The Cryosphere, 18, 451–474, 2024

[Link]
© Author(s) 2024. This work is distributed under
the Creative Commons Attribution 4.0 License.

Passive microwave remote-sensing-based high-resolution snow

depth mapping for Western Himalayan zones using multifactor
modeling approach
Dhiraj Kumar Singh1,4 , Srinivasarao Tanniru1 , Kamal Kant Singh3 , Harendra Singh Negi3 , and
RAAJ Ramsankaran1,2
1 Hydro-Remote Sensing Applications(H-RSA) Group, Department of Civil Engineering, Indian Institute of Technology
Bombay, Powai, Mumbai 400076, India
2 Interdisciplinary Program in Climate Studies, Indian Institute of Technology Bombay, Powai, Mumbai 400076, India
3 Defence Geoinformatics Research Establishment, Him Parisar, Sector 37A, Chandigarh 160036, India
4 Kalpana Chawla Centre for Research in Space Science & Technology (KCCRST), Chandigarh University,

Mohali 140413, India

Correspondence: RAAJ Ramsankaran (ramsankaran@[Link])

Received: 24 April 2023 – Discussion started: 16 June 2023

Revised: 29 November 2023 – Accepted: 30 November 2023 – Published: 31 January 2024

Abstract. Spatiotemporal snow depth (SD) mapping in the (UHZ). Multifrequency brightness temperature (TB) obser-
Indian Western Himalayan (WH) region is essential in many vations from Advanced Microwave Scanning Radiometer
applications pertaining to hydrology, natural disaster man- 2 (AMSR2), SCD data, terrain parameters (i.e., elevation,
agement, climate, etc. In situ techniques for SD measure- slope, and ruggedness), and geolocation for the winter pe-
ment are not sufficient to represent the high spatiotemporal riod (October to March) during 2012–2013 to 2016–2017 are
variability in SD in the WH region. Currently, low-frequency used for developing the SD models for dry snow conditions.
passive microwave (PMW) remote-sensing-based algorithms Different regression approaches (i.e., linear, logarithmic, re-
are extensively used to monitor SD at regional and global ciprocal, and power) are used to develop snow depth models,
scales. which are evaluated further to find if any of these models
However, fewer PMW SD estimation studies have been can address the heterogeneous association between SD ob-
carried out for the WH region to date, which are mainly con- servations and PMW TB. From the results, it is observed
fined to small subregions of the WH region. In addition, the from the analysis that the power regression SD model has
majority of the available PMW SD models for WH locations improved accuracy in all WH zones with the low root mean
are developed using limited data and fewer parameters and square error (RMSE) in the MHZ (i.e., 27.21 cm) compared
therefore cannot be implemented for the entire region. Fur- to the LHZ (32.87 cm) and the UHZ (42.81 cm). The spatial
ther, these models have not taken the auxiliary parameters distribution of model-derived SD is highly affected by SCD,
such as location, topography, and snow cover duration (SCD) terrain parameters, and geolocation parameters and has better
into consideration and have poor accuracy (particularly in SD estimates compared to regional and global products in all
deep snow) and coarse spatial resolution. zones. Overall results indicate that the proposed multifactor
Considering the high spatiotemporal variability in snow SD models have achieved higher accuracy in deep snowpack
depth characteristics across the WH region, region-wise mul- (i.e., SD >25 cm) of the WH region compared to previously
tifactor models are developed for the first time to estimate SD developed SD models.
at a high spatial resolution of 500 m × 500 m for three dif-
ferent WH zones, i.e., Lower Himalayan Zone (LHZ), Mid-
dle Himalayan Zone (MHZ), and Upper Himalayan Zone

Published by Copernicus Publications on behalf of the European Geosciences Union.

452 D. K. Singh et al.: Himalayan snow depth mapping

1 Introduction sive archive of historical data make spaceborne PMW remote

sensing data extensively useful for the retrieval of SD (Dietz
Snow is an essential land cover type and an important et al., 2012; Tedesco and Narvekar, 2010; Luojus et al., 2021;
cryosphere component. The snow cover encompasses an Chang et al., 1987).
aerial extent of approximately 45 × 106 km2 in the peak The historical PMW data and ongoing and planned mis-
winter over the Northern Hemisphere (Estilow et al., 2015; sions have paved the way for developing numerous SD in-
Lemke et al., 2007). Among many cryosphere regions in the version algorithms across the different cryosphere regions of
Northern Hemisphere, the Indian Western Himalayan (WH) the earth. Many studies of SD estimation have been carried
region is a unique snow-covered region with a complex to- out using multifrequency brightness temperature (TB) obser-
pography and high spatiotemporal variability in snow depth vations collected from PMW sensors on board different satel-
(SD) and diverse land cover types (Singh et al., 2018; Das lites (Chang et al., 1987; Saraf et al., 1999; Xiao et al., 2018;
and Sarwade, 2008; Thakur et al., 2019; Sharma et al., 2014; Kelly et al., 2005; Takala et al., 2011; Dai et al., 2018; Jiang
Singh et al., 2016). The WH region comprises three moun- et al., 2014; Singh et al., 2012). The volumetric PMW scat-
tain zones, e.g., Lower Himalayan Zone (LHZ), Middle Hi- tering increases while PMW TB decreases with an increase
malayan Zone (MHZ), and Upper Himalayan Zone (UHZ), in SD. The PMW brightness temperature difference (BTD)
and receives significant snowfall during winter (Dimri and of 18 and 36 GHz frequency increases with an increase in
Dash, 2012; Gurung et al., 2011; Kumar et al., 2019; Sharma SD up to a specific thickness and then saturates depending
and Ganju, 2000; Singh et al., 2016, 2014). The variation on snowpack conditions (Rango et al., 1979; Chang et al.,
in snow volume and its melt rate affects the availability of 1987; Tedesco and Narvekar, 2010). Hence, many studies of
fresh water for drinking, hydropower, irrigation facilities, PMW SD inversion relied on empirical models derived us-
and ecosystem conditions for millions of people residing in ing BTD between 18 and 36 GHz frequency from TB ob-
the foothills of WH zones (Singh et al., 2016; Thakur et al., servations (Chang et al., 1987; Saraf et al., 1999; Foster et
2019; Nüsser et al., 2019; Negi et al., 2020; Ahmad, 2020; al., 1997; Kelly et al., 2005, 2003; Das and Sarwade, 2008).
Mukherji et al., 2019; Vishwakarma et al., 2022). Further, Many of the empirical models for SD are developed by gen-
the variability in snow characteristics such as SD, density, eralizing the snowpack parameters such as snow density and
and volume and mountainous topography triggers frequent grain size (Chang et al., 1987, 1997; Kelly et al., 2003). How-
avalanches in the WH region, which have resulted in more ever, these parameters dynamically vary with space and time.
than 1000 casualties as reported in different studies (Ganju As a result, the applicability of many empirical SD models
et al., 2002; McClung, 2016; Gusain et al., 2016). There- (Chang et al., 1987; Foster et al., 1997; Aschbacher, 1989)
fore, quantifying snow variables, especially SD, is an essen- outside their study region is not good, as evident from several
tial field of study in the WH region. studies (Dai et al., 2018; Wang et al., 2019, 2020; Saraf et al.,
Traditionally SD information is acquired using in situ 1999; Xiao et al., 2018). Further, many PMW studies have
measurements from snow stakes, snow poles, ground pene- shown that the error in estimated SD using TB data varies
trating radar, automatic weather stations, etc. (Dong, 2018; with snow conditions (i.e., wetness, grain size, density), land
Kinar and Pomeroy, 2015). In situ methods provide accu- cover, topography, ground SD, etc. (Dai et al., 2018; Tedesco
rate SD; however, these techniques have several drawbacks, and Narvekar, 2010; Tedesco et al., 2010; Kelly et al., 2002;
such as limited spatial coverage, operational and mainte- Yang et al., 2021; Wang et al., 2010; Ansari et al., 2019).
nance constraints under harsh weather and complex terrain Different combinations of multifrequency PMW TB observa-
conditions, instrument calibration and malfunctioning issues, tions, snow information (i.e., snow cover fraction, grain size,
and high logistics and personnel requirements (Kinar and density), and auxiliary data such as topographical and land
Pomeroy, 2015; Gusain et al., 2016). In the WH region, cover information are used in the PMW-based SD model de-
because of the rocky terrain and harsh climatic conditions, velopment to account for these limitations (Dai et al., 2018;
a sparse network of snow monitoring stations is available Wang et al., 2020, 2019). Many SD modeling approaches
(Saraf et al., 1999; Singh et al., 2016; Gusain et al., 2016). comprising static empirical linear (Chang et al., 1987; Saraf
Apart from this, the available SD observations from the et al., 1999; Singh et al., 2012) and nonlinear models (Wang
in situ network are spatially and temporally discontinuous et al., 2020, 2019), dynamic models (Tedesco et al., 2010;
and inadequate for demonstrating the snowpack at a regional Grippa et al., 2004; Wei et al., 2021), snow emission mod-
scale, particularly in the high-altitude regions of the WH re- els (Dai et al., 2018; Yang et al., 2021), machine learning
gion. Spaceborne passive microwave (PMW) remote sens- algorithms (Xiao et al., 2018; Yang et al., 2020), assimilation
ing observations can partially compensate for these limita- schemes (Kwon et al., 2017; Graf et al., 2006), etc., have
tions and effectively monitor large areas for SD at a com- been developed using PMW TB and auxiliary datasets for
paratively low cost under all weather and terrain conditions different regions.
(Dietz et al., 2012; Amlien, 2008; Bernier, 1987; Xiao et al., Despite the significant progress in PMW-based SD esti-
2018). Sensitivity to snowpack characteristics, global cover- mation, very few studies have been carried out in the In-
age, daily temporal resolution, and availability of an exten- dian WH region using PMW data (Singh et al., 2012; Das

The Cryosphere, 18, 451–474, 2024 [Link]

