Operational algorithm for generation of snow depth maps from discrete data in Indian Western Himalaya

    Operational algorithm for generation of snow depth maps from discrete data in Indian Western Himalaya H.S. Gusain, V.D. Mishra, M.K. Arora, Shailesh Mamgain, Dhiraj Kumar Singh PII: DOI: Reference:

S0165-232X(16)30027-1 doi: 10.1016/j.coldregions.2016.02.012 COLTEC 2246

To appear in:

Cold Regions Science and Technology

Received date: Revised date: Accepted date:

18 November 2014 29 January 2016 27 February 2016

Please cite this article as: Gusain, H.S., Mishra, V.D., Arora, M.K., Mamgain, Shailesh, Singh, Dhiraj Kumar, Operational algorithm for generation of snow depth maps from discrete data in Indian Western Himalaya, Cold Regions Science and Technology (2016), doi: 10.1016/j.coldregions.2016.02.012

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

Operational algorithm for generation of snow depth maps from discrete data in Indian Western Himalaya

Department of Civil Engineering, IIT Roorkee, India and PEC university of Technology, Chandigarh, India c

SC

b

Snow and Avalanche Study Establishment (SASE), Chandigarh, India

RI

a

PT

H.S. Gusaina*, V.D. Mishraa, M.K. Arorab, Shailesh Mamgainc and Dhiraj Kumar Singha

Department of Architecture, IIT Roorkee, India

MA

NU

* Corresponding author: [email protected]

Abstract

D

In this paper, an algorithm is proposed for generation of snow depth maps. The efficacy of the

TE

algorithm has been established through a case study in lower and middle Himalayas, India. The

AC CE P

algorithm is a modified version of the spatial interpolation method proposed earlier in Swiss Alps. The method uses discrete point data supplemented with remotely sensed derived information data to create snow depth maps at spatial resolution of 0.5 km. In situ snow depth observations from 14 locations, Automatic Weather Station (AWS) recorded snow depth from 9 locations, Moderate Resolution Imaging Spectroradiometer (MODIS) images and Shuttle Radar Topographic Mission (SRTM) DEM form the database. The algorithm is based on the dependency of snow depth on elevation above mean sea level, which is later adjusted through the in situ snow depth observations to represent the local and regional characteristics of the snow distribution. The algorithm has been validated for different days of the winter season 2012-13 using leave one out station cross validation method. The mean absolute error (MAE) and Root Mean Square Error (RMSE) in estimation of snow depth have been observed as ~ 34 cm and ~ 42 cm respectively during the season. The snow depth maps generated from the proposed 1

ACCEPTED MANUSCRIPT algorithm may be useful in assessment of snow avalanche hazards as well as in various hydrological and glaciological studies in the inaccessible cryospheric region of the Western

PT

Himalaya.

SC

RI

Keywords: Snow depth, Western Himalaya, MODIS, DEM, Automatic Weather Station

NU

1. Introduction

Indian Western Himalaya consists of many parallel mountain ranges from south to north e.g Pir

MA

Panjal range, Great Himalayan range, Zanskar range, Laddakh range and Karakoram range (Gusain et al., 2009). These ranges receive snowfall during the winter season. During this time,

D

most part of the land gets snow covered. Lower and middle Himalayas receive the highest

TE

snowfall and are severely affected by snow avalanches. During the last decade and a half, more

AC CE P

than two hundred fatalities have been reported due to snow avalanches in this region. Therefore, knowledge of snow depth information along with other snow-meteorological parameters is vital for snow avalanche prediction in this region. In situ observations of snow depth are very sparse in this region. Therefore, remote sensing data supplemented with in situ observations are preferred to provide variation in snow depth at spatial level. Snow depth has been studied widely using in situ observations as well as remote sensing observations in different cryospheric regions by various researchers e.g. Shi and Dozier (2000), Brown et al. (2003), Kelly et al. (2003), Romanov and Tarpley (2007), Che et al. (2008), Das and Sarwade (2008), Marty (2008), Dai et al. (2012), Bühler et al. (2014, 2015), Grünewald et al. (2014) etc.. Various approaches have been adopted for mapping the snow depth. These include interpolation of in situ based measurements (Foppa et al., 2007), algorithms for space-borne passive and active microwave observations (Shi and Dozier, 2000; Kelly et al. 2003; Das and 2

ACCEPTED MANUSCRIPT Sarwade, 2008), assimilation of space-borne observations with in situ based measurements (Dai et al. 2012), snow depth estimation using LIDAR data processing before and after the snowfall,

PT

generation of Digital Surface Models (DSMs) of winter and summer terrain using

RI

photogrammetric image correlation technique (Bühler et al., 2014; 2015) etc.. Very few studies have been carried out to map snow depth in Indian Western Himalaya

SC

(Das and Sarwade, 2008; Singh et al., 2007). Singh et al. (2007) developed a regression equation

NU

to estimate snow depth in Patseo region of Great Himalaya using passive microwave SSM/I data along with in situ recorded snow depth. They estimated snow depth at spatial resolution coarser

MA

than 13 km and observed an RMSE of 37.5 cm. However, the regression equation obtained is region specific and can be used for only Patseo region of H.P (Himachal Pradesh) Himalaya, as

TE

D

high errors were obtained for other regions. Das and Sarwade (2008) used AMSR-E data to estimate snow depth in Indian Western Himalaya at a coarse spatial resolution of 5-km. They

AC CE P

used Microwave Emission Model of Layered Snowpacks (MEMLS) along with AMSR-E to understand the difference in the snowpack emitted and sensor received signals and modified algorithm proposed by Chang et al. (1987) for mountainous terrain of Indian Western Himalaya. They found algorithm to be useful for estimation of snow depth from 5 cm to approximately 60 cm with absolute error of 20.34 cm. The main limitation of the algorithm was estimation of snow depth only up to 1 m. However, a major part of the Indian Western Himalaya has snow depth more than 1 m during winter season (Gusain et al., 2009, 2004). These limitations of models in Indian Western Himalaya provide an opportunity to explore for alternate techniques to map snow depth with larger acceptance. In the present paper, a modified version of the spatial interpolation method proposed by Foppa et al. (2007) has been

3

ACCEPTED MANUSCRIPT developed to generate maps of snow depth for its operational use in the lower and middle

PT

Himalayas.

