. The
Theia Snow collection was recently upgraded to also provide
snow cover maps at 30 m resolution from Landsat-8 using
the same algorithm. The data are currently being produced
and freely distributed via the Theia land data centre.1The
snow retrieval algorithm is based on the Normalized Differ-
ence Snow Index (Dozier, 1989) but also uses a digital ele-
vation model to better constrain the snow detection. We first
describe the algorithm (Sect. 2) and its implementation in the
let-it-snow processor (LIS, Sect. 2.4.4). Then, we provide a
detailed description of the product characteristics (coverage,
period, format, etc.; Sect. 3). In Sect. 4, we present an evalu-
ation of Theia snow products using in situ snow depth mea-
surements, very high-resolution clear-sky snow maps and vi-
sual control. We finally discuss the main limitations and po-
tential applications of the Theia Snow collection.
2 Algorithm
2.1 Scope
An algorithm was designed to determine the snow presence
or absence from Sentinel-2 and Landsat-8 observations out-
side areas of dense forest with the following requirements:
It should be scalable, i.e. allow the processing of large
areas (104km2) with a reasonable computation cost
(typically less than 1 h for a single product).
It should be robust to seasonal and spatial variability of
the snow cover and land surface properties.
It should maximize the number of pixels that are classi-
fied as snow or no snow.
It is preferable to falsely classify a pixel as cloud than
falsely classify a pixel as snow or no snow.
2.2 Input
The algorithm works with multispectral remotely sensed im-
ages, which include at least a channel in the visible part
of the spectrum and a channel near 1.5 µm (referred to as
shortwave-infrared or SWIR). It takes the following as input:
a L2A product, including
the cloud and cloud shadow mask (referred to as an
“L2A cloud mask” in the following),
the green, red and SWIR bands from the flat-surface
reflectance product. These images are corrected
for atmospheric and terrain slope effects (Hagolle
et al., 2017). The slope correction is important in
mountain regions since it enables us to use the same
detection thresholds whatever the sun-slope geom-
A digital elevation model (DEM).
1Theia is a French inter-agency organization designed to foster
the use of Earth observation for land surface monitoring for aca-
demics and public policy actors.
Earth Syst. Sci. Data, 11, 493–514, 2019
S. Gascoin et al.: Theia Snow collection 495
2.3 Pre-processing
The red and green bands are resampled with the cubic
method to a pixel size of 20 m by 20 m to match the resolu-
tion of the SWIR band. The DEM is generated from the Shut-
tle Radar Topography Mission seamless DEM (Jarvis et al.,
2008) by cubic spline resampling to the same 20 m resolution
2.4 Algorithm description
2.4.1 Snow detection
The snow detection is based on the Normalized Difference
Snow Index (NDSI, Dozier, 1989) and the reflectance in the
red band. The NDSI is defined as
NDSI =ρgreen ρSWIR
ρgreen +ρSWIR
where ρgreen (or ρSWIR) is the slope-corrected surface re-
flectance in the green band (or SWIR at 1.6 µm). The NDSI
expresses the fact that only snow surfaces are very bright
in the visible and very dark in the shortwave infrared. Tur-
bid water surfaces like some lakes or rivers may also have
a high NDSI value; hence a additional criterion on the red
reflectance is used to avoid false snow detection in these ar-
eas. A cloud-free pixel is classified as snow if the following
condition is true:
(NDSI > ni)and (ρred > ri),(2)
where niand riare two parameter pairs (i.e. i∈ {1,2}) as ex-
plained below. Otherwise the pixel is marked as “no snow”.
The values of the parameters are provided in Table 1.
2.4.2 Snow line elevation
The algorithm works in two passes. The snow detection
(Sect. 2.4.1) is performed a first time using parameters n1and
r1, which are set in such way as to minimize the number of
false snow detections (Table 1). This first pass enables us to
estimate the minimum elevation of the snow cover zs, above
which a second pass will be performed with less conserva-
tive parameters values n2and r2(Table 1). An example of
the evolution of the snow line elevation over the tile 31TCH
in the Pyrenees is shown in Fig. 1 and the corresponding
snow maps are shown in Fig. 1. Based on regional analy-
ses of the snow line elevation variability in mountain ranges,
(e.g. Gascoin et al., 2015; Krajˇ
cí et al., 2016), we assume that
large altitudinal variations of the snow line elevation are not
likely on the scale of a tile of 110 km by 110 km. Therefore,
the minimum snow elevation zsis considered uniform on the
scale of the processed image.
Before proceeding to pass 2, the total snow fraction in the
image after pass 1 is computed. If this snow fraction is below
ft(Table 1), then pass 2 is skipped, as the sample of snow
pixels is not considered statistically significant for determin-
ing the snow line elevation. Otherwise, the pass 2 is activated.
For that purpose, the DEM is used to segment the image in
elevation bands with a fixed height of dz(Table 1). Then,
the fraction of the cloud-free area that is covered by snow
in each band (after pass 1) is computed. The algorithm finds
the lowest elevation band bat which the snow cover fraction
is greater than a given fraction fs(Table 1). The value of zs
is the lower edge of the elevation band that is two elevation
bands below band b.
2.4.3 Cloud mask processing
The cloud mask in the input L2A product is conservative
because (i) it is computed at a coarser resolution and (ii) it
was developed to remove surface reflectance variations due
to cloud contamination. However, the scattering of some thin
clouds is low in the SWIR, green and red bands, which are
used for snow detection (Sect. 2.4.1). Hence, the human eye
can see the snow (or the absence thereof) through these semi-
transparent clouds in a colour composite. In addition, the
L2A cloud mask tends to falsely classify the edges of the
snow cover as cloud. Therefore, the algorithm includes some
additional steps to recover those pixels from the L2A cloud
mask and reclassify them as snow or no snow. This step is
important because it substantially increases the number of
observations as specified in Sect. 2.1.
A pixel from the L2A cloud mask cannot be reclassified as
snow or no snow if any of these conditions are satisfied:
it is coded as “cloud shadow” in the L2A cloud mask;
it is coded as “high-altitude cloud” (or “cirrus”) in the
L2A cloud mask;
it is not a “dark cloud” (see below).
The cloud shadows are excluded because the signal-to-
noise ratio can be very low in these areas. The high clouds
are excluded because they can have a similar spectral signa-
ture as the snow cover, i.e. a high reflectance in the visible
and a low reflectance in the SWIR. This type of cloud are
detected in Sentinel-2 and Landsat-8 L2A products based on
the spectral band centred on the 1.38 µm wavelength (Gao
et al., 1993).
We select only the dark clouds as possible reclassifica-
tion areas, because the NDSI test is robust to the snow–
cloud confusion in this case. The dark clouds are defined us-
ing a threshold in the red band after downsampling the red
band by a factor rfusing the bilinear method. This resam-
pling is applied to smooth locally anomalous pixels, follow-
ing the MAJA algorithm, which performs the cloud detection
at 240 m for L2A products (Hagolle et al., 2017). Therefore,
if a (non-shadow, non-high-cloud) cloud pixel has a red re-
flectance at this coarser resolution that is lower than rD(Ta-
ble 1), then it is temporarily removed from the cloud mask Earth Syst. Sci. Data, 11, 493–514, 2019
496 S. Gascoin et al.: Theia Snow collection
Figure 1. Time series of snow and cloud cover area distribution by elevation band over tile 31TCH (tile location in Fig. 5). The dashed red
line indicates the position of zsas determined by the LIS algorithm. The corresponding snow maps are shown in Fig. 2.
Figure 2. Time series of snow maps over tile 31TCH (tile location in Fig. 5). The corresponding charts of snow and cloud cover area
distribution by elevation band are shown in Fig. 1.
Earth Syst. Sci. Data, 11, 493–514, 2019
S. Gascoin et al.: Theia Snow collection 497
and proceeds to the pass 1 snow detection. The new cloud
mask at this stage is the pass 1 cloud mask in Fig. 3.
After passing the passes 1 and 2, some former cloud pix-
els (pixels that were originally marked as cloud in the L2A
cloud mask) will not be reclassified as snow because they did
not fulfil the conditions of Eq. (2). These pixels are flagged
as cloud in the final snow product if they have a reflectance
in the red that is greater than rB(Table 1). Otherwise they
are classified as no snow. Here, the red band at 20 m resolu-
tion is used to allow an accurate detection of the snow-free
areas near the snow cover edges that the L2A product tend to
falsely mark as cloud. The resulting cloud mask is the pass 2
cloud mask in Fig. 3.
2.4.4 Implementation
The algorithm was implemented in an open-source processor
called let-it-snow (LIS). LIS is written in Python 2.7 and C++
and relies on the Orfeo ToolBox and GDAL libraries. GDAL
is used for input and output operations, image resampling
and metadata access (GDAL/OGR contributors, 2018). Or-
feo Toolbox enables us to make the computations with good
performances under memory constraints (i.e. the amount of
available memory can be predefined), which is critical for
operational production (Grizonnet et al., 2017). LIS takes
as input a digital elevation model and a Theia L2A prod-
uct from Sentinel-2 MSI, Landsat-8 OLI, SPOT-4 HRVIR or
SPOT-5 HRG (Gascoin et al., 2018). The spatial resolution
and central wavelength of the spectral bands used by LIS for
each sensor are given in Table 2. It is also compatible with
Landsat-8 USGS level 2 products (U.S. Geological Survey
Earth Resources Observation And Science Center, 2014) and
Sentinel-2 ESA’s Sen2Cor level 2A products (Louis et al.,
The parameter values were set based on previous studies
with Landsat (Hagolle et al., 2010; Zhu et al., 2015; Gascoin
et al., 2015) and by visually checking many snow maps and
snow fraction histograms. From this set of a priori parameter
values, only r2was adjusted based on the analysis of a first
batch of products to enhance the snow detection on shaded
3 Data description
The Theia Snow collection data can be freely accessed us-
ing a web browser by connecting to (last
access: 12 April 2019) or using this free command-line
tool: (last
access: 12 April 2019). The user must first create an account
in Theia to have the permission to download the data.
The Theia Snow collection is organized following the
tiling system of Sentinel-2 (Fig. 4). Each tile covers a square
area of 110 km by 110 km in the UTM coordinate system.