D. K. Singh et al.: Himalayan snow depth mapping 453

and Sarwade, 2008; Saraf et al., 1999; Singh et al., 2015). close to ground data with a root mean square error (RMSE)
The WH region, being a tropical region, experiences signifi- of ∼ 16 cm and MAE of ∼ 13.9 cm.
cant changes in temperature, leading to frequent melt–freeze Despite the development of various PMW SD models for
snow events causing snow grain growth, which introduces Himalaya in the last 2 decades (1999–2020), there are many
errors in the estimation of PMW SD (Singh et al., 2015). constraints in the spatiotemporal estimation of SD for the
Further, the limited availability of in situ SD observations, WH region. Many of the previous studies for SD estimation
very high SD (i.e., >1 m), and high spatiotemporal variabil- in the WH region have been carried out specifically for sub-
ity in snowpack characteristics pose numerous constraints for regions of the WH region, such as the Sutlej basin, Dhundi,
PMW SD estimation in the WH region. Consequently, no and Patseo. The PMW TB observations are affected by het-
studies were reported for PMW SD estimation in WH till erogeneity in snowpack properties, land cover, topography,
1999. For the first time, Saraf et al. (1999) estimated the av- etc. (Trujillo et al., 2007; Wang et al., 2010; Che et al., 2016;
erage monthly SD using Scanning Multichannel Microwave Derksen, 2008; Foster et al., 2005). However, previous stud-
Radiometer (SMMR) data on board Nimbus-7 during 1979– ies (Das and Sarwade, 2008; Saraf et al., 1999; Singh et al.,
1987 for the Sutlej valley region of Himalaya using the mod- 2012, 2020, 2007; Singh and Mishra, 2006) have not ac-
ified Chang model (Chang et al., 1992). However, the appli- counted for the aforementioned variables. Further, the accu-
cability of this model (Saraf et al., 1999; Chang et al., 1992) racy of SD retrievals from these models is also not evaluated
over the entire Himalaya cannot be justified as the model is with respect to the varying terrain and snow parameters. The
developed using less in situ data (from 11 stations) where accuracy of operational PMW SD products available in the
the stations are not distributed and is not tested outside the WH region, i.e., AMSR2 SD, has not been evaluated.
Sutlej basin. Singh and Mishra (2006) have proposed three Additionally, the AMSR2 SD product and previous PMW
empirical models using Advanced Scanning Microwave Ra- SD models have course resolution and have limitations
diometer for Earth (AMSR-E) data (horizontally polarized on their potential utility in various applications such as
TB of 18.7 and 36.5 GHz) for SD estimation in the Pir Panjal, avalanche susceptibility and hydrological modeling, espe-
Greater Himalaya, and Karakoram ranges of the WH region. cially at the regional scale. Considering these research gaps,
Following this study, Singh et al. (2007) used different empir- in the current study, different linear and nonlinear empirical
ical models for SD estimation using multifrequency Special models are developed to improve and estimate SD at high
Sensor Microwave/Imager (SSM/I) data (i.e., TB of 19, 22, resolution, i.e., 500 m for different WH zones, using a mul-
37, and 89 GHz during 1997–2002) over the Patseo region. tifactor approach. In this approach, multifrequency PMW
However, these studies (Singh and Mishra, 2006; Singh et observations from AMSR2 (during 2012–2019), terrain pa-
al., 2007; Saraf et al., 1999) have not provided any quanti- rameters, land cover parameters, and the Moderate Reso-
tative details about the accuracy of SD estimates and have lution Imaging Spectroradiometer (MODIS)-derived snow
not been evaluated using independent SD observations. Das cover product are statically correlated with the ground SD
and Sarwade (2008) used 18.7 and 36.5 GHz horizontally po- observations for the development and evaluation of the SD
larized data from AMSR-E and modified the coefficients of models. The accuracy of PMW multifactor SD models’ es-
Chang et al.’s (1987) model to suit the Indian Himalaya. The timates is compared with previous models and the AMSR2
modified model has shown a mean absolute error (MAE) of SD product. Further, in this study, the SD retrieval accuracy
20.34 cm in SD estimates but failed to estimate SD above is also analyzed with respect to different auxiliary parame-
60 cm. Singh et al. (2012) have developed multiple empir- ters. The present study has the following three objectives:
ical SD models for three SD classes, i.e., 1 to 5, 5 to 50,
– development of multifactor SD models to estimate SD
and 50 to 200 cm, in the Pir Panjal, Greater Himalayan, and
at high resolution for different WH zones;
Karakoram regions of the WH region using TB data of dif-
ferent frequencies from SSM/I. Their approach (Singh et al., – comparison and evaluation of the proposed multifactor
2012) has used the scattering index to estimate snow cover model, previous SD models, and AMSR2 SD products
and TB thresholds for identifying the SD class and estimation in different WH zones;
of SD. In another study, Singh et al. (2015) developed PMW
– analysis of multifactor SD retrieval accuracy with re-
SD models for the Dhundi and Patseo regions of Himalaya
spect to selected auxiliary variables (i.e., elevation,
using data from ground-based radiometers and in situ obser-
slope, land cover types, and snow cover duration –
vations. However, SD models were developed using observa-
SCD).
tions collected from only two field surveys, evaluated using
a single-day observation of AMSR-E TB data, and not tested Following this introduction section (Sect. 1), the current arti-
spatiotemporally. Recently, Singh et al. (2020) developed an cle is organized as follows. The topographical and geograph-
empirical algorithm for the Patseo region of the MHZ us- ical description of the study area is described in Sect. 2. The
ing Advanced Microwave Scanning Radiometer 2 (AMSR2) details of the in situ observation network and various remote
18.7 and 36.5 GHz TB (i.e., during 2012–2016) and in situ sensing datasets used for model development and evalua-
observations. They observed that the estimated SD is very tion are also given in the same section. Following that, the

[Link] The Cryosphere, 18, 451–474, 2024

454 D. K. Singh et al.: Himalayan snow depth mapping

methodology used in developing the multifactor model is

presented in Sect. 3. Subsequently, Sect. 4 describes the per-
formance of different multifactor models developed for the
three WH zones, a comparison of the different SD models,
and results from the analysis of multifactor SD model re-
trievals with respect to auxiliary parameters. The discussion
and summary are given in Sects. 5 and 6, respectively.

2 Study area and datasets

The topographic and environmental conditions prevailing in

the WH region are detailed in Sect. 2.1. This study makes use
of in situ data from the snow monitoring network and vari-
ous spaceborne data for the development of SD models for
different WH zones. These datasets along with their sources Figure 1. (a) Elevation variability in WH zones (i.e., LHZ: Lower
are listed in Table 1 and briefly discussed in the following Himalayan Zone; MHZ: Middle Himalayan Zone; UHZ: Upper Hi-
subsections from 2.2 to 2.7. malayan Zone) and DGRE observatory distribution (Note: J&K is
Jammu and Kashmir, HP is Himachal Pradesh).
2.1 Study area

Himalaya is the largest snow-covered territory outside the and the MHZ has a temperate climate, while the UHZ has po-
polar regions in the world (Gurung et al., 2011). The present lar climatic conditions with the presence of permanent snow.
study encompasses the entire WH region, which is a signif- Further, these zones have different timings and intensities
icant portion of the Indian Himalaya, situated in the states of precipitation. The LHZ has comparatively warmer condi-
of Jammu and Kashmir, Ladakh, and Himachal Pradesh (see tions, with mean monthly temperatures varying between (−3
Fig. 1). The WH region extends between longitudes from to 18 ◦ C), than the MHZ (−10 to 14 ◦ C) and UHZ (−25 to
73◦ 150 to 79◦ 450 E and latitudes from 30◦ 000 to 39◦ N and 0 ◦ C). As the latitude increase, the amount of precipitation
covers an area of 360 866 km2 . The WH region is unique deceases in the WH region. Negi et al. (2018) reported aver-
with its perennial snow-covered mountain peaks and sea- age winter precipitation (in terms of snow water equivalent)
sonal snow-covered valleys. Approximately 65 % of the ter- of ∼ 804, 549, and 431 mm in the LHZ, MHZ, and UHZ,
rain in the WH region is situated at an altitude of more respectively, during 1991–2015. Further, the snowpack per-
than 3000 m a.m.s.l. (above mean sea level) and is under- sistence varies based on the local weather conditions, which
lain by extremely steep and rugged mountains. The high- mainly changes with elevation across the three WH zones
altitude terrain and mountain topography influence both win- (Sharma et al., 2014).
ter precipitation (caused by western disturbances) and mon-
soon precipitation patterns (Dimri and Dash, 2012). Due to 2.2 Ground observatory stations data
prevailing topographical and weather conditions in the WH
region, forest cover is present only up to 3000 m a.m.s.l., and In the WH region, the Defence Geoinformatics Research
between 3000–4000 m a.m.s.l. thin vegetation consisting of Establishment (DGRE) (formally known as Snow and
shrubs and grass is present, whereas above 4000 m a.m.s.l. Avalanche Research Establishment) operates and maintains
altitude, vegetation is not present, and the land cover there a network of 43 observatory stations (see Fig. 1) which mea-
is predominantly comprised of barren land with snow and sures daily in situ SD twice (i.e., forenoon and afternoon)
ice. The WH region generally receives snow from October to along with other meteorological parameters such as temper-
March; from April onwards, snowmelt generates runoff con- ature and rainfall. Out of total 43 stations, 16 stations are lo-
tributing water to many rivers and streams within the region cated in the LHZ, 13 in the MHZ, and 14 in the UHZ. From
(Dimri and Dash, 2012; Sharma et al., 2014). the LHZ, MHZ, and UHZ observatories, elevation varies be-
In this study, three WH zones, i.e., LHZ, MHZ, and UHZ, tween 1652 and 3785 m a.m.s.l., 2440 and 4950 m a.m.s.l.,
defined based on the historical local meteorological and and 3960 and 5995 m a.m.s.l., respectively. In this study,
avalanche occurrence data (Sharma and Ganju, 2000), are in situ data comprising station name, date, latitude, longi-
used for developing multifactor SD models. The geomorphic tude, and SD for the 43 stations are obtained for the snow
and climate characteristics of these zones are given in Ta- period from 2012–2013 to 2018–2019. The in situ data are
ble 2. The three zones differ in regional topographical and grouped according to the WH zones for the development of
climatic conditions with varying elevations, temperatures, different multifactor SD models. The mean in situ SD of
rainfall, snowfall, etc. The LHZ has a subtropical climate, stations varies between 11 and 256 cm in the LHZ, 23 and

The Cryosphere, 18, 451–474, 2024 [Link]

D. K. Singh et al.: Himalayan snow depth mapping 455

Table 1. Sources of in situ, remote sensing datasets and their application in the present study.