RI

2. Study Area and Data

SC

The study area belongs to lower and middle Himalayas of Jammu & Kashmir from 33.48˚N to

NU

34.83˚N latitude and 73.68˚E to 76.12˚E longitude. It comprises of Kashmir valley, Pir Panjal, Shamshabari and Great Himalayan ranges in India (Figure 1). Kashmir valley is located between

MA

Pir Panjal and Great Himalayan ranges and is around 135 km long and 32 km wide. The elevation ranges between 1600 m to 1900 m a.s.l. approximately. Srinagar is the meteorological

D

station of Snow and Avalanche Study Establishment (SASE) in Kashmir Valley at an altitude of

TE

1652m a.s.l.. In the south of the valley lies Pir Panjal range. This range is generally thickly

AC CE P

forested below 2800 m a.s.l. and tree line exists up to 3300 m elevation. Beyond this elevation, the area is generally barren or rocky with few patches of seasonal grasses. Banihal and Gulmarg are the meteorological stations of SASE in this range at an elevation of 2830m and 2800m a.s.l respectively. Shamshabari range is an arc shape mountain range in the north-west of the Kashmir valley. This range is thickly forested in the lower elevation below 3000 m a.s.l and is generally barren or rocky above 3300 m a.s.l.. Haddan-Taj (3080 m), Stage-2 (2650 m), Ragini (3160 m), Z-Gali (3100 m) and Pharkiyan (2960 m) are meteorological stations in this range. Great Himalayan range lies in the north of the Kashmir Valley and shows a large spatial variability in the terrain. In this range, area is largely barren or rocky, strewn with boulders, glaciated in the higher reaches and vegetated in some valleys. Kanzalwan (2453 m), Sonamarg (2745 m), Drass (3092 m), Pathar (4250 m), Firmbase (4760 m) and Kilnala (3620 m) are the meteorological stations in this range. Pir Panjal and Shamshabari ranges receive higher snowfall amount during 4

ACCEPTED MANUSCRIPT winter compared to Great Himalayan range. However, temperatures are lower in Great

PT

Himalayan range compared to other two ranges.

34.83°N, 76.12°E

RI

Great Himalayan Range

SC

Pharkiyan Ragini Stage-II Haddan Taj

Kilnala Drass

Kashmir Valley

Firmbase Pathar

Sonamarg

Gulmarg

NU

N

Shamshabari Rannge Kanzalwan Z-Gali

Srinagar

< 1000 m 1000 m – 2000 m 2000 m – 3000 m 3000 m – 4000 m 4000 m – 5000 m

Banihal

> 5000 m

AWS locations Manual Meteorological Stations

TE

D

33.48°N, 73.68°E

MA

Pir Panjal Range

Elevation (m.s.l)

AC CE P

Figure 1: Study area and variation in elevation. Circles in the map show manual observation stations and triangles show AWS observations of snow depth.

Elevation of snow covered region varies from 1800 m to 6900 m a.s.l. during winter in these ranges. SASE has 14 manned snow-meteorological observation stations in the study area (shown as circles in Figure 1). Elevation range of the meteorological stations and AWS varies from 1652 m a.s.l in Kashmir Valley region to 4760 m a.s.l. in Great Himalayan range. Most of the stations are installed on a small plain area (approximately 10 m x 10m) on a mountain slope. Srinagar station is located in a large plain area of Kashmir valley. Banihal station is located at the mountain top and high wind is recorded at this station compared to other stations. Cornice formations are frequent during winter around Banihal station due to heavy wind drift. Ragini station in Shamshabari range is near the top of the mountain and high snow depths are recorded at this station compared to nearby stations due to heavy wind drift deposition. Drass station is

5

ACCEPTED MANUSCRIPT located in the valley of the Great Himalayan range and comparatively shallow snow pack observed at this station during peak winter compared to other stations at higher elevation in the

PT

same range. Snow-meteorological data including snow depth are recorded daily at the

RI

observation stations manually. Snow depth data has also been collected using ultrasonic sensors mounted on 9 AWSs at remote locations in the study area (shown as triangles in Figure 1). The

SC

elevation range of AWS varies from 2344 m a.s.l. to 4212 m a.s.l.. One AWS is located in Pir

NU

Panjal range at an altitude of 2615 m a.s.l., two AWSs are in Shamshabari range at an altitude of 2344m and 2736m a.s.l and six AWSs are in Great Himalayan range at an altitude of 2739m,

MA

2941m, 3006m, 3283m, 3850m and 4212m a.s.l..

Moderate Resolution Imaging Spectroradiometer (MODIS) sensor (Salomonson et al.,

TE

D

1989) data has been used for generation of snow cover maps at spatial resolution of 0.5 km in the study area by the method proposed by earlier researchers (Hall et al. 1995, Kulkarni et al., 2006;

AC CE P

Negi et al., 2009; Mishra et al., 2012; Gusain et al., 2014; Sharma et al., 2014). These snow cover maps have been used to determine the snow line in the region. The snow line locations have been used as additional data points having zero snow depth (Foppa et al., 2007) to supplement the in situ data base for spatial interpolation. The elevation information at all the locations (pixels) of the study area has been obtained from processed SRTM DEM freely available at http://srtm.csi.cgiar.org/. The SRTM DEM was compiled by Consultative Group for International Agriculture Research Consortium for Spatial Information (CGIAR-CSI) and made freely available to the public via internet. The vertical accuracy of the CGIAR-CSI processed SRTM DEM is better than a standard SRTM DEM accuracy of 16 m (Gorokhovich and Voustianiouk, 2006). The downloaded SRTM DEM was re-sampled at 0.5 km spatial resolution

6

ACCEPTED MANUSCRIPT to match the spatial resolution of MODIS derived snow cover maps. The re-sampled DEM was used for spatial interpolation of in situ snow depth data.