There are currently 127 tiles in the Theia Snow collection,
mostly over the mountain ranges of western continental Eu-
rope (France, Spain, Switzerland, Italy, western Austria).
The snow products are also provided for southern Quebec in
Canada, the Issyk-Kul lake catchment in Kyrgyzstan and the
Kerguelen Islands. These extra-European sites were selected
to support specific ongoing projects in relation with Theia.
An opportunistic tile is produced in central Nevada, USA,
because this tile is already available as a level-2A version
for calibration purposes. The Theia Snow collection covers
a number of mountain ranges with seasonal snow cover (Ta-
ble 3).
The Theia Snow collection products are available for 127
tiles, starting on December 2017. At the time of writing there
were over 17 000 available products. A set of 1300 products
tagged as version 1.0 are available between November 2015
and June 2017 for a subset of 15 tiles (Table 3). These prod-
ucts were produced in pre-operational using a different con-
figuration of LIS, which underestimated the snow-covered
area on shaded slopes. All other products were generated
with the same configuration as presented in this article. The
products after December 2017 can still have a different ver-
sion tag because of changing versions in the upstream L2A
product. However, the different L2A versions have a low
impact on the snow products. It is expected that the cloud
mask has improved gradually with time, due to algorithms
enhancements in the successive L2A versions and also be-
cause of the increasing revisit frequency of the Sentinel-2
mission (the full 5 d revisit became nominal above all land
masses in March 2018). The different versions of the Theia
Snow collection are listed and updated on this page: http:// (last ac-
cess: 12 April 2019).
The snow products are currently routinely generated at
CNES using the MUSCATE scheduler, which also manages
the L2A production for Theia. MUSCATE performs a series
of operations in a high-performance computing environment:
(i) download the level 1C product as soon it is available from
the French mirror of the Sentinel products (PEPS, Plateforme
d’Exploitation des Produits Sentinel), (ii) process and dis-
tribute the L2A product and (iii) process and distribute the
snow product. This workflow is optimized to reduce the lag
time between the acquisition and the distribution of the prod-
ucts. As an example, in April 2018, nearly 80 % of the snow
products were made available online 5 d or less after the date
of the Sentinel-2 acquisition (Fig. 6). It is expected that this
time lag should reach a median value of 2 d thanks to the
increasing performance of MUSCATE.
Each snow product is distributed as a zipped file which
contains raster and vector files. The file base name is the
product ID, i.e., where the productId
is a character string which is constructed as follows:
productId = Satellite_AcquisitionDate_
_L2B-SNOW_TileName_D_Version. Earth Syst. Sci. Data, 11, 493–514, 2019
498 S. Gascoin et al.: Theia Snow collection
Figure 3. Flow chart of the snow detection algorithm. Table 1 gives the description and value of the algorithm parameters (written in red in
this chart). MUSCATE is the scheduler which manages the L2A production for Theia.
For example, the eastern snow product in Fig. 5 was extracted
from a file named
which indicates that this product was generated from a
Sentinel-2A image acquired on 30 November 2015 at
10:56:41.486 UTC and covers tile 31TDH. The version of
the product is 1.0 (see above).
The size of each zipped product varies depending on the
complexity of the snow and cloud mask but generally ranges
between 10 and 100 Mb. After extracting the product, the fol-
lowing files are created:
Earth Syst. Sci. Data, 11, 493–514, 2019
S. Gascoin et al.: Theia Snow collection 499
Table 1. LIS algorithm parameters description and default values for Theia’s L2A Sentinel-2 (and Landsat-8) products.
Parameter Description Value
rfResize factor to produce the downsampled red band 12 (8)
rDMaximum value of the downsampled red-band reflectance used to define a dark cloud pixel 0.300
n1Minimum value of the NDSI for the pass 1 snow test 0.400
n2Minimum value of the NDSI for the pass 2 snow test 0.150
r1Minimum value of the red-band reflectance the pass 1 snow test 0.200
r2Minimum value of the red-band reflectance the pass 2 snow test 0.040
dzSize of elevation band in the DEM used to define zs100
fsMinimum snow fraction in an elevation band to define zs0.100
fctMinimum clear pixels fraction (snow and no snow) in an elevation band used to define zs0.100
ftMinimum snow fraction in the image to activate the pass 2 snow test 0.001
rBMinimum value of the red band reflectance to return a non-snow pixel to the cloud mask 0.100
Figure 4. Coverage of the Theia Snow collection tiles. Earth Syst. Sci. Data, 11, 493–514, 2019
500 S. Gascoin et al.: Theia Snow collection
Figure 5. Map of the first two products in the Theia Snow collection and comparison with MOD10A1.005 snow product of the same day.
The map background is the colourized version of the EU DEM.
Table 2. Spatial resolution and central wavelength of the spectral
bands used by LIS for each compatible sensor.
Sensor Green band Red band SWIR band
SPOT-4 HRV 20 m, 0.55 µm 20 m, 0.65 µm 20 m, 1.6 µm
SPOT-5 HRG 10 m, 0.55 µm 10 m, 0.65 µm 10 m, 1.6 µm
Sentinel-2 MSI 10 m, 0.56 µm 10 m, 0.66 µm 20 m, 1.6 µm
Landsat-8 OLI 30 m, 0.56 µm 30 m, 0.65 µm 30 m, 1.6 µm
productID_SNW_R2.tif is an 8 bit single-band
GeoTiff image which provides the following classifica-
tion for each pixel:
0: no snow,
100: snow,
205: cloud including cloud shadow,
254: no data.
is a vector file in the ESRI shapefile format,
which contains a vectorized version of the pro-
ductID_SNW_R2.tif using the polygon geometry.
productID_CMP_R2.tif is an 8 bit RGB GeoTiff
image which shows the outlines of the snow mask
(green lines) and cloud mask (magenta lines) on a
false-colour composite of the input L2A image (RGB
image made with SWIR, red and green bands). This
composition was chosen because RGB composites us-
ing the SWIR-band image are useful for discriminating
the snow cover from the snow-free areas and from the
clouds (Vidot et al., 2017). Hence, this image mainly
allows the user to visually check the consistency of the
snow and cloud mask.
productID_MTD_ALL.xml is a metadata file.
productID_QKL_ALL.jpg is a quick picture of the
productID_SNW_R2.tif image made using this
colour map (8 bit RGB code in parentheses): snow
in cyan (0,255,255), cloud in white (255,255,255), no
Earth Syst. Sci. Data, 11, 493–514, 2019
S. Gascoin et al.: Theia Snow collection 501
Table 3. Non-exhaustive list of mountain ranges covered by the Theia Snow collection and corresponding tiles. The tiles marked with an
asterisk were made available before the start of the operational production (see Sect. 3).
Mountain range Tiles
Alps 32TMS, 32TMR, 32TMT, 32TNS, 32TNR, 32TLQ, 32TLP, 32TLS, 32TLR, 32TQT,
32TQS, 33TUM, 31TGK, 32TNT, 31TGJ, 32TPR, 31TGL, 32TPT, 32TPS
Appenines 33TUH, 32TNQ, 33TUG, 33TVG, 33TWF, 32TPP, 33TWE, 32TPQ, 32TQP, 32TQN
Cantabrian 30TUN, 29TPH, 29TPG, 29TQH
Corsica 32TMM, 32TMN, 32TNM, 32TNM
Grant Range 11SPC
High Atlas 29RPQ, 29RNQ*, 29SQR, 29SPR, 30STA
Sistema Iberico 30TVM, 30TWK, 30TWM, 30TWL, 30TXK, 30TXL
Jura 31TGM, 32TLT
Lebanon T36SYD, T37SBU, T36SYC, T37SBT, T36SYB
Massif Central 31TDL,3 1TDK, 31TEK, 31TEL
Pyrenees 30TXN, 30TYM, 30TYN, 31TCG, 31TDG, 31TCH, 31TDH
Sierra de Cazorla 30SWH
Sierra Nevada 30SVG, 30SWG
Sistema Central 30TUL, 30TUK, 30TVL, 30TTK
Vosges 32TLT, 32ULU
These tiles can be viewed on an interactive map on this page:\T1\textbackslash#/sites (last access:
12 April 2019).
Figure 6. Cumulative probability of the time lags between the
Sentinel-2 acquisition and the distribution date of the correspond-
ing snow product, computed for all snow products with acquisition
date in April 2018.
snow in grey (119,119,119) and no data in black (0,0,0)
(see an example in Fig. 7).
DATA/productID_HIS_R2.txt (since product
version 1.4) is a text file indicating the cloud, snow and
no-snow fraction area by elevation bins (Sect. 2).
MASKS/productID_EXS_R2.tif (in product ver-
sion 1.0, it is located in the root of the zipped file) is an
8 bit GeoTiff image for expert evaluation, where each
bit indicates the value of an intermediate computation
mask (Sect. 2):
bit 1: snow (pass 1),
bit 2: snow (pass 2),
bit 3: clouds (pass 1),
bit 4: clouds (pass 2),
bit 5: clouds (initial all cloud).
For most applications, only files with the SNW suffix should
be useful.
The shapefile and raster images are referenced in WGS-
84 UTM system with the zone number given by the first two
digits of the tile name. All raster files have a 20 m resolution.
Note that no-snow pixels can be any surface, including water
surface. No data pixels are the pixels outside of the acquisi-
tion segment.
4 Evaluation
In this section we present an evaluation of the Theia Snow
collection. We first present the methods developed to do
this evaluation and then the results. The evaluation was
based only on Sentinel-2 products, because the production
for Landsat-8 had just started at the time of writing and thus
the Landsat-8 products represented a very minor fraction of
the Theia Snow collection. Earth Syst. Sci. Data, 11, 493–514, 2019
502 S. Gascoin et al.: Theia Snow collection
Figure 7. Example of a snow map (SNW file) and its corresponding colour composite (CMP file). In the snow map the snow is represented
in cyan, no snow in grey and cloud in white. In the RGB composite, the snow-covered areas are delineated in magenta, while the clouds are
delineated in green (Theia snow product of tile 32TNS on 27 May 2017).