Data Source Role/applications

In situ snow depth data DGRE Data Centre, Chandigarh, India Development and validation of multi-
factor SD models
AMSR2 [Link] (last access: 26 November 2023) Development, validation, and compari-
brightness temperature son of multifactor SD models
snow depth product
MODIS land cover [Link] Development of SD models
data (MCD12Q1) (last access: 26 November 2023)
Daily MODIS cloud- [Link] Development of models and model per-
free snow cover product 10.1594/PANGAEA.918198 (Muhammad, 2020) formance analysis
Digital elevation [Link] (last access: 21 January 2024) Development and model performance
model analysis

Table 2. Geomorphic characteristics of the WH zones.

Characteristics/ranges Lower Himalaya Middle Himalaya Upper Himalaya

Area (km2 ) 41 107 73 951 38 441
Elevation 1500–4800 1500–5700 1800–8100
Climate type Subtropical Temperate Polar
Winter snowfall High Moderate Scant
(from–to) (Dec–Mar) (Oct–Apr) (entire year)
Frequency of high-intensity Occasional Frequent Occasional
snowfall events
Vegetation cover presence Forest: <3000 m a.m.s.l. Grass: <3000 m a.m.s.l. –
Grass: 3000–4000 m a.m.s.l.
Snowpack persistence Early spring Spring Summer
up to

136 cm in the MHZ, and 52 and 356 cm in the UHZ during cess: 21 January 2024). In many locations of the WH region,
the study period. the temperatures exceed 0 ◦ C from April to September, lead-
ing to snowmelt (Negi et al., 2018; Sharma et al., 2014). The
2.3 AMSR2 brightness temperatures data resulting wet snow can lead to saturation of PMW TB (Dong
et al., 2005; Stiles and Ulaby, 1980; Tedesco et al., 2015), af-
AMSR2 is a PMW sensor on board the Japanese Aerospace fecting the accuracy of SD estimates from PMW SD models.
Exploration Agency’s (JAXA) Global Change Observation Therefore, in this study, the level-3 TB of ascending and de-
Mission 1st – Water (GCOM-W1) SHIZUKU, launched in scending orbital passes from the AMSR2 sensor are obtained
May 2012 (Imaoka et al., 2011). It is a follow-on instrument for the snow/winter period (October to March) from 2012 to
to AMSR and AMSR-E sensors and records upwelling mi- 2019 to develop the SD models.
crowave emissions from the earth’s surface in 14 channels
in the form of TB. AMSR2 TB observations are available 2.4 AMSR2 snow depth product
in seven frequencies (6.9, 7.3, 10.65, 18.7, 23.8, 36.5, and
89 GHz, hereafter referred to as 6, 7, 10, 18, 23, 36, and In this study, the AMSR2 SD products have been down-
89 GHz) at two polarizations (horizontal and vertical) for as- loaded from the website [Link] (last access:
cending and descending orbit passes with a temporal reso- 21 January 2024) from the snow season (October to March)
lution of 1 d. The multifrequency TB observations are re- from 2012 to 2019. The SD products corresponding to as-
gridded to 10 km spatial resolution (level-3 product) and are cending (13:30 ± 15 min) and descending (01:30 ± 15 min)
archived in the JAXA portal ([Link] last ac- passes have been used for comparison with the multifactor

[Link] The Cryosphere, 18, 451–474, 2024

456 D. K. Singh et al.: Himalayan snow depth mapping

SD model estimates. The standard AMSR2 SD algorithm pri- 2.5 SRTM digital elevation model
marily uses the daily brightness temperature data for the 10,
18, 23, 36, and 89 GHz frequencies and the surface physical Topography affects the rate of snow accumulation, ablation,
temperature (T ) data. In the development of the AMSR2 SD and redistribution. In the current study, Shuttle Radar To-
algorithm (Kelly, 2009), the following steps and conditions pography Mission (SRTM) digital elevation model (DEM)
have been considered. version 004 data at 90 m spatial resolution are used to ac-
count for the topographic effects in the SD model. The SRTM
– Step 1 – isolate wet and dry snow, non-snow-covered re- DEM for the entire earth is generated using the interfero-
gions. If dry snow is present in any region, it will satisfy metric synthetic aperture radar method (Farr et al., 2007;
conditions (1) and (2) (move to step 2); otherwise, there Jarvis et al., 2008) and can be downloaded from the web por-
is no snow-covered region, or only wet snow is present. tal ([Link] last access: 21 January 2024) in
GeoTIFF format. It has a minimum vertical accuracy of 16 m
Tb36H < 245 K (1) and RMSE of 9.73 m across the globe (Mukul et al., 2017).
Tb36V < 255 K (2) The SRTM DEM data are re-projected to the GCS-WGS-
1984 coordinate reference system and then mosaicked, ex-
– Step 2 – isolate moderate/deep and shallow snow- tracted, and resampled to 500 m spatial resolution. The ele-
covered areas. If moderate/deep snow is present, it will vation varies significantly across different WH ranges. The
satisfy conditions (3) and (4) (move to step 4) (Derksen, LHZ and MHZ have a lower elevated topography than the
2008); otherwise, shallow snow is present, or there is no UHZ (see Fig. 1).
snow-covered area (move to step 3).
2.6 Daily MODIS cloud-free snow cover day products
Tb10H − Tb36H > 0 K (3)
In the WH region, snow cover area (SCA) and snow cover
Tb10V − Tb36V > 0 K (4) pixels vary during different months of the year due to
changes in snowfall and snow ablation patterns. The lowest
– Step 3 – identify a shallow snow-covered area. If it satis- SCA has been observed during the month of August/Septem-
fies condition (Eq. 5), then shallow snow is present, and ber, and maximum SCA was observed during the month
a flag of 5.0 cm is set for the SD; otherwise, no snow is of February/March. Snow cover duration (SCD) depicts the
present. number of consecutive days snow cover is present for a
given pixel. It provides information regarding the persis-
Tb89V < 255 K, Tb89H < 265 K, Tb23V > Tb89V , tence of snowpack and is useful in improving PMW SD es-
Tb23H > Tb89H , and T < 267 K (5) timates (Singh et al., 2016; Wang et al., 2019; Dai et al.,
2018). In this study, the daily cloud-free MODIS snow cover
– Step 4 – estimation of moderate to deep SD using product (i.e., M*D10A1GL06) generated for high-mountain
Eq. (6). Asia (Muhammad and Thapa, 2020) at 500 m spatial reso-
  lution ([Link] Muham-
1 mad, 2020) has been used to generate the SCD product for
SD =
log10 (Tb36V − Tb36H ) X (Tb10V − Tb36V ) the study area during the data period. Previously, Sharma

1
 et al. (2014) and Singh et al. (2018) generated and eval-
+ (6) uated the SCD maps for the snow-covered Indian WH re-
log10 (Tb18V − Tb18H ) X (Tb10V − Tb18V )
gion. These studies (Sharma et al., 2014; Singh et al., 2018)
The developed SD algorithm was tested using World Me- revealed a higher average monthly SCD (>80 %) in high-
teorological Organization (WMO)-collected SD measure- altitude regions. These studies’ results further emphasize a
ments from 242 and 254 sites around the world during the strong longitudinal and altitudinal dependence on SCD, snow
2002–2003 and 2003–2004 winter season, respectively. In cover accumulation, and ablation in the WH region. There-
this only non-mountain stations with at least 30 d of mea- fore, SCD information can provide valuable insights to im-
sured snow were used in the comparison. In the recent prove the SD model. Daily binary snow cover maps prepared
study conducted over the mountainous terrain of the north- from M*D10A1GL06 are used to identify the snow cover
ern Xinjiang region, China, by Zhang et al. (2017), the presence for a given pixel. These binary snow cover maps are
AMSR2 SD products were compared with ground-collected used for computing the SCD information for each day from
SD data. They observed RMSEs of 18.5 cm (in AMSR2_A) 1 October of each year to 30 September of the following year
and 23.4 cm (in AMSR2_D) up to 30 cm for ground SD. during the study period. In this study, SCD of the WH region
However, AMSR2 SD products have not been evaluated for is retrieved only during the study period, i.e., from October
Indian Western Himalayan regions to date. to March for each year.