PT

The daily snow depth data collected at manual observation stations, snow depth data recorded by

RI

AWSs, SRTM DEM and MODIS sensor images forms the data base for the present study to generate snow depth maps using spatial interpolation.

SC

3. Generation of Virtual Snow Depth Data from Satellite Data

NU

Snow cover of Indian Western Himalaya has been mapped widely using remote sensing data obtained from a variety of sensors e.g. AWiFS, MODIS, LISS-3, LISS-4 etc. (Kulkarni et al.,

MA

2006; Negi et al., 2008, 2009; Sharma et al., 2014). SASE has also been involved in producing near real time binary snow cover maps of the Western Himalaya using MODIS data (Negi et al.,

TE

D

2008, 2009). An Earth Receiving Station (ERS) of MODIS has been established at SASE, Chandigarh, which receives daily real time images of Terra and Aqua MODIS at 0530 GMT and

AC CE P

0900 GMT. These images have been processed in near real time for the generation of binary snow cover maps by the method described in Negi et al. (2008, 2009) and Sharma et al. (2014). Lines differentiating between snow and no-snow area in snow cover maps are termed as snow lines. Figure 2 shows the retrieved snow cover map of the study area from the MODIS satellite image for 02-01-2013 (date format is dd-mm-yyyy). Snow depth at the edge of the snow line has been considered as zero. Five such points with zero snow depth have been selected from snow lines in different parts of the study area (shown with star symbol in Figure 2) and has been used to enhance the database. These five points have been chosen in the study area in such a way that at least one point fall in Pir Panjal range, Kashmir Valley, Shamshabari range and Great Himalayan range.

7

ACCEPTED MANUSCRIPT

Snow No-Snow

NU

SC

RI

PT

34.83°N, 76.12°E

33.48°N, 73.68°E

TE

D

MA

Figure 2: Binary snow cover map of the study area for 02-01-2013 retrieved from the MODIS image. Star symbol at five locations shows points with zero snow depth selected manually from different parts of the study area for input data.

AC CE P

A sensitivity analysis with 5, 10, 15, 20 and 25 virtual zero snow depth points were carried out. It was observed that snow depth values were best predicted with 5 virtual snow depth points. Error in prediction of snow depth at the meteorological stations was higher when virtual snow depth points were more than 5. It was observed that snow depth values were comparatively under estimated at meteorological stations when more than 5 virtual zero snow depth points were included.

4. Algorithm to Estimate Snow Depth Spatial interpolation method proposed by Foppa et al. (2007) uses snow depth data at discrete points to create a spatial model of snow depth, from which the snow depth at any location can be estimated. The method is based on the dependency of snow depth on elevation above mean sea level. This general dependency is later adjusted through the in situ snow depth observations to 8

ACCEPTED MANUSCRIPT represent the local and regional characteristics of the snow distribution. The spatial algorithm is implemented in two steps. First, the base value, which describes the correlation between the snow depth and

PT

Step 1:

RI

the elevation of a location via a power function, is determined. The general formulation is given

SC

as,

(1)

NU

where, HSj is the snow depth (cm) at grid cell (location) j, G is the base value (cm) obtained via a mathematical function, Aj is the compensation factor (cm), hej is the elevation or altitude of

MA

location j above mean sea level.

As the general description is an approximation that explains 50–70% of the total variance

TE

D

of the snow depth with elevation (Foppa et al., 2007), the base value is adjusted with a local to regional compensation factor in the second step. The compensation factor value is then added to

Step 2:

AC CE P

the base value to adjust the snow depth value for each grid cell (location) or image pixel.

In this step, the base value is adjusted with a compensation factor. The

compensation factor is calculated as the average of the difference of the base value of snow depth and the measured values of snow depth for the three locations which are nearest to the location where the snow depth has to be estimated. The formula for the compensation factor applied to three locations is given as, (2) here, HSi is the snow depth (cm), Gi is the base value (cm), j is the location or pixel at which the snow depth has to be estimated, i denotes the three nearest locations where base value of snow

9

ACCEPTED MANUSCRIPT depth has been estimated. It is important to note that the three nearest locations are not just planimetric distance-wise but elevation wise also. The distance equation is given as,

PT

(3)

RI

where, (Xj, Yj, hej) are the coordinates of the point where the snow depth has to be estimated and

SC

(Xi, Yi, hei) are coordinates of nearest three locations where base values have been estimated. p is a weighting factor applied to favour horizontal distance over vertical distance or elevation. Foppa

NU

et al. (2007) assigned the value of p=5000 and obtained this value empirically through a sensitivity analysis of the compensation value for various p values.

MA

The algorithm was successfully employed in Swiss Alps. However, in the context of Western Himalayan region where in situ snow depth observations are very sparse in comparison to Swiss

TE

D

Alps, therefore the model may not be directly applied in the Western Himalayan region. A few other limitations of the method proposed by Foppa et al. (2007) are, Only one mathematical function (i.e., power function) has been used to describe the

AC CE P

(1)

dependency of snow depth on elevation above mean sea level. (2)

For the determination of the compensation factor, equal weights have been assigned to

the three locations that are nearest to the location where the snow depth has to be estimated. (3)

The value of weighting factor p has been kept constant for all the days for which snow

depth has been estimated. In this paper, the spatial interpolation method proposed by Foppa et al. (2007) has been suitably modified to get over these limitations. The proposed method appears to be more flexible for generation of snow depth in Western Himalaya. The modifications have been explained in the following section.

10

ACCEPTED MANUSCRIPT 5. Modified Spatial Interpolation Method

PT

In Step 1, to determine the base value (equation 1), instead of using only a power function as the mathematical function, other mathematical functions such as linear, exponential, quadratic, etc.