4.1 Method
4.1.1 Comparison with in situ snow depth
We collected all available snow depth measurements within
tiles 31TGM, 31TGL, 31TGK, 32TLS, 32TLR, 32TLQ,
31TDH, 31TCH, 30TYN and 30TNX from the Météo-
France database between 1 September 2017 and 31 August
2018. These 10 tiles cover most of the French Alps and Pyre-
nees (Fig. 8). We obtained 120 snow depth time series with
at least one measurement. We gathered all available Sentinel-
2 snow products for these tiles over the same period, i.e.
a total of 1134 products. Then, we extracted the pixel val-
ues at the location of each snow measurement station for all
dates. When a station is located in two tiles we only kept the
data from the first tile. We selected the snow depth measure-
ments which were collected on the same day of a Theia snow
product and discarded the measurements corresponding to a
cloud detection in the Theia snow product. The snow depth
measurements were converted to snow presence and absence
using a threshold of SD0=0 m. This eventually allowed us
to compute a confusion matrix between a set of snow pres-
ence and absence data from in situ measurement and a set
of simultaneous snow presence and absence data from the
Theia Snow collection across the French Alps and Pyrenees.
However, previous studies comparing MODIS binary snow
products with ground measurements showed that a value of
0 m may not be optimal (Klein and Barnett, 2003; Gascoin
et al., 2015); therefore the sensitivity of the results to SD0
was tested by recomputing the confusion matrix for 1 cm in-
crements of SD0from 0 to 1 m.
4.1.2 Comparison with snow maps of higher spatial
We used SPOT-6 and SPOT-7 images with a resolution of
1.5 m in panchromatic and 6 m in multispectral (blue, green,
red, near-infrared) to evaluate the accuracy of the snow de-
tection in Theia snow products. For this comparison, the
LIS processor was run with its current operational config-
uration (i.e. the configuration used to routinely produce the
Theia snow products since December 2017 Table 1). SPOT-
6 and SPOT-7 sensors acquire images with a large radio-
metric depth coded in 16 bits, which enables us to identify
surface features in alpine regions with dark shaded slopes
and bright snow surfaces. We searched the catalogue of the
Kalideos database for available SPOT images that could
match a cloud-free (or nearly cloud-free) Theia snow map
over the French Alps region (, last ac-
cess: 12 April 2019). We identified six pairs of images in
2016 and 2017 with a maximum time lag of 6 d (Table 4,
Fig. 9). The SPOT images were obtained as orthorectified
Earth Syst. Sci. Data, 11, 493–514, 2019
S. Gascoin et al.: Theia Snow collection 503
Figure 8. Map showing the location of the stations in the réseau nivo-météorologique (Météo-France snow observation network) used for
this study (source: Météo-France) and the corresponding Sentinel-2 tiles.
products from the Kalideos database. The SPOT images were
used to generate reference snow maps by a pixel-based su-
pervised classification. For each SPOT image, we manually
drew about 15 polygons of homogeneous of snow and no-
snow surfaces using the panchromatic image. The few cloud
pixels in the SPOT images were also manually delineated
with a large buffer to restrict the classification to strictly
cloud-free areas. Colour composites made from the multi-
spectral images were also used as a visual support to help
the snow and cloud classification. These polygons were then
used to extract training samples from the SPOT products. The
samples were taken from both the panchromatic and the mul-
tispectral images. We tested a number of classification algo-
rithms by splitting the polygons in calibration and validation
data sets. The detail of this analysis is not shown here, but it
allowed us to conclude that the random forest classifier was
the best choice for our purpose, given its good accuracy and
its numerical efficiency (Bouchet, 2018).
In addition, we generated the snow maps from the same
Sentinel-2 level 1 products using ESAs Sen2Cor version 2.5
(Louis et al., 2016). The Sen2Cor output snow masks are ex-
pected to be nearly identical to the ones included in the dis-
tributed L2A products by ESA, since the same processor is
used. We have generated the L2A products ourselves with
Sen2Cor for this study, because the L2A products were not
yet available when we made this analysis (ESA started op-
erational production of Sentinel-2 data at level 2A in 2017).
The Theia and Sen2Cor snow maps were compared to the
reference snow maps from SPOT images using standard met-
Table 4. Pairs of Sentinel-2 and SPOT-6/7 products used for the
evaluation of the snow detection accuracy (see also Fig. 9).
Sentinel-2 product Reference product
Tile Date Sensor Date
31TGK 13 Aug 2016 SPOT-6 8 Aug 2016
32TLS 12 Oct 2016 SPOT-7 12 Oct 2016
31TGK 11 Dec 2016 SPOT-7 17 Dec 2016
31TGL 1 Dec 2016 SPOT-7 3 Dec 2016
31TGK 11 Mar 2017 SPOT-6 11 Mar 2017
31TGL 11 Mar 2017 SPOT-6 11 Mar 2017
rics derived from the confusion matrix (accuracy, F1 score,
kappa, false-detection rate and false-negative rate).
4.1.3 Evaluation metrics
From the confusion matrices of Sect. 4.1.1 and 4.1.2, we de-
rived the following metrics: accuracy (the proportion of the
total number of predictions that were correct), false-positive
rate (FPR, i.e. the proportion of no-snow pixels that were in-
correctly classified as snow), false-negative rate (FNR, i.e.
the proportion of snow pixels that were incorrectly classified
as no snow), F1 score (harmonic average of the precision and
recall) and kappa coefficient (Cohen, 1960). Earth Syst. Sci. Data, 11, 493–514, 2019
504 S. Gascoin et al.: Theia Snow collection
Table 5. Confusion matrix between the Theia snow products and in
situ snow depth data (SD).
Theia snow products
In situ snow depth no snow snow
SD =0 276 8
SD >0 76 1054
4.1.4 Visual verification
The evaluation methods presented above undersample the ac-
tual resolution of the Theia snow collection, both in space
(only 120 points of comparisons in Sect. 4.1.1) and time
(only six dates of comparisons in Sect. 4.1.2). Given that we
do not have a more extensive validation data set, we used a
time series of 64 Theia snow products from 6 July 2015 to
2 July 2017 over tile 31TCH to control the consistency of the
snow and cloud masks based on the visual inspection of the
false-colour composites. This approach is efficient for evalu-
ating the frequency of gross errors on the tile scale, i.e. large
patches of false snow or false no-snow detection (Vidot et al.,
4.2 Results
4.2.1 Comparison with in situ snow depth
The confusion matrix between the Theia snow products and
in situ snow depth measurements is given in Table 5. We find
1414 pairs of data for the study period (i.e. there are 1414
individual snow depth measurements for which a cloud-free
retrieval can be found in the Theia Snow collection on the
same day). From this confusion matrix, the accuracy (pro-
portion of correct classifications) is 94 % and the kappa co-
efficient is 0.83, which indicates excellent agreement accord-
ing to Fleiss et al. (2013). The false-positive rate (2.8 %) is
lower than the false-negative rate (6.7 %), which means that
the Theia snow product is more likely to underestimate than
to overestimate the snow detection at the station locations.
The false-negative rate decreases if SD0is set to a higher
value; however the false-positive rate also increases in such
a way that an optimum is reached at SD0=2 cm (Fig. 10).
The comparisons of each individual time series before
matching the data by date is shown in Appendix B. These
plots illustrate that the high revisit time of Sentinel-2 enables
us to capture the seasonal cycle of the snow cover well. Even
intermediate melt-out events at lower elevation stations can
be identified over the course of the snow season.
4.2.2 Comparison with snow maps of higher spatial
Table 6 shows that the MAJA–LIS workflow provides a bet-
ter detection of the snow cover than the Sen2Cor processor in
Figure 9. Location of the five SPOT-6/7 products used for the
evaluation of the snow detection accuracy and the corresponding
Sentinel-2 tiles (see also Table 4).
Figure 10. Sensitivity of the agreement between the in situ and
Theia snow product to SD0in metres (threshold to convert the mea-
sured snow depth to snow presence or absence). The inset shows a
close-up of the region of 0–0.2 m.
Earth Syst. Sci. Data, 11, 493–514, 2019
S. Gascoin et al.: Theia Snow collection 505
Table 6. Results of the validation of Theia (LIS 1.2.1) and ESA (Sen2Cor 2.5) snow products using SPOT-6/7 reference snow maps.
Sentinel-2 product Accuracy F1 Kappa FPR FNR
Tile Date Sen2Cor LIS Sen2Cor LIS Sen2Cor LIS Sen2Cor LIS Sen2Cor LIS
31TGK 11 Dec 2016 0.83 0.92 0.73 0.91 0.61 0.84 0.41 0.04 0.05 0.14
31TGL 11 Mar 2017 0.89 0.90 0.92 0.93 0.74 0.76 0.09 0.02 0.06 0.11
31TGL 1 Dec 2016 0.84 0.95 0.78 0.95 0.66 0.90 0.35 0.05 0.01 0.06
31TGK 11 Mar 2017 0.85 0.89 0.89 0.92 0.65 0.72 0.15 0.01 0.07 0.13
31TGK 13 Aug 2016 0.98 0.99 0.35 0.77 0.34 0.77 0.78 0.24 0.22 0.22
32TLS 12 Oct 2016 0.98 0.97 0.63 0.64 0.62 0.62 0.49 0.08 0.17 0.51
all the studied cases. Although the improvement in the accu-
racy coefficient is low in some cases (31TGL/2017-03-11,
31TGK/2016-08-13) or even slightly negative in one case
(32TLS/2016-10-12), the F1 score and kappa coefficient of
the Theia snow products are always greater than or equal to
those of Sen2Cor. The differences are significant for four
products among the six tested. The lower accuracy in the
case of the 32TLS/2016-10-12 product is due to a higher
false-negative rate. The false-positive rates in the Theia snow
products are generally lower than the false-negative rates,
which is consistent with the previous evaluation using in
situ data. The spatial analysis of the errors indicates that
the false-positives (i.e. overdetection of snow pixels) in the
Theia products are mostly located near the snow cover edges
(Fig. A1). This suggests that these errors might not be due
to a false detection of snow (e.g. confusion with a lake or a
cloud surface) but are rather due to the resolution discrepancy
with the reference data. By contrast, the spatial comparison
of the Sen2Cor products with the reference data shows large
patches of false negatives (omission of snow pixels) in the
snow-covered area, which are typically due to the omission
of snow pixels in shaded areas (Fig. A2).
4.2.3 Visual verification
We found 4 products among 64 with significant patches of
false-positive pixels (i.e. pixels falsely classified as snow).
