1. Contribution of high spatial resolution remote sensing data to the modeling of snow water equivalent in the Atlas Mountains
Wassim Mohamed Baba (1), Simon Gascoin (1), Lahoucine Hanich (2), Christophe Kinnard (3)
(1) Centre d’etudes spatiales de la biosphere (CESBIO) . Toulouse . France (2) Université Cadi Ayyad, Marrakech, Morocco (3) Université du Québec à Trois-Rivières
Hello,
I’am Mohamed Wassim BABA. I’m doing my Phd in CESIO at Toulouse in France funded by CNES and Midi-Pyrénées region. My researches aim to identify the contribution of high spatial resolution remote sensing data to the modeling of snow water equivalent in the Atlas Mountains.

2. Melt is a key variable that recharges oueds in Atlas. © Hertz.ma
Introduction
The importance of snowpacks in Morocco: Providing water during melt season for crop irrigation.
Key variables : Snow water equivalent (SWE), Melt.
Limitations : > Satellites do not provide direct measurement of SWE > The heterogeneity of distribution of SWE.
Snowmelt from the Mountains watersheds represent an important water resource for the semi arid cultivated low lands by providing water for crop irrigation.
Hence the need to monitor the key variables of snowpacks with a high accuracy .
As today, there is no satellite to measure SWE with accuracy but we they give SCA observations at high spatial resolution. In the other hand models simulate SWE … but to our knowledge many researches used them at high spatial resolution.
 We need to assimilate SCA to a snowpack model to assimilate SWE
Melt is a key variable that recharges oueds in Atlas. © Hertz.ma

3. Introduction Objectives:
Seeking the optimal resolution which ensure all at once the accuracy of simulations and consume reasonable time ?
Modeling SWE and assimilate it over semi-arid mountain (The high-Atlas in Morocco) .
Hence we need to seek the optimal resolution which will ensure all at once the accuracy of simulations and can be executed at reasonable time. This will make easier modeling SWE and assimilate it over the Atlas-Mountain in Morocco in the next studies.
Photo of Oukaimeden (Atlas mountains ). December 2016

4. The effect of spatial resolution
8 m
30 m
90 m
If we consider a mountainous area and represent it with 5 different DEMs ( 8m, 30m … and 500 m) we can observe that decreasing spatial resolution lead to smooth peaks and affect both slopes and orientations. This suggest that modeling snowpack will be influenced by the spatial resolution with which we simulate.
250 m
500 m

5. The effect of spatial resolution
If we consider a mountainous area and represent it with 5 different DEMs ( 8m, 30m … and 500 m) we can observe that decreasing spatial resolution lead to smooth peaks and affect both slopes and orientations. This suggest that modeling snowpack will be influenced by the spatial resolution with which we simulate.
The distribution of % aspects orientation at different scale
Vertical cross section : Elevations from different DEMs

6. Plan : Study area : High Atlas, Morocco Data : Pléiades & Formosat
Model : SnowModel.
Validation method
Results : What is the optimal resolution ?
To reply to theses questions we will address it in the Atlas area. Then, I will present the different high resolution data used to feed SnowModel. Afterwards, I will present the validation method.
adapt the following scheme :
Firstly, I will describe briefly the study area. Then, I will highlight some characteristics of the model and data. Afterwards I will present methods and results to finally close with a conclusion.

7. Study area The Rheraya basin. 225 km². Elevations : 1000 m to 4167 m.
5 AWS stations.
1 Snow measurement routine.
Marrakech
Our study area is located in Atlas-Mountains near Marrakech. We focus on the experimental watershed of the Rheraya river.
With 225 km² as a surface area and elevations varying from 1000 m to 4167 m ( at Toubkal highest peak in North Africa ). This catchement is mainly chosen because of the presence of 5 Automatic Weather Stations which provide meteorological forcing to simulate snowpack evolution.
(…  )

8. High resolution remote sensing data 1 : Pléiades DEM (4 m)
Alignment
Creating Point Clouds (Z)
Interpolation  DEM
Merging results
Filtering
Validation (GPS campaign measurements)
We used 3 Pléiades stereo to derive a high resolution DEM.
Validation MAE : 4.72 m
NMAD : 4.10 m
|Median| : 2.7 m 5 m precision : 75%
15 m precision : 94%