RI

may be used. The snow depth changes daily due to metamorphic processes and ablation. Spatial

SC

and temporal variation in snow depth may thus not be represented well by only one mathematical

NU

function, so other mathematical functions have also been investigated. The mathematical function that best fits the region may then be selected. In the present study, each of these

MA

functions has been applied to estimate the snow depth. The RMSE between the estimated and observed snow depth for each function is determined. The mathematical function that gives the

D

least RMSE has been selected and used as the function to estimate the base value of the snow

AC CE P

data.

TE

depth. Thus, the type of base equation is not fixed rather dynamic as per the nature of the input

In Step 2, modification in the estimation of compensation factor (equation 2) has been proposed. The compensation factor has been computed as the weighted average of the difference of the base value of the snow depth and the observed values of snow depth at three stations that are nearest to the location at which the snow depth has to be estimated. The weights have been introduced to reduce the errors in estimation of snow depth. After determining the three nearest snow depth stations, the weight to each station has been assigned on the basis of its distance from the location where snow depth has to be estimated. The highest weight is assigned to that snow depth station which has the least distance to the location where snow depth has to be estimated. Furthermore, a dynamic p value has been introduced in equation (3) to determine the nearest station. The weight (we) assigned to each location is given by,

11

ACCEPTED MANUSCRIPT (4) where, dsi is the distance given in equation (3). The normalised weight (nwe) for each station is

(5)

RI

PT

then calculated as,

SC

the formula for the compensation factor has thus been modified as,

(6)

NU

where, ui is the difference between the observed snow depth and estimated snow depth at three

MA

snow depth stations.

Thus, in this study, the compensation factor at each pixel has been estimated using

Optimum value of ‘p’ parameter in equation (3) has been obtained through a sensitivity

TE

i)

D

equation (6) in the following manner,

AC CE P

analysis by studying the RMSE between the observed snow depth and estimated snow depth. ii)

The value of p has been varied from 100 to 5000 at an interval of 100 and the value of p for which least RMSE between observed and estimated snow depth has been obtained, is selected.

iii)

Once p value has been selected, distances of the three nearest stations of a pixel have been computed.

iv)

Snow depth at the three nearest stations has been estimated using base equation (mathematical function).

v)

Difference between the observed snow depth and estimated snow depth ( ) for three nearest stations have been estimated. 12

ACCEPTED MANUSCRIPT vi)

Using equations (5) and (6), normalised weights and compensation factor have been estimated. Using the estimated base value of snow depth

and compensation factor ( ), snow

PT

vii)

RI

depth has been estimated for each pixel of the re-sampled SRTM DEM.

SC

It may be noticed that in the proposed method, the value of p is dynamic and changes with base

NU

equation as per every day data set. Snow depth maps in the study area have been generated for clear sky days at spatial resolution of 0.5 km and the model is applicable for all snow thicknesses

D

6. Validation of the Algorithm

MA

in the study area.

TE

Snow depth maps generated from the proposed algorithm have been validated for a number of

AC CE P

clear sky days of the winter season 2012-13 e.g. 02-01-2013, 07-01-2013, 07-02-2013, 03-032013, 07-03-2013, 25-03-2013 and 05-04-2013. In first step, appropriate base equation has been selected for these days using sensitivity analysis. Five mathematical functions linear, quadratic, power, power with offset and exponential were used to select the appropriate base equation. Selected base equation and optimum p-value has been used to generate the snow depth maps. Table 1 shows the RMSE associated with each mathematical function, selected base equation with least RMSE and optimum p-value associated with selected base equation for different days of the winter season. Sensitivity analysis was carried out with all input data points and snow depth was predicted by different mathematical functions. Mathematical function with least RMSE was selected as base equation. It can be seen from Table 1 that, for most of the days, quadratic equation has been selected as base equation, although power function with offset has also been selected as base equation for few days. It has been noticed that the difference in 13

ACCEPTED MANUSCRIPT optimum p-value for 02-01-2013 (p=200) and just 05 days later for 07-01-2013 (p=1800) is large (Table 1). For both the days, lowest two RMSE values were obtained for p=200 and p=1800. The

PT

difference in RMSE was marginal for both the p-values on both the days. RMSE was slightly

RI

lower by 0.2 for p=200 for 02-01-2013, while RMSE was marginally lower by 0.1 for p=1800 for 07-01-2013. During these 05 days a difference of ~ 10cm in the mean snow depth was also

NU

SC

observed.

02-01-2013

Linear (ahe+b)

D

Base equation

TE

Date

MA

Table 1: RMSE associated with mathematical functions, selected base equation and optimum p-value associated with selected base equation for different days

07-01-2013

07-02-2013

AC CE P

Quadratic (ahe2+bhe+c)

RMSE in Selected Base Optimum sensitivity Equation Value analysis 48.2 cm Quadratic 200 45.7 cm

Power (aheb)

48.2 cm

Power with offset (aheb + c)

46.6 cm

Exponential (a exp (bhe))

48.2 cm

Linear

45.9 cm

Quadratic

43.2 cm

Power

46 cm

Power with offset

44.8 cm

Exponential

46 cm

Linear

89.2 cm

Quadratic

83.2 cm

14

Quadratic

1800

Quadratic

200

p-

85.7 cm

Exponential

89.2 cm

Linear

70.4 cm

Quadratic

66 cm

Power

69.9 cm

Power with offset

64.6 cm

Exponential Linear

D

Quadratic

offset

66.7 cm

Power

62.8 cm

offset

with 4600

66.2 cm

Power with offset

61.3 cm

Exponential

AC CE P

05-04-2013

TE

Power

25-03-2013

with 3800

70.7 cm

MA

07-03-2013

Power

RI

Power with offset

SC

89.2 cm

NU

03-03-2013

Power

PT

ACCEPTED MANUSCRIPT

66.8 cm

Linear

75.7 cm

Quadratic

75.0 cm

Power

75.6 cm

Power with offset

79.1 cm

Exponential

76.1 cm

Linear

61.52 cm

Quadratic

61.4 cm

Power

61.54 cm

Power with offset

64.07 cm

Exponential

61.7 cm 15

Quadratic

1800

Quadratic

1800

ACCEPTED MANUSCRIPT The algorithm has been validated using leave-one-station out cross-validation method. Results of the modified algorithm have also been compared with the results from the original algorithm