These false snow areas are exclusively located in cloud areas
(Fig. 11). False-negative pixels (i.e. pixels falsely classified
as snow-free) can be found by looking at higher-resolution
data along the snow cover edges, but we did not find large
areas of false-negative pixels.
5 Discussion
The evaluation of the Theia Snow collection using both in
situ and remote-sensing reference data sets indicates an ex-
cellent accuracy of the snow detection, albeit with a tendency
to snow underdetection. The omission of snow pixels can be
due to the presence of trees or shadows. In the case of the
in situ comparison, part of the errors may be explained by
the uncertainty in the geolocalization of the Sentinel-2 data,
which is close to 11 m at 95 % for both satellites accord-
ing to the latest Sentinel-2 L1C data quality report (MPC
Team, 2018). In addition, in situ measurements may not be
representative of the snow conditions within the Sentinel-2
pixel. However, the sensitivity analysis of the accuracy and
kappa coefficient to the SD0value suggests that the discrep-
ancy between the scale of the in situ observations and the
scale of the pixel observation is not significant, since the op-
timal SD0is close to 0 m. In contrast, a similar analysis with
lower-resolution snow products from MODIS found an op-
timal SD0of 0.15 m Gascoin et al. (2015). This illustrates
how high-resolution snow products from Sentinel-2 enable
us to partly resolve the typical scale issue between in situ
and remote-sensing products (Blöschl, 1999).
The visual inspection of the products reveals that the main
issue in some Theia snow products is rather the occurrence
of false positives due to the confusion with cloud surfaces.
This problem can be due to two main factors:
Cold clouds: cold clouds have a spectral signature that
is close to the snow cover since they contain ice crys-
tals (high NDSI). If these clouds are not accurately de-
tected by the level 2A processor then the LIS algorithm
will also classify them as snow after pass 1 (Sect. 2).
As a result the snow line elevation zsis not well es-
timated, which can generate erroneous snow patches
within these clouds, even at low elevation (e.g. product
of 20 July 2017 in Fig. 11).
Shaded clouds: the three-dimensional structure of the
clouds can form shadows within the cloud cover area
(e.g. product of 4 March 2017 in Fig. 11). These shaded
cloud areas have a lower reflectance in the visible wave-
lengths and therefore can be considered as dark clouds
in the LIS algorithm, provided that the shaded cloud
cover area is large enough to significantly reduce the re-
flectance in the red band, even after the downsampling
step (Sect. 2).
Among 64 products, we find 4 cases with significant false
snow cover detection in this tile (Fig. 11). Although it is not
shown here, we have made this exercise over other tiles in the
Alps and Atlas mountains, and we estimate that this propor-
tion is similar in these areas. Regarding this issue we mainly Earth Syst. Sci. Data, 11, 493–514, 2019
506 S. Gascoin et al.: Theia Snow collection
Figure 11. Full time series of Theia snow products from 6 July 2015 to 2 July 2017 over tile 31TCH. Four products with visually evident
false snow detections are framed in red and shown at higher resolution. Red arrows indicate the location of the false snow patches, while the
green arrow indicates a true snow patch. The product of 16 April 2017 is shown as a reference to help the visualization of the errors.
rely on our visual judgment using colour composites to as-
sess the ability of the algorithm to discriminate the cloud
and the snow cover. However, to further quantify the algo-
rithm performance and eventually optimize its parameters,
we would need a database of classified imagery with labelled
regions of cloud and snow cover in many different cases.
Both Sen2Cor and LIS use the NDSI to detect the snow
cover; however the algorithms differ in many aspects. The
better performances of LIS in Sect. 4.2.2 can be related to the
mono-date approach for atmospheric correction and cloud
detection in Sen2Cor, which can cause snow–cloud confu-
sion in the L2A product, while LIS uses L2A products from
the more accurate multi-temporal MAJA processor (Baetens
et al., 2019). In addition, a unique feature of the LIS algo-
rithm is the use of the snow line elevation concept to im-
prove the robustness of the snow detection in mountain re-
gions. In addition, LIS was designed to retrieve snow below
thin clouds and to reclassify snow pixels which are frequently
found near the snow cover edges in L2A products. Apart
Earth Syst. Sci. Data, 11, 493–514, 2019
S. Gascoin et al.: Theia Snow collection 507
from the MAJA processing, however, the effect of these al-
gorithmic differences could be less evident in flat areas.
6 Code and data availability
The Theia snow collection can be accessed
and cited using this digital object identifier: (Gascoin
et al., 2018). The let-it-snow source code is available
under the GNU Affero General Public License v3.0 in this
let-it-snow/ (last access: 12 April 2019).
7 Conclusions
The Theia Snow collection is a free collection of snow prod-
ucts which indicate the presence or absence of snow on
the land based on Sentinel-2 (20 m) and Landsat-8 (30 m)
data. Most of the snow products are derived from Sentinel-
2. However, in late September 2018, Landsat-8 snow prod-
ucts started being routinely added to the Theia Snow col-
lection for the titles in France only. Previous Landsat-8 data
were also reprocessed back to March 2017. At the time of
writing (15 March 2019), for example in tile T31TCH in
the Pyrenees, 20.5 % of all available products were derived
from Landsat-8 data (63 among 244). In addition, the pro-
cessing is done to facilitate the combination of Landsat-8
and Sentinel-2 snow products, since the data format and the
tiling scheme are the same (only the pixel size differs). Al-
though the co-registration of Sentinel-2 and Landsat-8 data
can be still problematic (Storey et al., 2016), the combination
of “analysis ready” Landsat-8 and Sentinel-2 is important for
increasing the number of cloud-free observations.
The evaluation presented in this article indicate that the
Theia Snow product allows an accurate detection of the snow
cover (and of the absence of snow). However, it remains a
preliminary assessment that should be extended in the fu-
ture. For example, the evaluation was based only on Sentinel-
2 products, since the integration of Landsat-8 data started
during the writing of this paper and should be extended to
Landsat-8 products. However, the processing is done with
the same algorithm, which is based on the calculation of
the NDSI from slope-corrected surface reflectance (level 2A
data). In addition, the evaluation was limited to the French
Alps and French Pyrenees, while the collection covers other
mountain regions. Last, the spatial evaluation using higher-
resolution remote-sensing data was focused on the accuracy
of the snow detection using very high-resolution clear-sky
images, while, as discussed above, the main difficulty is not
to detect the snow when the sky is cloud-free but rather to
avoid false snow detection within clouds. The reduction of
the false-positive rate is the main challenge for the future de-
velopments in the LIS algorithm. In the meantime, we also
welcome feedback from other users.
The Theia Snow collection is based on optical observa-
tions; therefore it is not adapted to the detection of the snow
cover in dense forest areas where the ground is obstructed
by the canopy (Xin et al., 2012). This may typically occur in
evergreen conifer forests of alpine regions (e.g. Alps, Pyre-
nees). We have noted that the snow detection can be success-
ful even in alpine forests, but we lack data that can be used for
quantifying its accuracy. Therefore it is recommended to use
a land cover map to exclude these forest areas from the anal-
ysis. For the tiles in France, this can be done using the Theia
land cover map (Inglada et al., 2017), which is distributed at
the same spatial resolution. In the Pyrenees, the mean alti-
tude of zero-degree isotherm in winter is close to the treeline
elevation at 1600 m; therefore the impact of the forest cover
is limited (Gascoin et al., 2015). The LIS algorithm may also
fail to detect the snow cover on steep, shaded slopes if the
solar elevation is very low (typically below 20). This can
occur in midlatitude areas in winter. In this case, the slope
correction in the L2A product is generally not applied. This
is indicated in the L2A mask but not in the snow product. A
future release of the Theia Snow collection should include
this information. However, the visual inspection of many im-
ages in regions of complex topography gives us the impres-
sion that the snow detection algorithm already performs well
on shaded slopes. We are planning to investigate this issue
further using alternative approaches (Sirguey et al., 2009).
In spite of all these limitations, the Theia snow products
have already been successfully used for the evaluation of the
MODIS snow products (Masson et al., 2018) and the assim-
ilation of snow-covered area data into a snowpack model in
the High Atlas (Baba et al., 2018). Given that the snow cover
is a key driver of many natural processes in mountains re-
gions, we envision various potential applications of the Theia
snow products, including the modelling of the distribution of
the permafrost in mountain regions, the validation and cal-
ibration of hydrological models in snow-dominated catch-
ments and the spatial modelling of biodiversity and produc-
tivity of ecosystems in mountain regions. Earth Syst. Sci. Data, 11, 493–514, 2019
508 S. Gascoin et al.: Theia Snow collection
Appendix A: Spatial comparison between Theia and
Sen2Cor snow products with SPOT-6/7 reference data
Figure A1. Theia snow products vs. SPOT-6/7 snow maps (green: true negative, white: true positive, red: false positive, yellow: false
Figure A2. Sen2cor snow product vs. SPOT-6/7 snow maps (green: true negative, white: true positive, red: false positive, yellow: false
Earth Syst. Sci. Data, 11, 493–514, 2019
S. Gascoin et al.: Theia Snow collection 509
Appendix B: Time series of Theia snow products and
snow depth records at 120 snow observation stations
from 1 September 2017 to 31 August 2018
Figure B1. Magenta dots: snow presence (a) or absence (b) in the Theia snow collection. Black line: snow depth time series. Earth Syst. Sci. Data, 11, 493–514, 2019
510 S. Gascoin et al.: Theia Snow collection
Figure B2. Magenta dots: snow presence (a) or absence (b) in the Theia snow collection. Black line: snow depth time series.
Earth Syst. Sci. Data, 11, 493–514, 2019
S. Gascoin et al.: Theia Snow collection 511
Figure B3. Magenta dots: snow presence (a) or absence (b) in the Theia snow collection. Black line: snow depth time series. Earth Syst. Sci. Data, 11, 493–514, 2019
512 S. Gascoin et al.: Theia Snow collection
Figure B4. Magenta dots: snow presence (a) or absence (b) in the Theia snow collection. Black line: snow depth time series.
Earth Syst. Sci. Data, 11, 493–514, 2019
S. Gascoin et al.: Theia Snow collection 513
Author contributions. SG and OH defined the LIS algorithm.