The Cryosphere, 18, 451–474, 2024 [Link]

D. K. Singh et al.: Himalayan snow depth mapping 457

2.7 MODIS land cover product aspect, and surface roughness, are derived from the resam-
pled DEM product. The SCD product is already available in
The heterogeneity in land cover significantly impacts the 500 m spatial resolution. Therefore, it is processed only to
amount of upwelling PMW radiation, affecting the TB at match the extent and coordinate reference system (i.e., GCS-
different frequencies for a given pixel. The effect of differ- WGS 1984) of other datasets. Following the resolution and
ent types of land cover in PMW SD retrievals has been in- coordinate system matching process, for all DGRE observa-
vestigated in many studies (Friedl et al., 2002; Yu et al., tory locations, the data from remote sensing products (i.e.,
2012; Wang et al., 2016, 2019). In this study, the MODIS TB, elevation, slope, ruggedness, geographical locations, and
Level 3 yearly land cover product (i.e., MCD12Q1) for SCD) are extracted for the winter period from 2012–2013 to
the year 2019 is downloaded from [Link] 2018–2019. It is known that the forest cover intercepts the
[Link]/ (last access: 23 January 2024) at 500 m spa- upwelling radiation from the ground underneath the snow-
tial resolution. MCD12Q1 product depicts land cover in 17 pack and causes uncertainty in the snow depth estimates of
classes as per the International Geosphere–Biosphere Pro- PMW SD models (Che et al., 2008). Therefore, the forest
gram (IGBP) system. These 17 classes are further regrouped cover fraction has been calculated using the MODIS land
into four categories, i.e., bare land, grass land, forest, and wa- cover type product (i.e., MCD12Q1) for a 10 km point buffer
ter, which account for ∼ 55.9 %, 27.4 %, 16.3 %, and 0.29 % around each observatory site. The retrieved values are used to
of the total WH area in 2019, respectively. The reclassified minimize the forest cover impact by dividing the brightness
land cover data have been used along with other datasets for temperature observations by the value of non-forest fraction
the development of multifactor SD models for different WH (i.e., 1 − forest fraction) for a given pixel as suggested by
regions. Foster et al. (1997).
In this study, the forenoon SD observations, descend-
ing pass AMSR2 TB data, terrain parameters (i.e., slope,
3 Methodology
aspect, ruggedness), geographical locations, and SCD are
Different steps followed for developing and validating the paired based on date and station location. These data are then
multifactor SD model(s) are given in the following subsec- checked for discrepancies such as missing values, incorrect
tions from 3.1 to 3.5. The general outline of the methodology values, and outliers. There are no missing values for AMSR2
adopted is shown in Fig. 2. TB, SRTM elevation, and SCD observations for the in situ
stations over the WH region. However, samples containing
3.1 Data preprocessing any other discrepancies are removed. After data preprocess-
ing, a total of ∼ 13 242 samples, with each sample compris-
Different remote sensing datasets comprising PMW TB ing geographical location, TB, terrain parameters, and SCD,
(from AMSR2), SRTM DEM, MODIS land cover product, are retained. Using these samples, the data for the 5-year
and MODIS SCD are used in the current study. These prod- snow period, i.e., from 2012–2013 to 2016–2017, are used
ucts are natively present in different spatial resolutions and to develop multifactor SD algorithms for different zones of
coordinate systems. Hence, all remote sensing datasets are the WH region. The remaining 2 years of data of snow pe-
processed using ArcGIS software to match the spatial extent, riod, i.e., from 2017–2018 to 2018–2019, is used to compare
coordinate system, and spatial resolution. The brightness and validate the multifactor SD model results.
temperature and SD datasets downloaded from the JAXA
portal use the Northern Hemisphere polar stereographic co- 3.2 Identification of dry snow pixel
ordinate system and are present in the HDF5 format. These
are reprojected to the WGS 1984 coordinate system and are Along with snow cover, frozen ground, rainfall, and cold
converted to TIFF format with help of the format conversion desert conditions affect the upwelling microwave emission
tool developed by JAXA. Following that ArcGIS software from the earth’s surface and impact PMW TB recorded by
is used for resampling the brightness temperature (BT) im- spaceborne sensors (Ferraro et al., 1996; Grody and Ba-
agery to 500 m. No additional processing is carried out in the sist, 1996). Further, wet snow pixels and surface waterbodies
current work as the brightness temperature dataset acquired cause PMW absorption and reduce volume scattering from
from JAXA is the level-3 product. The brightness tempera- snow grains (Stiles and Ulaby, 1980). Consequently, the in-
ture from each image for all stations is then retrieved pro- clusion of TB values from these pixels in the development
grammatically using Python. The extracted TB data are used and evaluation of the model results in large uncertainty in
for calculating the BTD. The BTD is calculated between SD estimates (Tedesco et al., 2015; Dietz et al., 2012; Fos-
lower- and higher-frequency TB observations for each day ter et al., 2005; Dong et al., 2005). Therefore, before devel-
during the study period. oping SD algorithms, dry snow pixels must be segregated
Following the BTD calculation, the SRTM DEM prod- from other pixels. Grody and Basist (1996) have developed a
uct is re-projected, mosaicked, and resampled to 500 m spa- decision tree to identify dry snow pixels from other scatter-
tial resolution. Different terrain parameters, such as slope, ing pixels using TB of different frequencies. Grody and Ba-

[Link] The Cryosphere, 18, 451–474, 2024

458 D. K. Singh et al.: Himalayan snow depth mapping

Figure 2. Flowchart representing the methodology.

sist’s (1996) decision tree makes use of different filters (see information (Kelly et al., 2003). Therefore, the inclusion of
Fig. 3) based on the values of TB observations to separate higher frequencies (i.e., 89 GHz) and lower frequencies (i.e.,
snow from non-snow pixels. This study uses multifrequency 10, 23 GHz) is investigated in many studies, which has re-
AMSR2 TB data with Grody and Basist’s (1996) decision sulted in improved SD estimates (Kelly et al., 2003; Wang
tree to identify snow pixels. et al., 2019; X. Xiao et al., 2020; Wei et al., 2021). Hence,
PMW TB values of 10, 18, 23, 36, and 89 GHz are used in
3.3 Selection of multifactor SD model parameters this study. Apart from single-channel SD, 40 combinations of
TB, i.e., BTD of different frequencies and polarizations, are
Many of the initial PMW SD models have relied on TB from also considered. Terrain parameters (i.e., elevation, slope, as-
18 and 36 GHz channels for estimating SD (Chang et al., pect, surface roughness), location (latitude, longitude), land
1987; Saraf et al., 1999; Das and Sarwade, 2008; Kelly et al., cover, and SCD also affect characteristics of snowpack and
2003; Chang et al., 1997). However, these models have lim- PMW TB (Saydi and Ding, 2020; Sharma et al., 2014; Wang
itations in estimating shallow and deep snowpacks. The sen- et al., 2010; Ansari et al., 2019). Thus, overall, 57 parameters
sitivity of PMW TB to SD decreases once the SD reaches a (i.e., TB – 10, BTD – 40, terrain parameters – 4, location –
threshold depth (Wang et al., 2019; Dai et al., 2018; Das and 2, SCD – 1) are considered in the process of SD model de-
Sarwade, 2008; Kelly et al., 2003). TB values of higher fre- velopment. However, of the 57 parameters, it is likely that
quencies (i.e., 36, 89 GHz) saturate before lower frequencies some parameters are redundant and do not necessarily add
(i.e., 10, 18 GHz) as SD increases. The lower frequency (i.e., any value to the model. For example, TB values of 10H
10 GHz) has the potential to retrieve deep snow cover, while and 10V have a correlation of 0.9, and using both TB10H
the higher frequency (i.e., 89 GHz) can provide shallow snow

The Cryosphere, 18, 451–474, 2024 [Link]

D. K. Singh et al.: Himalayan snow depth mapping 459

sample is selected, and other observation samples are used

to develop the regression model. The overall performance of
the model is calculated by combining all predictions (for the
omitted samples) from the n models. The accuracy of the
models is calculated using the correlation coefficient (R) and
RMSE. The results of LOOCV from the models developed
with all 57 parameters (see Sect. 4.2) are analyzed to select
the most valuable features for SD model development.

3.4 Development of SD models

This study implements four different regression models (i.e.,

linear, logarithmic, power, and reciprocal) to develop SD
models. Different WH zones, i.e., LHZ, MHZ, and UHZ,
have different topographic, environmental, and snowfall con-
ditions. Hence, in this study, SD models are developed sep-
arately for each WH zone. Data from 2012–2013 to 2016–
2017 are used for the development of the different SD mod-
els. Further, out of 57 parameters, 13 parameters are selected
from the results of the LOOCV evaluation. These 13 param-
eters have a good correlation with in situ SD and are used
in developing the multifactor SD models using four types of
regression. The general form of the four types of regression
models is given in Eqs. (7)–(10).

Figure 3. Identification of pure snow pixels from AMSR2 TB using

y = α1 x1 + α2 x2 + . . . + αi xi + c, (7)
the decision tree.
y = α1 ln x1 + α2 ln x2 + . . . + αi ln xi + c, (8)
y = cx1α1 x2α2 . . . xiαi , (9)
and TB10V is not useful and can cause additional problems 1 1 1
y = α1 + α2 + . . . + αi + c, (10)
due to multicollinearity. Further, the use of a large number x1 x2 xi
of independent variables leads to a curse of dimensionality,
which poses challenges in model development by decreasing
where y is the ground-observed SD values; x1 , x2 , . . . , and
the model’s interpretability and increasing the computational
xi are the screened parameters; α1 , α2 , . . . , and αi are the
time and resources, often leading to overfitting (Velliangiri
regression coefficients of the multiparameter models; c is the
et al., 2019; Obaid et al., 2019). These problems can be ad-
offset constant; and i represents the number of parameters.
dressed by performing optimal parameter selection for model
development (Chandrashekar and Sahin, 2014). Optimal pa-
rameter selection reduces the data dimensionality and elimi- 3.5 Validation of SD model(s)
nates irrelevant data from the original dataset.
This study considers data from the snow period between The multifactor SD models for different WH zones are vali-
2012–2013 and 2016–2017 of the entire WH region for opti- dated using temporally independent in situ SD observations
mal parameter selection. To select the necessary parameters during 2017–2018 and 2018–2019. The accuracy of SD mod-
for the SD model, all 57 parameters are used independently els’ estimates is evaluated using standard regression metrics,
with in situ SD to develop single-parameter linear regression i.e., R and RMSE. Additionally, the efficacy of the proposed
models. While developing these models, evaluation is car- multifactor SD models is analyzed by comparing the accu-
ried out using the leave-one-out cross-validation (LOOCV) racy of the multifactor model with regional (Das and Sar-
method (Webb et al., 2011) for screening necessary param- wade, 2008; Singh et al., 2020) and heritage (Chang et al.,
eters. The LOOCV method is widely used by various re- 1987) SD models for different ranges of the WH region. The
searchers (Gusain et al., 2016; Joshi et al., 2017; Wang et SD models of Chang et al. (1987), Das and Sarwade (2008),
al., 2019) to conduct the validation of models and assess and Singh et al. (2020) are given in Eqs. (11), (12), and (13),
the model’s accuracy. In LOOCV, the observational dataset respectively.
is used to create n number of regression models (n is num- The comparison is carried out by estimating SD from all
ber of samples). In each of the n models, a different testing these stated models using the validation data present between

[Link] The Cryosphere, 18, 451–474, 2024

460 D. K. Singh et al.: Himalayan snow depth mapping

2017–2018 and 2018–2019.