PT

proposed by Foppa et al. (2007). Modified algorithm has performed better than the original

RI

algorithm. Figure 3 shows the estimated vs observed snow depth using original and modified algorithm for different days of the winter season 2012-13. It has been observed that the modified

SC

algorithm shows higher correlation coefficient (R2=0.696 vs R2= 0.588), better RMSE (42 cm vs

NU

48 cm) and better MAE (34 cm vs 37 cm) as compared to the original algorithm. Table 2 shows the statistics of comparison between original and modified algorithm for different days of the

MA

season. Modified algorithm had shown better results for all the days. The influence of weighted distance approach and optimum p-value can be seen from the

TE

D

RMSE values obtained. These RMSE values were obtained between estimated snow depth and observed snow depth from base equation, original algorithm and modified algorithm. RMSE in

AC CE P

prediction of snow depth by base equation only varies from ~43 cm to ~89 cm during different days (Table 1). The RMSE was reduced by original algorithm in which fix p-value of 5000 and compensation factor with mean difference of three nearest stations was applied. The RMSE and Mean Absolute Percentage Error (MAPE) in original algorithm were varied in the range of 33 cm to 66 cm and 25.7% to 64.1% respectively (Table 2). Modified algorithm shows further reduction in the errors in which dynamic p-value and weighted distance approach in compensation factor is applied. The RMSE and MAPE by modified algorithm vary between 28cm to 53cm and 22.8% to 57.4% respectively (Table 3). Thus, the results clearly show that modified approach has been able to produce better results in prediction of snow depths during different days of the season than the earlier algorithm.

16

TE

D

MA

NU

SC

RI

PT

ACCEPTED MANUSCRIPT

AC CE P

Figure 3: Estimated vs observed snow depth using original and modified algorithm for different days of winter season 2012-13. Table 2: Brief statistics of comparison between original and modified algorithm for different days of the winter season 2012-13. Date

02-012013 07-012013 07-022013 09-022013 03-032013 07-032013

Correlation Coefficient

RMSE

MAE

Original Modified

Original Modified

Original Modified Original Modified

0.6

0.77

40 cm

32cm

27 cm

25 cm

39.7%

36.8%

0.73

0.82

33 cm

28 cm

25 cm

23 cm

42.8%

39.4%

0.7

0.83

66 cm

53 cm

45 cm

37 cm

33.9%

29.9%

0.78

0.89

43 cm

40 cm

37 cm

35 cm

32.4%

30.6%

0.7

0.75

53 cm

48 cm

42 cm

41 cm

27.3%

26.6%

0.77

0.88

45 cm

41 cm

36 cm

32 cm

25.7%

22.8%

17

MAPE

ACCEPTED MANUSCRIPT 0.8

0.86

57 cm

53 cm

45 cm

44 cm

41.4%

40.4%

0.86

0.88

56 cm

49 cm

48 cm

43 cm

64.1%

57.4%

PT

25-032013 05-042013

RI

From the estimated snow depth data, spatial maps of snow depth have also been prepared and are

SC

shown in figures 4a, 4b, 4c and 4d for 07-01-2013, 07-02-2013, 07-03-2013 and 05-04-2013 days respectively. These maps show the distribution of snow depth and seasonal variations of

NU

snow cover in the study area. Table 3 provides the mean snow depth distribution in the study area during different days of the season 2013 and 2014. Snow cover generally starts building up

MA

in Western Himalaya from November month onwards with the start of snowfall events. Snowfall in Western Himalaya occurs due to Western Disturbances (WDs), which are cyclonic storms

TE

D

associated with the mid-latitude subtropical westerly jet (Dimri et al. 2015). WDs are synoptic weather systems producing precipitation over wide area in Himalayan ranges. Snow depth at a

AC CE P

particular location in the study area depends on many factors including number of snow storms, severity of the storms, wind deposition pattern, snow settlement and ablation pattern. Mean snow depth and maximum snow depth in the study area for 07-01-2013, 07-02-2013, 07-03-2013 and 05-04-2013 are given in Table 3. It can be seen that snow depth in most part of the Shamshabari range (north-west of the Kashmir Valley) for 07-01-2013 is more than 50 cm and maximum snow depth in the study area has also been observed in this range. Snow depth in most part of the Great Himalayan range is below 50 cm for 07-01-2013. Mean snow depth and maximum snow depth increase in the study area for 07-02-2013. Snow depth in most part of the Great Himalayan range is below 100 cm while it has found to be more than 100 cm in most part of Shamshabari range. Kashmir valley has snow depth below 50 cm and Pir Panjal range has a snow depth in the range of 50 cm to 150 cm.

18

ACCEPTED MANUSCRIPT

34.83°N, 76.12°E

PT

Shamshabari Range

RI

Great Himalayan Range

SC

Kashmir Valley

NU

Pir Panjal Range

MA

33.48°N, 73.68°E

D

Figure 4a: Snow depth map for 07-01-2013.

TE

AC CE P

Shamshabari Range

34.83°N, 76.12°E

Great Himalayan Range

Kashmir Valley

Pir Panjal Range

33.48°N, 73.68°E

Figure 4b: Snow depth map for 07-02-2013.

19

ACCEPTED MANUSCRIPT

34.83°N, 76.12°E

PT

Shamshabari Range

RI

Great Himalayan Range

SC

Kashmir Valley

MA

NU

Pir PanjalRange

33.48°N, 73.68°E

AC CE P

TE

D

Figure 4c: Snow depth map for 07-03-2013.