MG and GS worked on the implementation and development of the
LIS processor. MG, MB and SG worked on the validation of the
snow products. SG wrote the manuscript with inputs from all the
Competing interests. The authors declare that they have no con-
flict of interest.
Acknowledgements. The development and production of the
Theia Snow collection is supported by the CNES. We thank Tris-
tan Klempka for his contribution to earlier developments of the LIS
processor. We thank Marc Leroy, Arnaud Sellé and Philippe Pachol-
czyk for the coordination of the Theia project and Joelle Donadieu
and Céline l’Helguen for the transfer to the production at MUS-
CATE. This manuscript was greatly improved thanks to the con-
structive comments of Gaia Piazzi, Javier Herrero, Jeff Dozier and
an anonymous referee.
Review statement. This paper was edited by Alexander
Kokhanovsky and reviewed by Gaia Piazzi, Javier Herrero, Jeff
Dozier, and one anonymous referee.
Baba, M. W., Gascoin, S., and Hanich, L.: Assimilation of Sentinel-
2 Data into a Snowpack Model in the High Atlas of Morocco,
Remote Sensing, 10, 1982,,
Baetens, L., Desjardins, C., and Hagolle, O.: Validation of Coper-
nicus Sentinel-2 Cloud Masks Obtained from MAJA, Sen2Cor,
and FMask Processors Using Reference Cloud Masks Generated
with a Supervised Active Learning Procedure, Remote Sensing,
11, 433,, 2019.
Blöschl, G.: Scaling issues in snow hydrology, Hydrol. Process., 13,
2149–2175, 1999.
Bouchet, M.: Validation et amélioration des pro-
duits Surfaces Enneigées Sentinel-2, Zenodo,, 2018.
Cohen, J.: A coefficient of agreement for nominal scales, Educ. Psy-
chol. Meas., 20, 37–46, 1960.
Commissariat général au développement durable: Plan
d’applications satellitaires 2018, available at: https:
20d%E2%80%99applications%20satellitaires%202018.pdf (last
access: 12 April 2019), 2018.
Dozier, J.: Spectral signature of alpine snow cover from the Landsat
Thematic Mapper, Remote Sens. Environ., 28, 9–22, 1989.
Drusch, M., Bello, U. D., Carlier, S., Colin, O., Fernandez,
V., Gascon, F., Hoersch, B., Isola, C., Laberinti, P., Marti-
mort, P., Meygret, A., Spoto, F., Sy, O., Marchese, F., and
Bargellini, P.: Sentinel-2: ESA’s Optical High-Resolution Mis-
sion for GMES Operational Services, Remote Sens. Environ.,
120, 25–36,, 2012.
Fleiss, J. L., Levin, B., and Paik, M. C.: Statistical methods for rates
and proportions, John Wiley & Sons, 2013.
Frei, A., Tedesco, M., Lee, S., Foster, J., Hall, D., Kelly,
R., and Robinson, D. A.: A review of global satellite-
derived snow products, Adv. Space Res., 50, 1007–1029,, 2012.
Gao, B.-C., Goetz, A. F., and Wiscombe, W. J.: Cirrus cloud detec-
tion from airborne imaging spectrometer data using the 1.38 µm
water vapor band, Geophys. Res. Lett., 20, 301–304, 1993.
Gascoin, S., Hagolle, O., Huc, M., Jarlan, L., Dejoux, J.-F.,
Szczypta, C., Marti, R., and Sánchez, R.: A snow cover climatol-
ogy for the Pyrenees from MODIS snow products, Hydrol. Earth
Syst. Sci., 19, 2337–2351,
2015, 2015.
Gascoin, S., Grizonnet, M., Klempka, T., and Salgues, G.: Algo-
rithm theoretical basis documentation for an operational snow
cover product from Sentinel-2 and Landsat-8 data (Let-it-snow),, 2018.
Gascoin, S., Grizonnet, M., Hagolle, O., and Salgues, G.:
Theia Snow collection, CNES for THEIA Land data center,, 2018.
GDAL/OGR contributors: GDAL/OGR Geospatial Data Abstrac-
tion software Library, Open Source Geospatial Foundation, http:
// (last access: 12 April 2019), 2018.
Grizonnet, M., Michel, J., Poughon, V., Inglada, J., Savinaud, M.,
and Cresson, R.: Orfeo ToolBox: open source processing of re-
mote sensing images, Open Geospatial Data, Software and Stan-
dards, 2, 15,, 2017.
Hagolle, O., Huc, M., Pascual, D. V., and Dedieu, G.: A multi-
temporal method for cloud detection, applied to FORMOSAT-2,
VENµS, LANDSAT and SENTINEL-2 images, Remote Sens.
Environ., 114, 1747–1755, 2010.
Hagolle, O., Huc, M., Desjardins, C., Auer, S., and Richter,
R.: MAJA Algorithm Theoretical Basis Document,, 2017.
Hall, D., Riggs, G., Salomonson, V., DiGirolamo, N., and Bayr, K.:
MODIS snow-cover products, Remote Sens. Environ., 83, 181–
194, 2002.
Inglada, J., Vincent, A., Arias, M., Tardy, B., Morin, D., and Rodes,
I.: Operational High Resolution Land Cover Map Production at
the Country Scale Using Satellite Image Time Series, Remote
Sensing, 9, 95,, 2017.
Jarvis, A., Guevara, E., Reuter, H. I., and Nelson, A. D.: Hole-filled
SRTM for the globe: version 4: data grid, CGIAR Consortium
for Spatial Information, Washington D.C., USA, 2008.
Klein, A. and Barnett, A.: Validation of daily MODIS snow
cover maps of the Upper Rio Grande River Basin for the
2000-2001 snow year, Remote Sens. Environ., 86, 162–176,, 2003.
cí, P., Holko, L., and Parajka, J.: Variability of snow line eleva-
tion, snow cover area and depletion in the main Slovak basins in
winters 2001–2014, J. Hydrol. Hydromech., 64, 12–22, 2016.
Li, J. and Roy, D. P.: A Global Analysis of Sentinel-2A,
Sentinel-2B and Landsat-8 Data Revisit Intervals and Impli-
cations for Terrestrial Monitoring, Remote Sensing, 9, 902,, 2017.
Louis, J., Debaecker, V., Pflug, B., Main-Korn, M., Bieniarz,
J., Mueller-Wilm, U., Cadau, E., and Gascon, F.: Sentinel-2
Sen2Cor: L2A Processor for Users, in: Proc. ’Living Planet Sym- Earth Syst. Sci. Data, 11, 493–514, 2019
514 S. Gascoin et al.: Theia Snow collection
posium 2016’, Prague, Czech Republic, 9–13 May 2016, ESA
SP-740, 9–13, 2016.
Malnes, E., Buanes, A., Nagler, T., Bippus, G., Gustafsson, D.,
Schiller, C., Metsämäki, S., Pulliainen, J., Luojus, K., Larsen,
H. E., Solberg, R., Diamandi, A., and Wiesmann, A.: User re-
quirements for the snow and land ice services – CryoLand, The
Cryosphere, 9, 1191–1202,
2015, 2015.
Masson, T., Dumont, M., Mura, M., Sirguey, P., Gas-
coin, S., Dedieu, J.-P., and Chanussot, J.: An Assess-
ment of Existing Methodologies to Retrieve Snow Cover
Fraction from MODIS Data, Remote Sensing, 10, 619,, 2018.
Metsämäki, S., Pulliainen, J., Salminen, M., Luojus, K., Wiesmann,
A., Solberg, R., Böttcher, K., Hiltunen, M., and Ripper, E.: Intro-
duction to GlobSnow Snow Extent products with considerations
for accuracy assessment, Remote Sens. Environ., 156, 96–108,
MPC Team: Sentinel-2 L1C Data Quality Report, European Space
Agency, Tech. Rep. 31, 80 pp., 2018.
Painter, T. H., Rittger, K., McKenzie, C., Slaughter, P., Davis, R. E.,
and Dozier, J.: Retrieval of subpixel snow covered area, grain
size, and albedo from MODIS, Remote Sens. Environ., 113, 868–
879, 2009.
Parajka, J. and Blöschl, G.: Spatio-temporal com-
bination of MODIS images–potential for snow
cover mapping, Water Resour. Res., 44, W03406,, 2008.
Ramsay, B. H.: The interactive multisensor snow and ice mapping
system, Hydrol. Process., 12, 1537–1546, 1998.
Sirguey, P., Mathieu, R., Arnaud, Y., Khan, M., and Chanussot, J.:
Improving MODIS Spatial Resolution for Snow Mapping Using
Wavelet Fusion and ARSIS Concept, IEEE Geosci. Remote S.,
5, 78–82,, 2008.
Sirguey, P., Mathieu, R., and Arnaud, Y.: Subpixel monitoring of the
seasonal snow cover with MODIS at 250m spatial resolution in
the Southern Alps of New Zealand: Methodology and accuracy
assessment, Remote Sens. Environ., 113, 160–181, 2009.
Storey, J., Roy, D. P., Masek, J., Gascon, F., Dwyer, J., and Choate,
M.: A note on the temporary misregistration of Landsat-8 Op-
erational Land Imager (OLI) and Sentinel-2 Multi Spectral In-
strument (MSI) imagery, Remote Sens. Environ., 186, 121–122,, 2016.
U.S. Geological Survey Earth Resources Observation And Science
Center: Landsat OLI Level-2 Surface Reflectance (SR) Science
Product,, 2014.
Vidot, J., Bellec, B., Dumont, M., and Brunel, P.: A daytime VIIRS
RGB pseudo composite for snow detection, Remote Sens. En-
viron., 196, 134–139,,
Xin, Q., Woodcock, C. E., Liu, J., Tan, B., Melloh, R. A.,
and Davis, R. E.: View angle effects on MODIS snow map-
ping in forests, Remote Sensing of Environment, 118, 50–59,, 2012.
Zhu, Z., Wang, S., and Woodcock, C. E.: Improvement and expan-
sion of the Fmask algorithm: Cloud, cloud shadow, and snow de-
tection for Landsats 4–7, 8, and Sentinel 2 images, Remote Sens.
Environ., 159, 269–277, 2015.
Earth Syst. Sci. Data, 11, 493–514, 2019
... Optical remote sensing is a measurement technique that is widely used to analyze snow-cover dynamics over large areas; see, e.g., [6][7][8][9]. In comparison to ground-based station measurements, satellite-based data can capture the spatial distribution of the snow extent. ...