SDChang et al. (1987) = 1.59 × (TB18H − TB36H), (11)

SDDas and Sarwade (2008) = 3.16 × (TB18H − TB36H) + 24.25, (12)
SDSingh et al. (2020) = −7.58 × (TB18V − TB36V) + 233.71, (13)

where TB denotes the brightness temperature values, 18 and

36 indicate the frequency of TB (in GHz), and V and H are
the vertical and horizontal polarization, respectively.
Apart from the aforementioned comparative analysis, a
random sample image from the study area for a single day
(3 February 2019) is taken. Then, the estimated SD over the
selected area using the multifactor SD model(s) is spatially
compared with AMSR2 operational products (see Sect. 4.5).
This spatial comparison helps in understanding how the de-
veloped multifactor SD model(s) differs from the AMSR2 Figure 4. Spatial distribution of mean SD in WH zones along
operational SD products in representing SD information over DGRE stations during 2012–2019 (October to March months)
(Note: J&K is Jammu and Kashmir, HP is Himachal Pradesh).
the WH region. The magnitude of in situ SD, terrain param-
eters, and SCD can significantly affect the accuracy of the
PMW SD model in the study region. Therefore, the accu- observed that out of nine stations that have a mean SD greater
racy of operational AMSR2 SD products and multifactor SD than 150 cm, five are present in the UHZ.
models with respect to varying ground SD, topographic el- The overall analysis of in situ SD measurements indicates
evation, and SCD is determined in different WH zones (see the mean and standard deviation (µ ± σ ) are observed as
Sect. 4.6). ∼ 121.5 cm ± 122.5 cm in the LHZ, ∼ 85.9 cm ± 83.5 cm in
the MHZ, and 176.6 cm ± 208.9 cm in the UHZ. A higher
4 Results and analysis mean SD is observed in the UHZ compared to the other two
ranges. A total of 95 % of overall SD values in the LHZ are
The insights from the analysis of in situ SD observations below 350 cm, with the remaining 5 % having SD between
in WH zones are reported in Sect. 4.1. Following that, the 350 and 650 cm. However, in the MHZ and UHZ, 95 % of to-
results from the LOOCV evaluation of multiple parameters tal SD observations are below 200 and 500 cm, respectively,
are given in Sect. 4.2. The outcomes from the accuracy as- with the remaining 5 % ranging between 200 and 500 cm and
sessment and comparison of different PMW SD estimates 500 and 2030 cm.
are described in Sect. 4.3 and 4.4, respectively. The spatial
comparison of the high-resolution SD map from the multi- 4.2 SD parameters screening and evaluation over the
factor model and AMSR2 products is shown in Sect. 4.5. In WH region
Sect. 4.6, the analysis of multifactor SD model performance
with respect to different parameters is detailed. The expressions for linear regression models developed with
each parameter and regression metric, i.e., RMSE and R re-
4.1 Spatial analysis of the in situ SD observations in sults obtained from LOOCV analysis, are shown in Table 3.
WH ranges In terms of geographical location, latitude has a higher cor-
relation (i.e., 0.24) and lower RMSE (97.96 cm) than lon-
The mean of in situ SD at each of the 43 DGRE stations gitude. Among the terrain parameters, SCD has the highest
is estimated for the winter period (October to March) dur- R (i.e., 0.45) and the lowest RMSE (i.e., 90.27 cm) and is
ing 2012–2013 to 2018–2019 (see Fig. 4). The results in- followed by elevation (R = 0.30 and RMSE = 96.12), slope
dicate that during the data period, the mean SD values var- (R = 0.26 and RMSE = 97.59), and ruggedness (R = 0.25
ied between ∼ 11 cm (elevation: 1664 m) and ∼ 256 cm (el- and RMSE = 97.85), making it highly important for the de-
evation: 3160 m) in the LHZ, ∼ 21 cm (elevation: 3250 m) velopment of multifactor SD models. The SD models built
and ∼ 136 cm (elevation: 4950 m) in the MHZ, and ∼ 49 cm with TB observations from descending orbital passes have
(elevation: 3250 m) and ∼ 365 cm (elevation: 5995 m) in the a relatively higher correlation and lower RMSE than those
UHZ. The analysis also demonstrates that out of 43 manual from ascending pass TB data when analyzed with in situ SD.
stations, 4 stations have a mean SD between 11 and 50 cm, This is mainly because descending orbital passes occur in the
18 stations have a mean SD between 50 and 100 cm, 7 sta- morning time with no melting of snow; however, ascending
tions have a mean SD between 100 and 150 cm, 5 stations orbital passes occur in the afternoon time with substantial
have mean SD between 150 and 200 cm, and the remaining 4 melting of snow in the study area. Therefore, only descend-
have mean SD >200 cm during the data period. Further, it is ing pass TB observations are used in the study.

The Cryosphere, 18, 451–474, 2024 [Link]

D. K. Singh et al.: Himalayan snow depth mapping 461

Table 3. Results of LOOCV evaluation for SD models developed using single parameters.

Parameters Independent variable (x) Linear regression model RMSE (cm) R

Geographical location Latitude y = 32x − 1046.81 97.96 0.24
Longitude y = 21x − 1573.72 99.30 0.17
Terrain Elevation y = 0.029x − 45.81 96.12 0.30
Aspect y = −0.16x + 109.58 99.75 0.14
Slope y = −3.42x + 119.67 97.59 0.26
Ruggedness y = −0.31x + 128.15 97.85 0.25
Cloud-free snow SCD y = 1.14x − 19.58 90.27 0.45
product
Brightness 10H y = −4.5x + 1210.78 106.28 0.34
temperature 10V y = −5.77x + 1553.53 105.58 0.36
(ascending data) 18H y = −3.9x + 1047.89 104.59 0.39
18V y = −4.8x + 1293.91 103.74 0.40
23H y = −3.59x + 963.28 105.22 0.37
23V y = −4.13x + 1109.64 104.65 0.38
36H y = −2.1x + 587.17 108.81 0.28
36V y = −2.22x + 620.55 109.08 0.27
89H y = −0.24 + 161.87 113.29 0.03
89V y = −0.08x − 90.99 113.35 0.01
Brightness 10H y = −3.2x + 858.65 94.28 0.35
temperature 10V y = −4.01x + 1071.98 93.96 0.36
(descending data) 18H y = −2.76x + 741.75 93.96 0.36
18V y = −3.4x + 910.08 93.44 0.38
23H y = −2.67x + 706.15 94.05 0.36
23V y = −3.05x + 811.49 93.74 0.37
36H y = −1.78x + 480.19 96.38 0.29
36V y = −1.9x + 522.28 96.42 0.29
89H y = −0.58x + 207.88 100.29 0.10
89V y = −0.45 + 181.84 100.54 0.07

y = SD (cm).

Apart from the single-channel PMW TB, 40 different com- GHz), and H and V represent horizontal and vertical po-
binations of descending pass orbital BTDs are tested using larization, respectively), three terrain parameters (elevation,
linear regression (in the LOOCV approach). The selected slope, ruggedness), latitude, and SCD are used to develop
and rejected BTD parameters with their RMSE and R with multifactor PMW SD models for the three WH ranges.
the in situ SD are shown in Table 4. The selected parame-
ters have the lowest RMSE and highest correlation, and they 4.3 Evaluation and comparison of different multifactor
pass the F test at a significance level of 0.001. It is observed SD models in WH zones
that descending pass BTD models exhibit higher correlation
and accuracy metrics compared to single-channel descend-
The details of the developed models are given in Table 5. The
ing pass models. From the overall results (R, RMSE), de-
results from the regression analysis indicate multifactor SD
scending pass BTD parameter-based SD models have higher
models developed with a power regression approach have a
R (0.24 to 0.39) and lower RMSE (91.63 to 93.92 cm) com-
better fit with the in situ data and outperformed other regres-
pared to single-channel TB-based SD models, which have
sion models with R (RMSE in cm) values of 0.62 (49.17),
R values of 0.07 to 0.35 and RMSE values of 93.44 to
0.78 (37.72), and 0.76 (55.12) in the LHZ, MHZ, and UHZ,
100.54 cm. Therefore, different BTD data from descending
respectively.
passes (see Table 4) are selected instead of single-channel
The developed multifactor models in each WH zone are
TB to develop multifactor SD models. Along with the eight
evaluated (with R and RMSE metrics) using temporally inde-
descending BTD parameters (BTD of 10H18H, 10H23H,
pendent data from 2017–2018 to 2018–2019. Comparison of
18H89V, 36H89V, 36V89V, 23H89V, 10V89V, and 10V23H,
the four types of multifactor regression models in the LHZ,
where 10, 18, 23, 36, and 89 represent the frequency (in
MHZ, and UHZ is carried out with the help of the Taylor

[Link] The Cryosphere, 18, 451–474, 2024

462 D. K. Singh et al.: Himalayan snow depth mapping

Table 4. BTD SD model (with descending observations) relation with SD and evaluation using LOOCV method.