34.83°N, 76.12°E

Shamshabari Range

Great Himalayan Range

Kashmir Valley

Pir PanjalRange

33.48°N, 73.68°E

Figure 4d: Snow depth map for 05-04-2013.

20

ACCEPTED MANUSCRIPT Mean snow depth increases for 07-03-2013, although maximum snow depth decreases. Snow depth in most part of the Shamshabari range appears to be decreasing and probably this may be

PT

the reason for decrease in maximum snow depth as highest snow depth appeared in this range.

RI

Snow depth in most part of the Great Himalayan range appears to be increasing for 07-03-2013. Total area under snow in Great Himalayan range is higher as compared to other ranges and

SC

probably this is the region for higher mean snow depth in the study area for 07-03-2013. Mean

NU

snow depth and maximum snow depth decreases for 05-04-2013 due to melting of the snow. Snow depths in Pir Panjal and Shamshabari ranges are generally higher than Great Himalayan

MA

range during January and February months due to higher snowfall in these two ranges. January onwards, mean snow depth increases in the study area with the accumulation of snow from each

TE

D

snow storm. March onwards temperature in the Pir Panjal and Shamshabari ranges rises above 0°C and ablation of the snow cover starts in these ranges. Due to low temperatures during

AC CE P

January, February and March months, ablation of snow cover starts April onwards in Great Himalayan range. This causes thick snowpack during February and March months in Great Himalayan range. April onwards less number of snow storms are observed in Western Himalaya compared to January, February and March months. Rising temperatures and less snow storms lead to accelerated depletion of snow cover April onwards. Mean snow depth in the study area (Table 3) indicates the build-up of thick snowpack during February and March months in the Himalayan ranges. Deep snow cover in the Himalayan mountain ranges produces frequent snow avalanches during these months. Mean snow depth on 05-04-2013 and 06-05-2014 has been found to be 67.9 + 41.5 cm and 63 + 38 cm respectively, which is less than the mean snow depth observed during February and March months. This indicates significant depletion of snow cover during April and May. Generally, by the end of June month, most of the meteorological stations

21

ACCEPTED MANUSCRIPT of SASE become snow free and snow cover depleted in most part of the study area, except

PT

regions with high elevation above ~ 4500 m a.s.l.

Year

Date

Mean

Snow

depth

07-01-2013

43.9 + 27.9 cm

213 cm

07-02-2013

102.8 + 63.5 cm

417 cm

139.7 + 54.6 cm

291.7 cm

123.4 + 49 cm

283.7 cm

NU

54.3 + 33.6 cm

D

TE

07-03-2013

221 cm

25-03-2013

96.8 + 52.3 cm

274.5 cm

05-04-2013

67.9 + 41.5 cm

216 cm

02-01-2014

37.1 + 17.7 cm

171 cm

30-01-2014

62.5 + 29.7 cm

183 cm

08-02-2014

142.5 + 63.7 cm

347 cm

06-03-2014

124 + 68.1 cm

363 cm

06-05-2014

63 + 38 cm

200 cm

AC CE P

Snow

depth

02-01-2013

03-03-2013

2014

+ Maximum

MA

SC

Standard deviation 2013

RI

Table 3: Mean snow depth with standard deviation and maximum snow depth values

Figure 5 shows the error in estimation of snow depth at different elevation level (elevation of meteorological stations and AWS) during different days of the season. It can be seen that errors in snow depth estimation are observed at all elevation levels from 1652 m a.s.l to 4760 m a.s.l.. Maximum error up to 189 cm has been observed for the station Ragini in

22

ACCEPTED MANUSCRIPT Shamshabari range having elevation of 3160 m a.s.l.. This may be because of high snow depth values recorded at this station due to wind deposited snow compared to other stations in the same

PT

mountain range. This also highlights that locations with heavy wind deposition or wind erosion

RI

patterns are not good measurement sites to be used in generation of spatial snow depth maps. Minimum errors have been observed for the station z-Gali in the Shamshabari range having

SC

elevation of 3100 m a.s.l. The snow depth values for this station have been observed in

NU

consistent with the snow depth values of other nearest stations leading to low errors. This shows that errors in estimation of snow depth are not elevation dependent rather errors are region

MA

specific. Regions having high numbers of meteorological stations with consistent snow depth values are predicted well. Regions with sparse observation stations and high differential snow

TE

D

depth values have shown high errors e.g. Srinagar station at lower elevation (1652 m a.s.l) shows high errors as this is the only meteorological station in the wide Kashmir Valley. Error in

AC CE P

estimation of snow depth in the valley region of Great Himalayan range increases with increase in thickness of snow cover at higher elevation stations due to high differential snowfall at valley region and higher elevations. Drass station lies in the valley region of Great Himalaya and errors during January months observed are generally low for this station as snow cover thickness was in consistent with other higher elevation stations. During the month of February and March, deep snow cover develops at the higher elevations of Great Himalaya due to higher snowfall as compared to valley region. Differential snow depth in valley and high elevations of Great Himalaya become high during February and March months, hence, Drass station has shown higher errors for these months. Well distributed measurement stations in a regular grid over the space and elevation in the study area can improve the errors in snow depth maps.

23

ACCEPTED MANUSCRIPT

PT

Ragini

Z-Gali

TE

D

MA

NU

SC

Drass

RI

Srinagar

AC CE P

Figure 5: Error in estimation of snow depth with respect to elevation on different days of the season.

7. Comparison with Other Studies The results of the present study have been compared with other studies in Western Himalaya and other mountain ranges of the world. Singh et al. (2007) predicted snow depth in Patseo region of Great Himalaya using regression equation developed between brightness temperature of SSM/I microwave images and in situ recorded snow depth. They obtained RMSE of 37.5 cm which was about 30% of the mean snow depth in the study area. Das and Sarwade (2008) estimated snow depth at spatial resolution of 5 km using AMSR-E data and reported an absolute error of 20.34 cm. They found algorithm to be useful for estimation of snow depth from 5 to 60 cm only.