9. High resolution remote sensing data 2 : Formosat-2 snow cover maps (8 m)
Satellite
Formosat-2
Spatial Resolution
8 m
Swath width
24 km
Pixel Quantization
12 bits
Bands
Blue,Green,Red
Talking about high resolution data, we also used a time series of Formosat-2 images at 8 m of resolution during snow-season to extract SCA by supervised classification. These SCA images will be used to qualify the quality of simulations later.
A time series of Formosat-2 images used to derive SCA by supervised classification.

10. SnowModel
Micromet
Snowpack
EnBal
SnowTran-3D
Micromet : Distributing meteorological variables (T,RH,WS,WD,Prec,Swi,Lwi) over the area using Barnes interpolation (Barnes 1964).
T, RH, P = f(Elevation)
Swi,Lwi = g(slope, orientation)
WS,WD = h(topography)
In order to study carefully the effect of topography on modeling snow in mountainous area we had to use a distributed snow model which take into account the difference of elevations, aspects and orientations when it models snowpacks evolution…
Snowmodel, is composed from 4 sub-models : Micromet which distributes meteorological variables over the area. The distribution of Temperature, RH and Precipitation depend on Elevation. Short and Longwaves depend on slopes and orientation while Wind speed and direction depend on topography
Liston et ..

11. SnowModel
Micromet
Snowpack
EnBal
SnowTran-3D
EnBal : Energy balance model (Liston 1995, Liston et al. 1999)
Qsi: shortwave incoming
Qli: longwave incoming
Qle: longwave emitted
Qh: turbulent sensible heat
Qe: turbulent latent heat
Qc: conductive energy transport = 0
Qm: energy available for melt
α: albedo = f(snow depth)
EnBal : Computes the flux exchanges.

12. 3 D view of simulated SWE (m) by SnowModel ( 1X January 2009 )
Micromet
Snowpack
EnBal
SnowTran-3D
Snowpack : 1-layer snowpack model (Liston & Hall 1995)
Compaction-based snow density evolution.
=> snow depth and SWE
Snow pack : a 1-layer snowpack model simulates the key variables of snowpack based on EnBal and Micromet outputs.
3 D view of simulated SWE (m) by SnowModel ( 1X January 2009 )

13. SnowModel
Micromet
Snowpack
EnBal
SnowTran-3D
SnowTran-3D : Redistribute SD & SWE depending on wind speed and direction. (Optional sub-model). Not activated
And finally, SnowTran-3D an optional sub-model which redistribute SD and SWE depending on Wind speed and direction. It is not activated in our case, due to the lack of information about speed direction.
(Liston et al. 2007)

14. Methods SWE to SCA threeshold = 28 mm
After simulating snowpack over the basin at different resolutions ( 8m to 500 m ) we derive spatialized SD. And we compare locally simulated SD to in-situ SD at Oukaimeden.

15. Methods SWE to SCA threeshold = 28 mm
After simulating snowpack over the basin at different resolutions ( 8m to 500 m ) we derive spatialized SD. And we compare locally simulated SD to in-situ SD at Oukaimeden.

16. Methods SWE to SCA threeshold = 28 mm
After simulating snowpack over the basin at different resolutions ( 8m to 500 m ) we derive spatialized SD. And we compare locally simulated SD to in-situ SD at Oukaimeden.

17. Methods Observed Modeled Confusion matrix : Modeled Snow Snow free
TP (a)
FN (b)
FP (c)
TN (d) b d c a
Here a small description to confusion matrix. We compare modeled and observed SCA matrix pixel by pixel. If both of them define a pixel a snow covered we increment the True Positive score by one. In case of both of them define a pixel as snow free we increment TN … and if there is a dismatch we increment FN or TP. Afterwards, we compute statistical scores related to these scores as described here. The main statistics are global Accuracy AC and Heidke Skill Score HSS.
Statistical scores :
No snow
Snow
AC = (TP + TN ) / Total
HSS = 2 (ad – bc) / [(a+c)(c+d) + (a+b)(b+d)]

18. HSS ( at the end of snowing season )
Results
Score / Resolution 8m
30 m
90 m
250 m
500 m
HSS
0.72
0.71
0.70
HSS ( at the end of snowing season )
0.69
0.66
RMSE(SD)
Correlation (SD)
Before melting season the mean of Heidke Skill Score for 8 m to 250 m yield similar. On the contrary 500 m underestimates SCA and the HSS score is inferior to one obtained from 8 m to 250 m.