... First, an L2A product is generated using the MAJA processor [43], which applies an atmospheric correction, a topographic normalization, and a conservative cloud-masking. Second, binary snow cover maps are generated with the let-it-snow (LIS) processor [9] using a combination of different NDSI thresholds and a digital elevation model. The LIS processor additionally refines the cloud mask by recovering snow pixels from the conservative cloud mask [9]. ...
... Second, binary snow cover maps are generated with the let-it-snow (LIS) processor [9] using a combination of different NDSI thresholds and a digital elevation model. The LIS processor additionally refines the cloud mask by recovering snow pixels from the conservative cloud mask [9]. FSC TOC is calculated from the pixels classified as snow and is based on an empirical relationship that was calibrated using high-resolution satellite data [11]. ...
Full-text available
Snow cover is of high relevance for the Earth’s climate system, and its variability plays a key role in alpine hydrology, ecology, and socioeconomic systems. Measurements obtained by optical satellite remote sensing are an essential source for quantifying snow cover variability from a local to global scale. However, the temporal resolution of such measurements is often affected by persistent cloud coverage, limiting the application of high resolution snow cover mapping. In this study, we derive the regional snow line elevation in an alpine catchment area using public webcams. We compare our results to the snow line information derived from the Moderate-Resolution Imaging Spectroradiometer (MODIS) and Sentinel-2 snow cover products and find our results to be in good agreement therewith. Between October 2017 and the end of June 2018, snow lines derived from webcams lie on average 55.8 m below and 33.7 m above MODIS snow lines using a normalized-difference snow index (NDSI) of 0.4 and 0.1, respectively, and are on average 53.1 m below snow lines derived from Sentinel-2. We further analyze the superior temporal resolution of webcam-based snow cover information and demonstrate its effectiveness in filling temporal gaps in satellite-based measurements caused by cloud cover. Our findings show the ability of webcam-based snow line elevation retrieval to complement and improve satellite-based measurements.
... Previous studies explore various established approaches to evaluate variable influence in estimating runoff [2,5,[8][9][10][11][12][13][14][15][16][17][18]; however, very few of these studies have addressed the dynamic nature of the influences from a statistical approach. Consequently, there is relatively little information available about the variability of variable influence on streamflow at a catchment scale, but it is also critical to understand the dynamics of parameters' influence on runoff when focusing on improving prediction accuracy. ...
... Theia Snow collection routinely generates some high-resolution maps where snow cover is accurately detected through Sentinel-2 and Landsat-8 observations. Theia Snow products have been successfully applied for the evaluation of MODIS snow products; Gascoin (2019) discussed its potential applications in permafrost distribution modeling, hydrologic modeling, and spatial modeling of ecosystems in mountain regions [17]. However, as an optical-based observation system, the Theia Snow collection can't always adequately define high spatial variability of snowpack properties, because optical sensors (e.g., MODIS) can't capture snow cover below the canopy. ...
Full-text available
Since the middle of the 20th century, the peak snowpack in the Upper Rio Grande (URG) basin of United States has been decreasing. Warming influences snowpack characteristics such as snow cover, snow depth, and Snow Water Equivalent (SWE), which can affect runoff quantity and timing in snowmelt runoff-dominated river systems of the URG basin. The purpose of this research is to investigate which variables are most important in predicting naturalized streamflow and to explore variables' relative importance for streamflow dynamics. We use long term remote sensing data for hydrologic analysis and deploy R algorithm for data processing and synthesizing. The data is analyzed on a monthly and baseflow/runoff basis for nineteen sub-watersheds in the URG. Variable importance and influence on naturalized streamflow is identified using linear standard regression with multi-model inference based on the second-order Akaike information criterion (AICc) coupled with the intercept only model. Five predictor variables: temperature, precipitation, soil moisture, sublimation, and SWE are identified in order of relative importance for streamflow prediction. The most influential variables for streamflow prediction vary temporally between baseflow and runoff conditions and spatially by watershed and mountain range. Despite the importance of temperature on streamflow, it is not consistently the most important factor in streamflow prediction across time and space. The dominance of precipitation over streamflow is more obvious during baseflow. The impact of precipitation, SWE, sublimation, and minimum temperature on streamflow is evident during the runoff season, but the results vary for different sub-watersheds. The association between sublimation and streamflow is positive in the runoff season, which may relate to temperature and requires further research. This research sheds light on the primary drivers and their spatial and temporal variability on streamflow generation. This work is critical for predicting how warming temperatures will impact water supplies serving society and ecosystems in a changing climate.
... We used remote-sensed snow cover data to assess the relationship between snow cover at hatching date and use of nest site. Snow cover was retrieved from a raster-based, binary snow cover map with a spatial resolution of 20 m (Gascoin et al. 2018(Gascoin et al. , 2019. This analysis was based on broods located during nest surveys in 2018-2020 for which we had sufficient data to calculate the hatching date (n nests 2018 = 12, n nests 2019 = 25, n nests 2020 = 102). ...
Full-text available
Timing and location of reproduction are central to reproductive success across taxa. Among birds, many species have evolved specific strategies to cope with environmental variability including shifts in timing of reproduction to track resource availability or selecting suitable nest location. In mountain ecosystems, complex topography and pronounced seasonality result in particularly high spatiotemporal variability of environmental conditions. Moreover, the risk of climate‐induced resource mismatches is particularly acute in mountain regions given that temperature is increasing more rapidly than in the lowlands. We investigated how a high‐elevation passerine, the white‐winged snowfinch Montifringilla nivalis, selects its nest site in relation to nest cavity characteristics, habitat composition and snow condition. We used a combination of field habitat mapping and satellite remote sensing to compare occupied nest sites with randomly selected pseudo‐absence sites. In the first half of the breeding season, snowfinches preferred nest cavities oriented towards the morning sun while they used cavities proportional to their availability later on. This preference might relate to the nest microclimate offering eco‐physiological advantages, namely thermoregulatory benefits for incubating adults and nestlings under the harsh conditions typically encountered in the alpine environment. Nest sites were consistently located in areas with greater‐than‐average snow cover at hatching date, likely mirroring the foraging preferences for tipulid larvae developing in meltwater along snowfields. Due to the particularly rapid climate shifts typical of mountain ecosystems, spatiotemporal mismatches between foraging grounds and nest sites are expected in the future. This may negatively influence demographic trajectories of the white‐winged snowfinch. The installation of well‐designed nest boxes in optimal habitat configurations could to some extent help mitigate this risk.
... We performed this analysis with level-2A surface reflectance products processed with the MAJA atmospheric correction and cloud screening software (Hagolle et al., 2015(Hagolle et al., , 2017. We used the snow maps derived from the same Sentinel-2 images to exclude from the analysis the pixels that were not covered by snow in both images (data available from Gascoin et al., 2019). ...
Full-text available
Saharan dust outbreaks have profound effects on ecosystems, climate, human health and the cryosphere in Europe. However, the spatial deposition pattern of Saharan dust is poorly known due to a sparse network of ground measurements. Following the extreme dust deposition event of February 2021 across Europe, a citizen science campaign was launched to sample dust on snow over the Pyrenees and the European Alps. This somewhat improvised campaign triggered wide interest since 152 samples were collected in the snow in the Pyrenees, the French Alps and the Swiss Alps in less than four weeks. An analysis of the samples showed a large variability in the dust properties and amount. We found a decrease in the deposited mass and particle sizes with distance from the source along the transport path. This spatial trend was also evident in the elemental composition of the dust as the iron mass fraction decreased from 11 % in the Pyrenees to 2 % in the Swiss Alps. At the local scale, we found a higher dust mass on south facing slopes, in agreement with estimates from high-resolution remote sensing data. This unique dataset, which resulted from the collaboration of several research laboratories and citizens, is provided as an open dataset to benefit a large community and enable further scientific investigations.
... The images were processed following the workflow described in Richiardi et al. [67]: first extracting the cloud cover with the Fmask (function of mask) processor [68], then, performing the atmospheric and topographic correction with Sen2Cor [69]. Finally, the Snow Cover Extent (SCE) is extracted through the revised-Let It Snow (rLIS) algorithm, which exploits the approach firstly proposed by Gascoin et al. [32] and then modified by Richiardi et al. [67]. rLIS, which is specifically aimed at improving the discrimination between snow and clouds in mountain areas, is a tree algorithm based on thresholds on the normalized difference snow index (NDSI) and the red and the shortwaveinfrared (SWIR) reflectance, and on a Digital Elevation Model (DEM). ...
Full-text available
In the Alpine environment, snow plays a key role in many processes involving ecosystems, biogeochemical cycles, and human wellbeing. Due to the inaccessibility of mountain areas and the high spatial and temporal heterogeneity of the snowpack, satellite spatio-temporal data without gaps offer a unique opportunity to monitor snow on a fine scale. In this study, we present a random forest approach within three different workflows to combine MODIS and Sentinel-2 snow products to retrieve daily gap-free snow cover maps at 20 m resolution. The three workflows differ in terms of the type of ingested snow products and, consequently, in the type of random forest used. The required inputs are the MODIS/Terra Snow Cover Daily L3 Global dataset at 500 m and the Sentinel-2 snow dataset at 20 m, automatically retrieved through the recently developed revised-Let It Snow workflow, from which the selected inputs are, alternatively, the Snow Cover Extent (SCE) map or the Normalized Difference Snow Index (NDSI) map, and a Digital Elevation Model (DEM) of consistent resolution with Sentinel-2 imagery. The algorithm is based on two steps, the first to fill the gaps of the MODIS snow dataset and the second to downscale the data and obtain the high resolution daily snow time series. The workflow is applied to a case study in Gran Paradiso National Park. The proposed study represents a first attempt to use the revised-Let It Snow with the purpose of extracting temporal parameters of snow. The validation was achieved by comparison with both an independent dataset of Sentinel-2 to assess the spatial accuracy, including the snowline elevation prediction, and the algorithm’s performance through the different topographic conditions, and with in-situ data collected by meteorological stations, to assess temporal accuracy, with a focus on seasonal snow phenology parameters. Results show that all of the approaches provide robust time series (overall accuracies of A1 = 93.4%, and A2 and A3 = 92.6% against Sentinel-2, and A1 = 93.1%, A2 = 93.7%, and A3 = 93.6% against weather stations), but the first approach requires about one fifth of the computational resources needed for the other two. The proposed workflow is fully automatic and requires input data that are readily and globally available, and promises to be easily reproducible in other study areas to obtain high-resolution daily time series, which is crucial for understanding snow-driven processes at a fine scale, such as vegetation dynamics after snowmelt.