Serial no. Independent variable (x) Linear regression model RMSE (in cm) R
Selected parameters 1 BTD (36H–89V) y = −2.24x + 107.05 91.63 0.39
2 BTD (36V–89V) y = −2.16x + 81 92.24 0.37
3 BTD (10V–23H) y = 4.12x + 31.05 92.45 0.35
4 BTD (23H–89V) y = −1.78x + 122.17 92.46 0.36
5 BTD (10V–18V) y = 7.43x + 52 92.58 0.25
6 BTD (10H–23H) y = 4.12x + 56 93.78 0.20
7 BTD (10H–18H) y = 5.66x + 58 93.47 0.21
8 BTD (18H–89V) y = −1.61x + 122.34 93.92 0.24
Rejected parameters 9 BTD (10H–36H) y = 0.85x + 70.11 102.20 0.17
10 BTD (10H–89H) y = −0.91x + 114.15 102.16 0.18
11 BTD (10H–18V) y = 2.84x + 89.85 102.20 0.17
12 BTD (10H–23V) y = 3.29x + 78.28 102.15 0.18
13 BTD (10H–36V) y = 0.55x + 77.67 102.21 0.16
14 BTD (10H–89V) y = −1.15x + 177.36 102.14 0.19
15 BTD (10V–18H) y = 5.10x + 30.11 102.04 0.20
16 BTD (10V–36H) y = 1.21x + 56.24 102.17 0.18
17 BTD (10V–89H) y = −0.66x + 110.16 102.19 0.17
18 BTD (10V-23V) y = 4.93x + 44.79 102.03 0.20
19 BTD (10V–36V) y = 1.08x + 63.64 102.18 0.17
20 BTD (10V–89V) y = −0.92x + 116.53 102.17 0.18
21 BTD (18H–23H) y = 3.18x + 77.95 102.19 0.17
22 BTD (18H–36H) y = 0.18x + 83 102.23 0.16
23 BTD (18H–89H) y = −1.4x + 122.75 102.09 0.20
24 BTD (18H–23V) y = −3.92x + 75.13 102.26 0.17
25 BTD (18H–36V) y = −0.51x + 90.15 102.24 0.16
26 BTD (18V–23H) y = 6.1x + 32.36 102.08 0.20
27 BTD (18V–36H) y = 0.86x + 68.45 102.21 0.17
28 BTD (18V–89H) y = −1.14x + 122.94 102.13 0.19
29 BTD (18V–23V) y = 6.35x + 61.65 102.13 0.18
30 BTD (18V–36V) y = 0.43x + 78.64 102.22 0.16
31 BTD (18V–89V) y = −1.4x + 126.53 102.10 0.20
32 BTD (23H–36H) y = −0.09x + 86.33 102.23 0.16
33 BTD (23H–89H) y = −1.57x + 123.73 102.06 0.20
34 BTD (23H–36V) y = −1.16x + 93.49 102.23 0.17
35 BTD (23V–36H) y = 0.68x + 74.51 102.22 0.16
36 BTD (23V–89H) y = −1.36x + 125.26 102.09 0.20
37 BTD (23V–36V) y = −0.07x + 86.26 102.23 0.16
38 BTD (23V–89V) y = −1.62x + 126.92 102.06 0.20
39 BTD (36H–89H) y = −2.10x + 113.67 102.01 0.20
40 BTD (36V–89H) y = −1.91x + 118.52 102.03 0.19

Note: y = SD (cm).

diagram (see Fig. 5). The R and RMSE (in cm) for the met- 4.4 Comparative analysis of multifactor and other SD
rics of power, linear, logarithmic, and reciprocal models in models in different zones of the WH region
different WH zones are given in Fig. 5 and Table 6. The re-
sults from the comparison indicate that in all WH zones, the In order to compare the performance of different SD models
multifactor SD model developed using power regression has in the WH region, the SD values are estimated using differ-
exhibited higher accuracy, i.e., better correlation and lower ent models with the help of PMW and other auxiliary data
RMSE compared to the models built using linear, logarith- during the study period (i.e., 2017–2018 to 2018–2019). Dif-
mic, and reciprocal regression approaches. Therefore, in each ferent models used in the comparative analysis are the mul-
WH zone, the multifactor SD model from power regression tifactor SD model(s) from this study, regional SD models of
is used to estimate PMW SD at 500 m spatial resolution. the WH region (i.e., Das et al., 2008; Singh et al., 2020), and
the heritage SD model provided by Chang et al. (1987). The

The Cryosphere, 18, 451–474, 2024 [Link]

D. K. Singh et al.: Himalayan snow depth mapping 463

Table 5. Multifactor SD model regression coefficient for WH zones during 2012–2017 (October to March).

WH zones Model type Models R (RMSE)

Lower Himalayan Zone Linear y = −51.16−0.09x1 +0.08x2 +0.94x3 −0.24x4 +1.15x5 −0.37x6 +5.21x7 −1.6x8 −0.40x9 +1.14x10 + 0.59 (64.14)
0.001x11 − 0.45x12 − 4.07x13
Logarithmic y = −528.08 − 733.37 ln(x1 ) + 678.83 ln(x2 ) + 8.80 ln(x3 ) − 230.65 ln(x4 ) + 84.84 ln(x5 ) + 0.45 (81.12)
19.32 ln(x6 )−0.82 ln(x7 )+30.21 ln(x8 )+22.66 ln(x9 )−11.51 ln(x10 )−132.43 ln(x11 )+12.23 ln(x12 )−
38.37 ln(x13 )

Power y = 4.49 × 10−26 x110.12 x24.93 x31.28 x4−2.29 x50.59 x6−0.25 x70.56 x8−2.79 x90.04 x10
0.22 x 1.70 x 0.001 x −0.68
11 12 13 0.62 (49.17)
Reciprocal y = 487.86−5073.85/x1 −835884.29/x2 −619.77/x3 +45566.55/x4 −226.44/x5 +1.04/x6 +2/x7 + 0.49 (78.11)
24.12/x8 + 0.19/x9 − 1.13/x10 + 69.02/x11 − 0.37x12 − 75.87/x13
Middle Himalayan Zone Linear y = 1285.89 − 34.66x1 + 0.001x2 + 1.57x3 − 0.05x4 + 1.50x5 − 4.10x6 + 7.73x7 − 1.77x8 + 2.90x9 − 0.69 (42.04)
1.39x10 + 0.001x11 + 4.70x12 − 8.84x13
Logarithmic y = 2281.76−1364.46 ln(x1 )+0.24 ln(x2 )+27.81 ln(x3 )−35.38 ln(x4 )+120.18 ln(x5 )+4.53 ln(x6 )+ 0.62 (51.11)
23.03 ln(x7 )−49.38 ln(x8 )+15.60 ln(x9 )+26.05 ln(x10 )−83.25 ln(x11 )+28.42 ln(x12 )−124.63 ln(x13 )

Power y = 3.7 × 1013 x1−7.72 x20.03 x30.03 x40.07 x51.09 x6−0.11 x70.57 x8−0.01 x90.05 x10
−0.12 −0.73 0.19 −1.51
x11 x12 x13 0.78 (37.72)
Reciprocal y = −26.58+4283.25/x1 −2060.51/x2 +20.08/x3 −6150.16/x4 −268.45/x5 −0.04/x6 −0.35/x7 + 0.66 (45.72)
23.07/x8 − 0.19/x9 − 0.03/x10 + 2.63/x11 − 0.16/x12 + 5.72/x13
Upper Himalayan Zone Linear y = −3754.98+99.59x1 +0.07x2 −1.25x3 +0.06x4 +1.23x5 +10.46x6 +0.001x7 +4.90x8 −17.63x9 + 0.74 (58.07)
19.77x10 − 6.58x11 + 6.06x12 − 16.11x13
Logarithmic y = −8280.01 + 4430.22 ln(x1 ) + 378.71 ln(x2 ) − 102.97 ln(x3 ) + 68.20 ln(x4 ) + 116.51 ln(x5 ) + 0.68 (69.08)
31.14 ln(x6 )−6.29 ln(x7 )−60.95 ln(x8 )−16.924 ln(x9 )+42.65 ln(x10 )−9.88 ln(x11 )+15.42 ln(x12 )−
74.45 ln(x13 )

Power y = 6.4 × 10−53 x126.26 x22.91 x3−0.20 x40.62 x50.64 x60.11 x7−0.18 x8−0.31 x9−0.13 x10
0.23 x 0.10 x 0.001 x 0.44
11 12 13 0.76 (55.12)
Reciprocal y = 5308.54 − 170 987.89/x1 − 1 378 163.61/x2 + 121.52/x3 − 5412.19/x4 − 197.55/x5 + 0.08/x6 + 0.33 (59.61)
0.51/x7 − 0.45/x8 − 0.60/x9 + 1.15/x10 − 0.26/x11 − 0.09/x12 − 63.24/x13
Note: in Table 5, x1 to x5 are latitude, elevation, slope, ruggedness, and SCD, respectively; x6 to x13 are the BTD of 10H18H, 10H23H, 18H89V, 36H89V, 36V89V, 23H89V, 10V89V, and 10V23H, respectively; V is the
vertical polarization; H is the horizontal polarization; and 10, 18, 23, 36, and 89 are the frequency (in GHz) of the corresponding BT channels.

Table 6. Comparative analysis of multifactor SD models during 2017–2019 for WH zones.