19. Results The evolution of total accuracy during Formosat-2 acquisitions
The total accuracy
The evolution of total accuracy during Formosat-2 acquisitions

20. Results The evolution of total accuracy during Formosat-2 acquisitions
These 2 graphs present the evolution of total accuracy score during snow-season, and we can observe again that the AC score for high resolution simulations is better before melting season.
The evolution of total accuracy during Formosat-2 acquisitions

21. HSS ( at the end of snowing season )
Best resolution ?
Score / Resolution 8m
30 m
90 m
250 m
500 m
HSS
0.72
0.71
0.70
HSS ( at the end of snowing season )
0.69
0.66
RMSE (SD)
0.17
0.18
0.19
0.13
0.33
Correlation (SD)
0.90
0.85
Comparing SD with in-situ measurements confirms precedent results. Simulations at 500 m underestimates SD

22. Best resolution ?
Comparing SD with in-situ measurements confirms precedent results. Simulations at 500 m underestimates SD . The bias is too high !
The evolution of simulated SD (8 m to 500 m) compared to measured SD at Oukaimeden (3200 m.a.s.l)

23. Analysis : SWE distribution 22-02-2009
SWE (m)
8 m
30 m
90 m
To better understand the difference in the quality of simulations let’s analyze the distribution of Modeled SWE over a sub-region and compare it to Formosat-2 image. We can see that shaded area in Formosat-2 image correspond to area with maximum SWE in 8 m to 90 m and sometimes for 250 m also. While 500 m simulation consider over the sub-region as homogenous.
250 m
500 m
Formosat-2

24. Analysis : Temporal mean over N-S transect
Distance(m)
Distance(m)
Distance(m)
Distance(m)
Distance(m)
Distance(m)
Moreover, we analyzed a transect N-S as shown in these figures. We can see that incoming solar radiation and snow precipitation are very sensitive to the spatial resolution and therefore the mean SWE is also sensitive to the spatial resolution.
Distance(m)
Distance(m)
Distance(m)
Distance(m)
Distance(m)
Distance(m)
Temporal mean incoming shortwave (W/m²) over the snow-season. From the top to down, right to left : 8m/30m, 8m/90m, 8m/250m and 8m/500m.
Sum Snow Precipitations (mm) over the snow-season. From the top to down, right to left : 8m/30m, 8m/90m, 8m/250m and 8m/500m.
Mean SWE (mm) over the snow-season. From the top to down, right to left : 8m/30m, 8m/90m, 8m/250m and 8m/500m.

25. Diagnostic Melt Timinoutine DAM capacity : 5 Millions m3
Finally, we analyzed cumulative Swemelt over 10 days. As shown in the figure decreasing topographical resolution lead to increase the bias of Swemelt estimation. For example during melting season, this difference can reach 1.5 m3 in 10 days between 8 m simulations and 500 m simulations : It’s equivalent to 33 % of the capacity of medium DAM in Morocco like Timinoutine.
Timinoutine DAM capacity : 5 Millions m3
Total difference between 8 m and 500 m is 5 Millions m3.
During melting season : 10 days difference is higher than 1 Million m3

26. Summary and perspectives
Simulating SWE depends on topographical resolution :
Coarse resolution smooths peaks and  underestimates snowfall in elevated area and overestimated them in low lands.
It smooths orientations and aspects and mask shaded area  Create bias in energy balance, therefore periods of melting are not in the right timing.
Simulations up to 250 m seems to be the best trade off between accuracy and computing time in the Rheraya basin.
Ongoing work : Assimilate SCA by using a global DEM and Sentinel-2 / Landsat data in order to simulate SWE in the Atlas.
To sum up :

27. Questions ?