... Yet, estimating SWE spatial distribution is important to make accurate predictions of snowmelt runoff in alpine catchments. Satellite remote sensing provides spatial information about snow-related variables including (i) snow cover spatial extent at various spatiotemporal scales (Aalstad et al., 2020;Gascoin et al., 2019;Hall et al., 2002;Hüsler et al., 2014), (ii) snow depth (Lievens et al., 2019;Marti et al., 2016), (iii) albedo (Kokhanovsky et al., 2020), and (iv) land surface temperature (Bhardwaj et al., 2017). However, the direct estimation of key variables such as the snow water equivalent (SWE) or density by means of remote sensing techniques remains challenging (Dozier et al., 2016). ...
Full-text available
Accurate knowledge of the seasonal snow distribution is vital in several domains including ecology, water resources management, and tourism. Current spaceborne sensors provide a useful but incomplete description of the snow-pack. Many studies suggest that the assimilation of remotely sensed products in physically based snowpack models is a promising path forward to estimate the spatial distribution of snow water equivalent (SWE). However, to date there is no standalone, open-source, community-driven project dedicated to snow data assimilation, which makes it difficult to compare existing algorithms and fragments development efforts. Here we introduce a new data assimilation toolbox, the Multiple Snow Data Assimilation System (MuSA), to help fill this gap. MuSA was developed to fuse remotely sensed information that is available at different timescales with the energy and mass balance Flexible Snow Model (FSM2). MuSA was designed to be user-friendly and scal-able. It enables assimilation of different state variables such as the snow depth, SWE, snow surface temperature, binary or fractional snow-covered area, and snow albedo and could be easily upgraded to assimilate other variables such as liquid water content or snow density in the future. MuSA allows the joint assimilation of an arbitrary number of these variables , through the generation of an ensemble of FSM2 simulations. The characteristics of the ensemble (i.e., the number of particles and their prior covariance) may be controlled by the user, and it is generated by perturbing the meteorological forcing of FSM2. The observational variables may be assimilated using different algorithms including particle filters and smoothers as well as ensemble Kalman filters and smoothers along with their iterative variants. We demonstrate the wide capabilities of MuSA through two snow data assimilation experiments. First, 5 m resolution snow depth maps derived from drone surveys are assimilated in a distributed fashion in the Izas catchment (central Pyrenees). Furthermore, we conducted a joint-assimilation experiment, fusing MODIS land surface temperature and fractional snow-covered area with FSM2 in a single-cell experiment. In light of these experiments , we discuss the pros and cons of the assimilation algorithms , including their computational cost.
... Finally, we implemented two random forest classification analyses to assess relationships between anomalies of TL0.075 and GR0.075 and predictors (Breiman, 2001). We classified the two indicators into three categories: positive anomalies Snow-free growing degree days (SF-GDD) maps were calculated through the combined use of the Snow-melt out date (SMOD) product at 20-m resolution derived from Sentinel-2 time series analysis (Gascoin et al., 2019, Barrou Dumont et al., 2021, and the SAFRAN-CROCUS climatological model from Météo-France (Vernay et al., 2022), which provides the average daily temperature for each French massif and for 300 m elevation bands (for Swiss and Italian sites, we applied data from the French Mont-Blanc and Vanoise massifs, respectively). To obtain the SF-GDD for each plot, we computed the 260 cumulative sum of daily average air temperature above 0°C between snow melt-out date and August 1 for the year 2019. ...
Full-text available
Glacier forefields have long provided ecologists with a model to study patterns of plant succession following glacier retreat. While plant survey-based approaches applied along chronosequences provide invaluable information on plant communities, the “space-for-time” approach assumes environmental uniformity and equal ecological potential across sites and does not account for spatial variability in initial site conditions. Remote sensing provides a promising avenue for assessing plant colonisation dynamics using a so-called “real-time” approach. Here, we combined 36 years of Landsat imagery with extensive field sampling along chronosequences of deglaciation for eight glacier forefields in the south-western European Alps to investigate the heterogeneity of early plant succession dynamics. Based on the two complementary and independent approaches, we found strong variability in the time lag between deglaciation and colonisation by plants and in subsequent growth rates, and in the composition of early plant succession. All three parameters were highly dependent on the local environmental context, i.e., local vegetation surrounding the forefields and energy availability linked to temperature and snowmelt gradients. Potential geomorphological disturbance did not emerge as a strong predictor of succession parameters, perhaps due to insufficient spatial resolution of predictor variables. Notably, elapsed time since deglaciation showed no consistent relationship to plant assemblages, i.e., we did not identify a consistent order of successional species across forefields as a function of time. Overall, both approaches converged towards the conclusion that early plant succession is not stochastic as previous authors have suggested but rather deterministic. We discuss the importance of scale in deciphering the unique complexity of plant succession in glacier forefields and provide recommendations for improving botanical field surveys and using Landsat time series in glacier forefields systems. Our work demonstrates complementarity between remote sensing and field-based approaches for both understanding and predicting future patterns of plant succession in glacier forefields.
... In recent years, these characteristics of S2 have enabled high spatial resolution monitoring of snow cover in mountain areas [34][35][36]. However, long-term studies have been limited due to the late availability of S2 images. ...
Full-text available
Mapping seasonal snow cover dynamics provides essential information to predict snowmelt during spring and early summer. Such information is vital for water supply management and regulation by national stakeholders. Recent advances in remote sensing have made it possible to reliably estimate and quantify the spatial and temporal variability of snow cover at different scales. However, because of technological constraints, there is a compromise between the temporal, spectral, and spatial resolutions of available satellites. In addition, atmospheric conditions and cloud contamination may increase the number of missing satellite observations. Therefore, data from a single satellite is insufficient to accurately capture snow dynamics, especially in semi-arid areas where snowfall is extremely variable in both time and space. Considering these limitations, the combined use of the next generation of multispectral sensor data from the Landsat-8 (L8) and Sentinel-2 (S2), with a spatial resolution ranging from 10 to 30 m, provides unprecedented opportunities to enhance snow cover mapping. Hence, the purpose of this study is to examine the effectiveness of the combined use of optical sensors through image fusion techniques for capturing snow dynamics and producing detailed and dense normalized difference snow index (NDSI) time series within a semi-arid context. Three different models include the enhanced spatial and temporal adaptive reflectance fusion model (ESTARFM), the flexible spatio-temporal data fusion model (FSDAF), and the pre-classification flexible spatio-temporal data fusion model (pre-classification FSDAF) were tested and compared to merge L8 and S2 data. The results showed that the pre-classification FSDAF model generates the most accurate precise fused NDSI images and retains spatial detail compared to the other models, with the root mean square error (RMSE = 0.12) and the correlation coefficient (R = 0.96). Our results reveal that, the pre-classification FSDAF model provides a high-resolution merged snow time series and can compensate the lack of ground-based snow cover data.
The management and conservation of Pyrenean high mountain lakes within the current context of climate change and increasing anthropogenic pressure require detailed knowledge of their biogeochemical functioning. In this doctoral thesis, five sampling campaigns were carried out (2017-2019) in more than 20 alpine lakes. The analysis of water samples allowed us to study the occurrence, the depth profiles, the geographical distribution and the seasonal trends of a large array of physico-chemical and biogeochemical parameters. Specifically, the cycle of carbon dioxide (CO2) and the fate of Potentially Harmful Trace Elements (PHTEs) were investigated. The mercury (Hg) was specially studied through the development of an analytical procedure for the measurement of trace concentrations in natural waters and through biogeochemical investigations on the distribution and the fate of Hg species in the water column, as well as in sediment archives.The new and robust procedure developed in this work to measure the total alkalinity (TA) and the dissolved inorganic carbon (DIC) allowed us to determine the other two parameters of the CO2 system, the pH and the fugacity of CO2 (fCO2). The bedrock characteristics of the watershed appear to be the most important parameters influencing the acid status of the studied lakes. Moreover, obtained fCO2 values indicate that lakes are sources of CO2 for the atmosphere.The measurement of various physico-chemical parameters allowed us to discriminate and classify the studied lakes according to their water geochemistry, highlighting the importance of the trophic status of the lakes, the geological background and the atmospheric inputs. The occurrence, sources and behaviour of the PHTEs were investigated with evidence of a contrast between geological and atmospheric inputs. Intensive monitoring revealed some PHTEs to be highly sensitive to environmental changes such as temperature and redox conditions.Monitoring natural concentrations of total Hg in aquatic systems remains a difficult challenge and there is a need for the development of low cost and easy handling analytical methods. The method for analysis of trace Hg concentrations developed and optimized in this work was successfully operational and exhibits a suitable limit of detection and an excellent reproducibility. Hg speciation results in the water column demonstrated the pristine state and the dynamic of the Pyrenean lakes. The homogeneity in the non-gaseous total Hg concentrations in the studied lakes confirmed the absence of local sources and the potential use of these ecosystems as sentinels of regional to global Hg contamination. While inorganic mercury (iHg) did not show seasonal variations, monomethylmercury(MMHg) was significantly higher in autumn 2018 and dissolved gaseous mercury (DGM) varied strongly within and among lakes. In-situ experiments confirmed the conditions that promote Hg methylation (stratified anoxic waters), demethylation and photoreduction (intense UV light).The historical Hg record in sediment archives highlighted temporal trends in Hg accumulation rates (HgARs) with a progressive increase since the 16th Century and the industrialization, mirroring the Hg production in Almadén mines (Southern Spain). Besides, Hg stable isotopes allow the identification of distinct anthropogenic sources as well as past climate variability.Overall, environmental changes in lake ecosystems, induced by either climatic conditions (temperature, light intensity) or anthropogenic pressure (atmospheric inputs, eutrophication, atmospheric CO2) are likely to produce significant impacts among CO2, specific PHTEs and Hg biogeochemical cycles in mountainous ecosystems.