24

ACCEPTED MANUSCRIPT Estimation of snow depth has also been reported in other mountain ranges around the globe. Foppa et al. (2007) reported correlation coefficient between observed and estimated snow

PT

depth in the range of 0.61 to 0.74 during different days of the year 2005 in Swiss Alps. They

RI

observed RMSE and MAE in the range of 21 cm to 67 cm and 15 cm to 45 cm respectively, with in situ recorded snow depth. Romanov and Tarpley (2007) compared snow depth maps derived

SC

from satellite images with surface observations. They have reported an error of 30% for snow

NU

depth below 30 cm, and 50% for snow depths ranging from 30 to 50 cm over mid latitude areas of North America for winter seasons 2003-04 and 2004-05. Recently Buhler et al. (2015)

MA

mapped snow depth in high alpine catchments of Swiss Alps in the neighborhood of Davos, Switzerland using digital photogrammetric technique. They have developed digital surface

TE

D

models (DSMs) of summer and winter terrains at high spatial resolution of 0.25 m using aerial stereo images and generated snow depth maps. They have reported an average accuracy better

AC CE P

than 15% compared to the average snow depth of 2.2 m over the entire study area. The technique produced accurate snow depth maps at high spatial resolution, although the technique was comparatively expensive.

The results of the present study reveal that the modified algorithm may be used to produce snow depth maps at higher spatial resolution with comparable accuracy in Western Himalaya. The study also overcomes the limitations of being region specific and applicable for shallow snow cover. Moreover, modified algorithm has also shown improvement in snow depth estimation over the original algorithm proposed by Foppa et al. (2007) in data sparse region of Western Himalaya.

25

ACCEPTED MANUSCRIPT 8. Conclusion

PT

Indian Western Himalaya receive wide spread snowfall during winter season due to Western Disturbances. Deep snow cover develops in Pir Panjal, Shamshabari and Great Himalayan ranges

RI

during winter months. There are sparse in situ observations of snow depth in these Himalayan

SC

ranges. Sparse in situ observations and lack of appropriate models to map snow depth at spatial

NU

level provided an opportunity to develop a scheme for mapping snow depth in these ranges. A modified spatial interpolation algorithm was proposed in this paper. The algorithm is a modified

MA

version of an algorithm proposed earlier in Swiss Alps. Snow depth observations from 14 meteorological stations, 9 AWSs and 5 virtual zero snow depth points derived from satellite

D

images were used to generate snow depth maps. The algorithm was validated with leave-one-out

TE

cross-validation method. The mean absolute error, Root Mean Square Error and correlation

AC CE P

coefficient between observed and estimated snow depth were obtained as 34 cm, 42 cm and 0.83 respectively for the winter season 2013. The results from the proposed algorithm show improvement in the accuracy of snow depth estimation as compared to the original algorithm. It also overcomes the limitations of other algorithms proposed for Western Himalaya. Although, regions having high numbers of meteorological stations with consistent snow depth values are predicted well, while regions with low numbers of observation stations and high differential snow depth values have shown high errors. Nevertheless, the modified algorithm can map snow depth at higher spatial resolution and is applicable for all snow thicknesses. The snow depth maps have been generated at spatial resolution of 0.5 km for clear sky days, which are quite helpful in observations of temporal and spatial variation of snow cover build-up during winter season. Mean snow depth in the study area increases January onwards and deep snow cover develop during February and March months having mean snow depth more than 100 cm. March 26

ACCEPTED MANUSCRIPT onwards snow cover start depleting due to accelerated melting of snow and by the end of June most part of the study area become snow free. The proposed algorithm fulfill a wide existing gap

PT

in production of snow depth spatial maps for operational purposes and various snow related

RI

applications in Western Himalaya. SASE has been using this algorithm with success to produce snow depth maps of the study area during winter at operational level and on a regular basis.

SC

These maps have been found to be useful in operational avalanche forecasting, hydrological and

NU

glaciological applications in Pir Panjal range, Shamshabri range, Kashmir valley and Great Himalayan ranges. Although, the quality of maps can be improved by well distributed

MA

measurement stations in a regular grid (at least one station in a 30 km x 30 km grid and 0.4 km elevation difference) over the space and elevation. However, such an arrangement will require

TE

D

more than 100 meteorological stations in the study area. The algorithm can also be used in other

AC CE P

cryospheric regions of the globe after validation for the respective regions.

Acknowledgment

Authors are thankful to Shri Ashwagosha Ganju, Director, Snow and Avalanche Study Establishment for constant encouragement and SASE DATA Center for providing in situ and remote sensing data. Website http://srtm.csi.cgiar.org/ is duly acknowledged for providing SRTM DEM for the study.

References Brown, R. D., Brasnett, B., Robinson, D. 2003. Gridded North American monthly snow depth and snow water equivalent for GCM evaluation. Atmosphere-Ocean. 41, 1-14.

27

ACCEPTED MANUSCRIPT Bühler, Y., Marty, M., Egli, L., Veitinger, J., Jonas, T., Thee, P., Ginzler, C., 2014. Spatially continuous mapping of snow depth in high alpine catchments using digital photogrammetry.

PT

The Cryosphere Discussion, 8, 3297-3333.

RI

Bühler, Y., Marty, M., Egli, L., Veitinger, J., Jonas, T., Thee, P., Ginzler, C., 2015. Snow depth mapping in high alpine catchments using digital photogrammetry. The Cryosphere, 9, 229-

SC

243.

NU

Chang, A. T. C., Foster, J. L., Hall, D. K., 1987. Nimbus-7 derived global snow cover parameters. Annals of Glaciology. 9, 39-44.

MA

Che, T., Li, X., Jin, R., Armstrong, R., Zhang., T., 2008. Snow depth derived from passive microwave remote sensing data in China. Annals of Glaciology. 49, 145-154.