Western Himalayan zones

Lower Himalayan Middle Himalayan Upper Himalayan
Models R RMSE (in cm) R RMSE (in cm) R RMSE (in cm)
Power 0.65 22.7 0.76 19.2 0.89 22.6
Linear 0.64 29 0.68 22.8 0.75 33.5
Logarithmic 0.38 52 0.14 41 0.73 36.9
Reciprocal 0.09 121.3 0.47 26.7 0.61 43.2

estimated SD from each model is compared with in situ SD in situ SD with an R value of 0.65 (see Fig. 6). The RMSE is
observations within the respective WH zones to understand observed to be 36.32, 49.82, and 119.79 cm for the Chang
the accuracy of SD retrievals. Singh et al.’s (2020) model et al. (1987) model, Das et al. (2008) model, and Singh
is proposed only for the MHZ. Therefore, it is not used for et al. (2020) model, respectively. The proposed multifactor
SD estimation in the LHZ and UHZ when doing comparative SD model has shown good accuracy with a lower RMSE of
analysis. 27.21 cm compared to other SD models. The SD model pro-
In the LHZ, both Chang’s model and Das’s model have posed by Singh et al. (2020) for the MHZ is developed using
poor correlation with in situ SD and have shown RMSEs (R) data from a single observatory location. Hence, the Singh et
of 39.51 (−0.16) and 49.66 (−0.14) (see Fig. 6), whereas al. (2020) model cannot represent the spatial variability in
the proposed multifactor SD model has shown a good corre- SD and shows significant errors with higher bias.
lation with RMSE (R) of 32.87 (0.75). In the MHZ, Chang Similar to the results observed in the LHZ and MHZ, the
et al. (1987), Das et al. (2008), and Singh et al. (2020) have Chang et al. (1987) model and Das et al. (2008) model have
exhibited poor correlation with in situ SD with R values of shown a poor correlation in the UHZ with R values of 0.18
0.22, 0.21, and −0.22, respectively, whereas the proposed and 0.19, respectively. In comparison, the multifactor SD
multifactor SD model has shown a good correlation with model has shown a good correlation with an R value of 0.67

[Link] The Cryosphere, 18, 451–474, 2024

464 D. K. Singh et al.: Himalayan snow depth mapping

ascending SD product, i.e., AMSR2_A (see Fig. 7b), and de-

scending SD product, i.e., AMSR2_D (see Fig. 7c), of the
same region for the given day at 10 km resolution are also
prepared.
According to MODIS-derived SCA at 500 m resolution
(see Fig. 7a) and DGRE observatories’ in situ SD informa-
tion, this is snow cover with varying thickness on 3 Febru-
ary 2019 in WH zones. However, in both the AMSR2 SD
products, i.e., AMSR2_A and AMSR2_D, the majority of
pixels have zero SD value, resulting in the underestimation
of SD information by AMSR2 products. The maximum SD
values observed in different products are as follows: AMSR2
ascending SD, 58 cm; AMSR2 descending SD, 78 cm; mul-
tifactor SD model, 476 cm. The multifactor SD model shows
high heterogeneity in SD across the selected region in WH
zones compared to AMSR2 SD products. Further, the mul-
tifactor SD model offers good detail about snow cover and
provides SD data in the region at a high resolution of 500 m.

4.6 Comparison of performance of multifactor SD

product with operational AMSR2 SD product
Figure 5. Taylor diagram for the evaluation of multifactor SD mod-
els during 2017–2019 for (a) the LHZ, (b) the MHZ, and (c) the
Though regions with higher mean SD (i.e., LHZ, UHZ) have
UHZ.
a higher error than regions with a lower mean SD (i.e.,
MHZ), it is important to assess how the SD products’ ac-
(see Fig. 6). The RMSE values are observed to be 60.95, curacy varies with changes in in situ SD. Hence, Sect. 4.6.1
51.74, and 42.81 cm for the Chang et al. (1987) model, Das analyzes the operational AMSR2 SD products and AMSR2
et al. (2008) model, and proposed multifactor SD model in multifactor model performances in different SD classes. Fur-
the UHZ. Overall results from the comparative analysis indi- ther, it is also essential to understand how the model’s accu-
cate that in each WH zone, the multifactor SD model has racy is affected with respect to different auxiliary parameters,
higher accuracy with good correlation (i.e., R) and lower i.e., topographical and land cover parameters. Therefore, the
errors when compared with other models. Further, the de- model accuracy is evaluated with respect to different topo-
veloped model has exhibited better accuracy metrics in the graphical and land cover parameters, and the results are pre-
MHZ compared to other WH zones. The mean SDs observed sented in Sect. 4.6.2.
during the study period (i.e., 2017–2018 to 2018–2019) in
Pir Panjal and UHZ are higher than the mean SD of the MHZ. 4.6.1 Analysis of operational AMSR2 SD products and
Further, the LHZ has a forest canopy which can affect the multifactor SD model in different SD classes
PMW TB observations, whereas, in the MHZ, most of the
region is devoid of forest vegetation except for some patchy In each WH zone, the AMSR2 SD products and multifactor
grass vegetation. Hence it is expected that an increased error SD model estimates are grouped into five SD classes, i.e., 0–
for SD models in the LHZ and UHZ compared to the MHZ 25, 25–50, 50–75, 75–100, and >100 cm based on in situ SD
will be observed. observations during 2017–2018 to 2018–2019. Along with
the multifactor SD, the operational AMSR2 SD product (i.e.,
4.5 Spatial comparison of SD from multifactor model from both ascending and descending pass data) is also ana-
and operational AMSR2 SD products: a case study lyzed in the SD classes by comparing it with in situ SD ob-
servations. RMSE of each SD class is calculated to evaluate
The spatial comparison of SD maps from AMSR2 SD prod- the accuracy of SD estimates. Other models (i.e., Chang et
ucts and the multifactor SD model is performed to under- al., 1987; Das et al., 2008; and Singh et al., 2020) were not
stand the improvement in the AMSR2 multifactor SD model considered in this analysis as they were not operational SD
over the operational AMSR2 SD products in the WH region. models. The effect of variation in ground SD on the accu-
For this purpose, the SD maps of operational AMSR2 SD racy of the AMSR2 multifactor SD model and AMSR2 SD
products and the multifactor SD model for WH zones for products is shown in Table 7.
3 February 2019 are considered (see Fig. 7). The SD spa- The results indicate that the magnitude of RMSE of
tial map at 500 m resolution for the WH zones is generated AMSR2 SD products and the multifactor SD model in-
using the multifactor SD model (see Fig. 7d). The AMSR2 creased with an increase in SD. When in situ SD <25 cm,

The Cryosphere, 18, 451–474, 2024 [Link]

D. K. Singh et al.: Himalayan snow depth mapping 465

Figure 6. Comparison of model-estimated and field-observed SD values in the LHZ (Pir Panjal), MHZ (Greater Himalayan), and UHZ
(Karakoram) during 2012–2019 (October to March).

Figure 7. Spatial map of SD variation on 3 February 2019. (a) MODIS SCA, (b) AMSR2_A SD product map at 10 km, (c) AMSR2_D SD
product map at 10 km, and (d) multifactor SD model map at 500 m.

AMSR2 SD products have shown relatively lower error in of 43 stations, only four stations have a mean SD <25 cm,
all three zones compared to the developed multifactor SD and for the remaining stations, the mean in situ SD was more
model. However, the observed error in this class (i.e., 0– than 25 cm during the study period. Hence AMSR2 SD prod-
25 cm) is still large and varies between 11–15 cm in the ucts are less useful for spatial monitoring of SD in the WH
AMSR2 SD product and 14–27 cm in the multifactor SD region. Though the developed AMSR2 multifactor model has
model across the three zones, whereas for classes with in situ shown higher error when in situ SD <25 cm, it is more useful
SD >25 cm, the proposed multifactor SD model has a lower for the WH region as the RMSE is lower when SD >25 cm.
RMSE than both AMSR2 products in all the zones of the WH Overall, the multifactor SD model in the MHZ has a low
region. This analysis clearly shows that for shallow snow re- RMSE compared to the LHZ and UHZ in all SD classes. This
gions in the WH region, operational AMSR2 products can be could be due to prevailing dry snow conditions, lower mean
used. However, the AMSR2 SD products show a large error SD, and the absence of forest in this range, which can favor
for deep and moderate snow regions. In the WH region, out the PMW SD algorithms to retrieve better SD estimates from

[Link] The Cryosphere, 18, 451–474, 2024

466 D. K. Singh et al.: Himalayan snow depth mapping

Table 7. RMSE of operational AMSR2 SD products and the proposed multifactor SD model across different WH zones in different in situ
SD classes.

WH zones Model Snow depth class (in cm)

0–25 25–50 50–75 75–100 >100
Lower Himalayan Zone AMSR2_A 14.60 38.88 62.31 88.63 159.16
AMSR2_D 14.35 35.68 58.65 85.75 152.84
Multifactor SD model 27.64 21.62 37.27 40.48 63.73
Middle Himalayan Zone AMSR2_A 13.67 35.97 61.39 86.34 200.25
AMSR2_D 18.00 33.45 58.59 82.77 193.76
Multifactor SD model 20.99 20.29 27.84 41.95 81.04
Upper Himalayan Zone AMSR2_A 13.23 38.13 61.58 87.84 375.61
AMSR2_D 14.34 36.93 60.19 86.69 372.67
Multifactor SD model 37.12 41.54 38.62 40.12 161.01