Considering the tradeoff between spatial resolution and temporal resolution, spatiotemporal fusion has become a promising technique to monitor snow cover dynamics with both high spatial and temporal resolutions. The representative spatiotemporal fusion methods, e.g. Spatial Temporal Data Fusion Approach (STDFA), usually exist obvious phenomenon of spectral distortion when the surface reflectance changes nonlinearly, which affects the quality of the spatiotemporal fusion image. To address this issue, an effective STDFA-Matching-Pix2pix-Generative Adversarial Network (SMPG) algorithm combining the unmixing-based method, deep learning method, pre-matching and post-matching module is proposed to reduce the spectral distortion of STDFA fusion image. The high-temporal-low-spatial (HTLS) resolution MOD09GA data and high-spatial-low-temporal resolution (HSLT) Landsat 8 data are selected in this study. SMPG algorithm is firstly employed to obtain daily high-spatial-high-temporal (HSHT) images, and then daily snow cover results with a spatial resolution of 30 m are obtained by calculating the normalized difference snow index (NDSI). SMPG algorithm is further compared with STDFA, Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM), Flexible Spatiotemporal DAta Fusion (FSDAF), Swin SpatioTemporal Fusion Model (SwinSTFM), and Generative Adversarial Network-based SpatioTemporal Fusion Model (GAN-STFM). The experimental results indicate that the proposed algorithm yields better overall performance in daily spatiotemporal fusion image and snow cover result with a spatial resolution of 30 m. The mean correlation coefficient (CC) of SMPG can achieve 0.962, which is 0.06-0.36 higher than that of other spatiotemporal fusion methods. The error between the percentage of snow cover area obtained through SMPG and validation data is within 0.84%.
Full-text available
The Sentinel-2 satellite mission, developed by the European Space Agency (ESA) for the Copernicus program of the European Union, provides repetitive multi-spectral observations of all Earth land surfaces at a high resolution. The Level 2A product is a basic product requested by many Sentinel-2 users: it provides surface reflectance after atmospheric correction, with a cloud and cloud shadow mask. The cloud/shadow mask is a key element to enable an automatic processing of Sentinel-2 data, and therefore, its performances must be accurately validated. To validate the Sentinel-2 operational Level 2A cloud mask, a software program named Active Learning Cloud Detection (ALCD) was developed, to produce reference cloud masks. Active learning methods allow reducing the number of necessary training samples by iteratively selecting them where the confidence of the classifier is low in the previous iterations. The ALCD method was designed to minimize human operator time thanks to a manually-supervised active learning method. The trained classifier uses a combination of spectral and multi-temporal information as input features and produces fully-classified images. The ALCD method was validated using visual criteria, consistency checks, and compared to another manually-generated cloud masks, with an overall accuracy above 98%. ALCD was used to create 32 reference cloud masks, on 10 different sites, with different seasons and cloud cover types. These masks were used to validate the cloud and shadow masks produced by three Sentinel-2 Level 2A processors: MAJA, used by the French Space Agency (CNES) to deliver Level 2A products, Sen2Cor, used by the European Space Agency (ESA), and FMask, used by the United States Geological Survey (USGS). The results show that MAJA and FMask perform similarly, with an overall accuracy around 90% (91% for MAJA, 90% for FMask), while Sen2Cor’s overall accuracy is 84%. The reference cloud masks, as well as the ALCD software used to generate them are made available to the Sentinel-2 user community.
Full-text available
The snow melt from the High Atlas is a critical water resource in Morocco. In spite of its importance, monitoring the spatio-temporal evolution of key snow cover properties like the snow water equivalent remains challenging due to the lack of in situ measurements at high elevation. Since 2015, the Sentinel-2 mission provides high spatial resolution images with a 5 day revisit time, which offers new opportunities to characterize snow cover distribution in mountain regions. Here we present a new data assimilation scheme to estimate the state of the snowpack without in situ data. The model was forced using MERRA-2 data and a particle filter was developed to dynamically reduce the biases in temperature and precipitation using Sentinel-2 observations of the snow cover area. The assimilation scheme was implemented using SnowModel, a distributed energy-balance snowpack model and tested in a pilot catchment in the High Atlas. The study period covers 2015-2016 snow season which corresponds to the first operational year of Sentinel-2A, therefore the full revisit capacity was not yet achieved. Yet, we show that the data assimilation led to a better agreement with independent observations of the snow height at an automatic weather station and the snow cover extent from MODIS. The performance of the data assimilation scheme should benefit from the continuous improvements of MERRA-2 reanalysis and the full revisit capacity of Sentinel-2.
Full-text available
The characterization of snow extent is critical for a wide range of applications. Since 1966, snow maps at different spatial resolutions have been produced using various satellite sensor images. Nowadays, the most widely used products are likely those derived from Moderate-Resolution Imaging Spectroradiometer (MODIS) data, which cover the whole Earth at a near-daily frequency. There are a variety of snow mapping methods for MODIS data, based on different methodologies and applied at different spatial resolutions. Up to now, all these products have been tested and evaluated separately. This study aims to compare the methods currently available for retrieving snow from MODIS data. The focus is on fractional snow cover, which represents the snow cover area at the subpixel level. We examine the two main approaches available for generating such products from MODIS data; namely, linear regression of the Normalized Difference Snow Index (NDSI) and spectral unmixing (SU). These two approaches have resulted in several methods, such as MOD10A1 (the NSIDC MODIS snow product) for NDSI regression, and MODImLAB for SU. The assessment of these approaches was carried out using higher resolution binary snow maps (i.e., showing the presence or absence of snow) at spatial resolutions of 10, 20, and 30 m, produced by SPOT 4, SPOT 5, and LANDSAT-8, respectively. Three areas were selected in order to provide landscape diversity: the French Alps (117 dates), the Pyrenees (30 dates), and the Moroccan Atlas (24 dates). This study investigates the impact of reference maps on accuracy assessments, and it is suggested that NDSI-based high spatial resolution reference maps advantage NDSI medium-resolution snow maps. For MODIS snow maps, the results show that applying an NDSI approach to accurate surface reflectance corrected for topographic and atmospheric effects generally outperforms other methods for the global retrieval of snow cover area. The improvements to the newer version of MOD10A1 (Collection 6) compared to the older version (Collection 5) are significant. Products based on SU provide a good alternative and more accurate retrieval of the snow fraction where wider ranges of land covers are concerned. The fusion process and its resulting 250 m spatial resolution product improve snow line retrieval. False detection in mixed pixels, probably due to the spectral variability associated with the various materials in the spectral mixture, has been identified as an area that will require improvement.
Full-text available
Combination of different satellite data will provide increased opportunities for more frequent cloud-free surface observations due to variable cloud cover at the different satellite overpass times and dates. Satellite data from the polar-orbiting Landsat-8 (launched 2013), Sentinel-2A (launched 2015) and Sentinel-2B (launched 2017) sensors offer 10 m to 30 m multi-spectral global coverage. Together, they advance the virtual constellation paradigm for mid-resolution land imaging. In this study, a global analysis of Landsat-8, Sentinel-2A and Sentinel-2B metadata obtained from the committee on Earth Observation Satellite (CEOS) Visualization Environment (COVE) tool for 2016 is presented. A global equal area projection grid defined every 0.05° is used considering each sensor and combined together. Histograms, maps and global summary statistics of the temporal revisit intervals (minimum, mean, and maximum) and the number of observations are reported. The temporal observation frequency improvements afforded by sensor combination are shown to be significant. In particular, considering Landsat-8, Sentinel-2A, and Sentinel-2B together will provide a global median average revisit interval of 2.9 days, and, over a year, a global median minimum revisit interval of 14 min (±1 min) and maximum revisit interval of 7.0 days.
Full-text available
Orfeo ToolBox is an open-source project for state-of-the-art remote sensing, including a fast image viewer, applications callable from command-line, Python or QGIS, and a powerful C++ API. This article is an introduction to the Orfeo ToolBox’s flagship features from the point of view of the two communities it brings together: remote sensing and software engineering.
Full-text available
RGB pseudo composite imagery is a robust and fast processing technique for satellite observations to provide near real-time synthesized weather information to forecasters. The technique consists of superposing three images resulting from a linear combination of multi-spectral observations and corresponding them to red, green and blue. A new algorithm for the detection of snow during daytime has recently been developed for the Suomi-NPP VIIRS instruments. The algorithm uses five VIIRS bands including that centered at 2.25 μm to enhance snow detection and that centered at 1.24 μm to highlight the temporal variability of the snow cover. Furthermore, this paper shows the capability of the algorithm to detect snow under thin cloud conditions. This product is already provided in near real-time to forecasters of the French national weather service, Météo-France.
Full-text available
A detailed and accurate knowledge of land cover is crucial for many scientific and operational applications, and as such, it has been identified as an Essential Climate Variable. This accurate knowledge needs frequent updates. This paper presents a methodology for the fully automatic production of land cover maps at country scale using high resolution optical image time series which is based on supervised classification and uses existing databases as reference data for training and validation. The originality of the approach resides in the use of all available image data, a simple pre-processing step leading to a homogeneous set of acquisition dates over the whole area and the use of a supervised classifier which is robust to errors in the reference data. The produced maps have a kappa coefficient of 0.86 with 17 land cover classes. The processing is efficient, allowing a fast delivery of the maps after the acquisition of the image data, does not need expensive field surveys for model calibration and validation, nor human operators for decision making, and uses open and freely available imagery. The land cover maps are provided with a confidence map which gives information at the pixel level about the expected quality of the result.
Technical Report
This document provides the theoretical basis that hides behind the modules of MAJA processor. MAJA stands from MACCS-ATCOR Joint Algorithm, where MACCS was a Multi-Temporal Atmospheric Correction and Cloud Screening software, developed by CNES and CESBIO, and ATCOR is an Atmospheric Correction software developed by DLR. MAJA is based on MACCS architecture and includes modules that come from ATCOR. Link to document:
Conference Paper
Sen2Cor is a Level-2A (L2A) processor which main purpose is to correct single-date Sentinel-2 Level-1C products from the effects of the atmosphere in order to deliver a Level-2A surface reflectance product. This paper provides a description of L2A products contents and format. It presents also the different ways to run and configure the Sen2Cor processor and provides up-to-date information about the Sen2Cor release status and early validation results at the time of the Living Planet Symposium 2016.