TE

D

Dai, L., Che, T., Wang, J., Zhang, P., 2012. Snow depth and snow water equivalent estimation from AMSR-E data based on a priori snow characteristics in Xinjiang, China. Remote

AC CE P

Sensing of Environment, 127, 14-29. Das, I., Sarwade, R. N., 2008. Snow depth estimation over north-western Indian Himalaya using AMSR-E. International Journal of Remote Sensing. 29, 4237-4248. Dimri, A. P., Niyogi, D., Barros, A. P., Ridley, J., Mohanty, U. C., Yasunari, T., Sikka, D. R., 2015.

Western

Disturbances:

A

Review.

Reviews

of

Geophysics.

53,

doi:10.1002/2014RG000460. Foppa, N., Stoffel, A., Meister, R., 2007. Synergy of in situ and space borne observation for snow depth mapping in the Swiss Alps. International Journal of Applied Earth Observation and Geoinformation. 9, 294-310.

28

ACCEPTED MANUSCRIPT Gorokhovich, Y. Voustianiouk, A., 2006. Accuracy assessment of the processed SRTM-based elevation data by CGIAR using field data from USA and Thailand and its relation to the

PT

terrain characteristics. Remote Sensing of Environment. 104, 409-415.

RI

Grünewald, T., Bühler, Y., Lehning, M., 2014. Elevation dependency of mountain snow depth. The Cryosphere, 8, 2381-2394.

SC

Gusain, H. S., Chand, D., Thakur, N. K., Singh, A., Ganju, A., 2009. Snow avalanche

NU

climatology of Indian Western Himalaya. Proceedings of International Symposium on Snow and Avalanches. 6-10 April 2009, SASE Manali, India. 85-93.

MA

Gusain, H. S., Singh, A., Ganju, A., Singh, D., 2004. Characteristics of the seasonal snow cover of Pir Panjal and Great Himalayan ranges in Indian Himalaya. International Symposium on

TE

D

Snow Monitoring and Avalanches, 12-16 April, 2004, Manali, India. 97-102. Gusain, H. S., Mishra, V. D., Arora, M. K., 2014. Estimation of net shortwave radiation flux of

AC CE P

western Himalayan snow cover during clear sky days using remote sensing and meteorological data. Remote Sensing Letters. 5, 83-92. Hall, D. K., Riggs, G. A., V. V., Salomonson, 1995. Development of methods for mapping global snow cover using Moderate Resolution Imaging Spectroradiometer (MODIS) data. Remote Sensing of Environment. 54, 127–140. Jonas, T., Marty, C., Magnusson, J., 2009. Estimating the snow water equivalent from snow depth measurements in the Swiss Alps. Journal of Hydrology. 378, 161-167. Kelly, R. E., Chang, A. T., Tsang, L., Foster, J. L., 2003. A prototype AMSR-E global snow area and snow depth algorithm. IEEE Transactions on Geoscience and Remote Sensing, 41, 230242.

29

ACCEPTED MANUSCRIPT Kulkarni, A. V., Singh, S. K., Mathur, P., Mishra, V. D., 2006. Algorithm to monitor snow cover using AWiFS data of RESOURCESAT-1 for the Himalayan region. International Journal of

PT

Remote Sensing, 27, 2449-2457.

RI

Marty, C. 2008. Regime shift of snow days in Switzerland. Geophysical Research Letters. 35, doi:10.1029/2008GL033998.

SC

Mishra, V. D., Gusain, H. S., Arora, M. K., 2012. Algorithm to derive narrow band to broad

NU

band albedo for snow using AWiFS and MODIS imagery of Western Himalaya-validation. International Journal of Remote Sensing Applications, 2, 52-62.

MA

Negi, H. S., Snehmani, Thakur, N. K., 2008. Operational Snow Cover Monitoring in NWHimalaya using Terra and Aqua MODIS Imageries. Proceedings of International workshop

TE

D

on Snow, Ice, Glacier and Avalanches. 07-09 Jan 2008, IIT Bombay, Mumbai, India. Negi, H. S., Kulkarni, A. V., Semwal, B. S., 2009. Estimation of snow cover distrubution in Beas

AC CE P

Basin, Indian Himalaya using satellite data and ground measurements. Journal of Earth System Science. 118, 525-538. Romanov, P., Tarpley, D., 2007. Enhanced algorithm for estimating snow depth from geostationary satellites. Remote Sensing of Environment. 108, 97-110. Salomonson, V. V., Barnes, W. L., Maymon, P. W., Montgomery, H. E., Ostrow, H., 1989. MODIS: Advanced Facility Instrument for Studies of the Earth as a System, IEEE Transactions on Geoscience and Remote Sensing, 27, 145–153. Sharma V., Mishra, V. D., Joshi, P. K., 2014, Topographic controls on spatio-temporal snow cover distribution in Northwest Himalaya. International Journal of Remote Sensing. 35, 3036-3056.

30

ACCEPTED MANUSCRIPT Shi, J., Dozier, J., 2000. Estimation of snow water equivalence using SIR-C/X-SAR, Part II: Inferring snow depth and particle size. IEEE Transactions on Geoscience and Remote

PT

Sensing. 38, 2475-2488.

RI

Singh, K. K., Mishra, V. D., Negi, H. S., 2007. Evaluation of snow parameters using Passive

AC CE P

TE

D

MA

NU

SC

Microwave Remote Sensing. Defence Science Journal. 57, 271-278.

31

ACCEPTED MANUSCRIPT

Highlights (for review)

PT

RI

SC

NU MA D



TE



An algorithm for snow depth estimation in Western Himalaya is proposed. Proposed algorithm improves upon the limitations of earlier published snow depth interpolation algorithm. It also has advantages over the previous models for estimation of snow depth in Western Himalaya. Snow depth maps generated from the algorithm are being used operationally in Western Himalaya.

AC CE P

 

32