PMW TB, whereas higher temperatures, moist snow condi- tion classes for the multifactor SD model and AMSR2 as-
tions, forest vegetation in the LHZ, and deep snow conditions cending and AMSR2 descending SD products in different
in both the LHZ and UHZ can negatively affect the accuracy WH regions is as follows: LHZ, 21.38–47.27, 23.05–113.21,
of the AMSR2 multifactor SD model in these regions by af- and 18.44–93.72 cm; MHZ, 17.82–54.79, 39.10–107.72, and
fecting the TB observations. 37.98–103.38 cm; UHZ, 11.73–126.13, 17.71–188.67, and
19.50–182.12 cm. Though overall RMSE variation is high
4.6.2 Multifactor model performance analysis with across the different elevation classes, both AMSR2 SD prod-
respect to auxiliary parameters ucts have similar RMSEs for any given elevation class within
each WH zone. However, the multifactor SD model has
Among all the factors considered in the AMSR2 multifactor lower RMSE than both AMSR2 SD products for elevation
SD model development, elevation and SCD have good het- classes across the three WH zones. Other than elevation, the
erogeneity across the stations in each WH zone. The other amount of snowfall and snow conditions vary widely with
terrain factors, such as slope and land cover, do not have SCD across the different WH zones. This can lead to varying
much variation and are similar for many of the stations within accuracy trends in SD retrievals for a given factor in different
a WH zone (see Fig. 8). Though a large variation in land WH zones.
cover is observed across the entire WH region, in the LHZ the The SD generally increases with an increase in SCD, af-
majority of stations are surrounded by forest cover, whereas fecting the PMW SD retrieval from different models. Across
in the MHZ, stations are mainly over grassland and barren all WH regions, the RMSE values of the AMSR2 ascend-
land. The UHZ is devoid of vegetation, and all stations are ing and AMSR2 descending SD products and the multifactor
present over barren land and glaciers. Thus, the variation in SD model increased with an increase in SCD. However, the
land cover within a range is not significant. Therefore, in this RMSE of the multifactor SD model is significantly low com-
section, only the effect of varying elevation and SCD on the pared to the AMSR2 SD products in all SCD classes in each
accuracy of SD from the AMSR2 multifactor SD model and WH zone. The SCD variation at the end of the snow year
operational AMS2 SD products is evaluated. (30 September 2013), along with RMSEs in different sta-
Slope and SCD are divided into different classes consid- tions calculated for the time period, i.e., 2017–2018 to 2018–
ering the overall variation in the WH region. Within these 2019, is represented in Fig. 8d. The RMSE variation associ-
classes, the SD retrievals from the AMSR2 products and ated with SCD classes for the multifactor SD model, AMSR2
multifactor model are compared with in situ SD measure- ascending product, and AMSR2 descending product in dif-
ments during the winter period between 2017–2018 and ferent WH ranges are as follows: LHZ, 25.18–61.80, 47.32–
2018–2019. The accuracy of model estimates in each class 158.11, and 45.47–140.59 cm; MHZ, 19.71–70.92, 33.53–
is evaluated by calculating RMSE (see Table 8). The RMSE 140.03, and 31.61–137.66 cm; UHZ, 83.40–122.22, 97.63–
associated with each station for different factors (i.e., ele- 205.89, and 92.54–204.15 cm.
vation, slope, SCD, and land cover) is depicted in Fig. 8.
The results indicate that in the LHZ and UHZ, with an in-
crease in elevation, RMSE increased for both AMSR2 prod- 5 Discussion
ucts and the multifactor model, whereas in the MHZ, there
is no specific trend in the variation in accuracy with respect The Indian WH region has the highest mountain peaks in
to elevation. The RMSE (in cm) variation across all eleva- Asia that separate the plane regions of the Indian subcon-

The Cryosphere, 18, 451–474, 2024 [Link]

 Table 8. RMSE (cm) variation in AMSR2 SD products and the multifactor SD model for different elevations, as well as SCD classes across three WH zones (for snow period during
2017–2018 to 2018–2019).

Parameter Elevation (in m) SCD (in days)

[Link]
Parameter range 2000–2500 2500–3000 3000–3500 3500–4000 4000–4500 4500–5000 >5000 0–30 30–60 60–90 >90
Lower Himalayan Zone Mean in situ SD 18.44 63.09 93.72 – – – – 37.32 84.33 96.95 140.38
AMSR2_A 23.05 76.97 113.21 – – – – 47.32 101.59 114.23 158.11
AMSR2_D 26.51 78.43 107.99 – – – – 45.47 95.71 111.62 140.59
D. K. Singh et al.: Himalayan snow depth mapping

Multifactor model 21.38 41.13 47.27 – – – – 25.18 42.12 54.22 61.8

Middle Himalayan Zone In situ SD 80.53 59.96 31.8 69.42 83.02 80.67 66.49 22.49 53.39 83.96 127.9
AMSR2_A 85.04 86.84 39.1 98.53 107.01 107.72 84.72 33.53 66.73 99.12 140.03
AMSR2_D 79.69 88.13 37.98 88.32 103.38 101.86 82.48 31.61 65.06 98.25 137.66
Multifactor model 54.79 45.46 17.82 32.54 54.67 52.41 37.48 19.71 35.67 49.63 70.92
Parameter Elevation (in m) SCD (in days)
Parameter range 2000–2500 2500–3000 3000–3500 3500–4000 4000–4500 4500–5000 >5000 0–30 30–60 60–90 >90
Upper Himalayan Zone In situ SD – – – 15.38 67.8 80.33 130.94 52.79 78.2 108.24 167.86
AMSR2_A – – – 17.71 77.6 104.97 188.67 97.63 130.2 157.39 205.89
AMSR2_D – – – 19.5 69.38 104.42 182.12 92.54 123.51 158.46 204.15
Multifactor model – – – 11.73 34.22 55.45 126.13 83.4 89.9 108.71 122.22

The Cryosphere, 18, 451–474, 2024

467
468 D. K. Singh et al.: Himalayan snow depth mapping

Figure 8. Spatial distribution of RMSE of the multifactor SD model for varying (a) elevations, (b) slopes, (c) land cover types, and (d) SCD
among the 43 ground stations.

tinent from the Tibetan Plateau with a mean elevation of MHZ, and UHZ for the proposed model is 4.5, 2, and 6.3 cm,
∼ 3116 m a.m.s.l. The discussion about the performance of respectively, whereas the bias for the legacy model and other
SD models and factors affecting the multifactor SD model is regional models is considerably higher with significant over-
given in Sect. 5.1 and 5.2, respectively. estimates in the lower depth values and underestimates in
higher-depth regions. Further, it must be emphasized that
5.1 SD models’ performance these models have very poor correlation with the in situ snow
depth and the SD estimates mainly confined in a range irre-
Four multifactor SD models are developed for each WH spective of the magnitude of the ground snow depth observa-
zone using different regression approaches (i.e., linear, log- tion values.
arithmic, reciprocal, and power). These models are com- In general, with an increase in SD, the accuracy of mul-
pared with regional SD models, Chang’s SD model, and op- tifactor models declined in all WH zones. However, the ac-
erational AMSR2 SD products. The overall analysis of the curacy of developed multifactor SD models is distinct for a
results indicates that the power-regression-based multifac- given SD class in different WH zones. This is because the
tor SD model has higher accuracy compared to other mul- spatial distribution of snowfall and snow characteristics (i.e.,
tifactor SD models, regional approaches, Chang’s model, SD, snow wetness, density) are not uniform at different geo-
and AMSR2 SD products in all WH regions. However, graphic locations of observatories distributed across the three
AMSR2 SD products have shown comparable to better ac- WH zones. The SD model developed in the MHZ has shown
curacy (i.e., similar to the multifactor SD model) under shal- better accuracy metrics than those developed for the LHZ and
low snow conditions (SD <25 cm). Nevertheless, once SD UHZ.
exceeds 25 cm, the performance of AMSR2 SD products de- The topographical parameters in the WH region play a vi-
clined considerably (see Table 8). Further, AMSR2 SD prod- tal role in affecting the local climate as well as snow dis-
ucts have a large amount of missing data over the WH re- tribution. The inherent weakness of PMW TB in capturing
gion, highlighting its poor utility for various regional appli- deeper snowpack thickness is overcome to a certain extent
cations. The regression modeling approach attempts to find by considering SCD in model development. Thus, the over-
a better fit by optimizing the loss function, i.e., mean error. all improved performance of the multifactor model over the
Over the WH region, the majority of the observations have previously developed models and AMSR2 products can be
SD >25 cm. Therefore, understandably the model estimates attributed to the consideration of topographical parameters
are better in higher-SD regions compared to shallow-SD re- and SCD in the model development. Further, a combination
gions. of multiple lower- and higher-frequency TBs is considered in
The proposed model has an overall positive bias with over- the model for capturing both deeper and shallower snowpack
estimated SD values for shallower SD and underestimation
in the case of higher-SD observations. The bias for the LHZ,

The Cryosphere, 18, 451–474, 2024 [Link]

D. K. Singh et al.: Himalayan snow depth mapping 469

thickness. Different factors affecting the performance of the tively affect the accuracy of the SD estimates from the PMW
multifactor SD model are discussed in Sect. 5.2. multifactor SD model compared to the estimates observed
in the MHZ, whereas, in the MHZ, the absence of forest
5.2 Factors affecting the performance of multifactor cover, relatively low mean SD (compared to both the LHZ
SD model and UHZ), and stable snow conditions result in relatively
better conditions for SD estimation using PMW data. There-
A total of 57 parameters comprising multifrequency PMW fore, compared to other ranges, the multifactor SD model has
TB, BTD, terrain parameters, and SCD are screened using shown improved accuracy in the MHZ, whereas, in the UHZ,
LOOCV to determine the suitable factors to develop the higher SD is present due to which the PMW signal saturates;
PMW multifactor SD model. Finally, the SD algorithm is de- hence larger errors in SD are observed for the multifactor
veloped only using the selected parameters, i.e., geograph- SD model in this region. Thus, location (i.e., latitude), land
ical location parameters (latitude), terrain parameters (ele- cover, elevation, SCD, and magnitude of in situ SD play a
vation, slope, and ruggedness), SCD, and BTDs (36H89V, vital role in the accuracy of multifactor SD model estimates
36V89V, 10V23H, 23H89V, 10V18V, 10H23H, 10H18H, in the WH region.
and 18H89V). In the present study, different combinations The developed model has shown improved performance
of frequencies in vertical and horizontal polarization have compared to other tested approaches in the WH region. How-
been used to estimate shallow to deep snow across different ever, the transferability of the multifactor model to other re-
zones of the WH region. Only descending pass PMW ob- gions, specifically mountainous regions, is uncertain. This is
servations are employed in this study to avoid the problems due to the fact that the relationship of SD with topographi-
pertaining to wet snow, which is more prominent during the cal conditions and SCD can potentially change in the other
ascending pass. Among the different factors (i.e., other than regions. The proposed multifactor model coefficients attempt
PMW data) evaluated using LOOCV, SCD has shown a rel- to improve SD estimates as per prevailing snow conditions in
atively strong correlation (i.e., R = 0.45) with in situ SD ob- the WH region. Understanding the influence of topographi-
servations. Higher SCD generally indicates longer snow per- cal conditions, snow persistency, and snowpack dynamics is
sistence which leads to an increase in snow accumulation, essential for using the model outside the WH region.
whereas shorter SCD indicates the absence/melt of snow,
which leads to lower SD.
Apart from SCD, terrain parameters, i.e., elevation and 6 Summary and conclusions
slope, have an impact on the spatial distribution of SD within
an area (Saydi and Ding, 2020; Trujillo et al., 2007; Sharma The contrasting climate and snow conditions prevailing in
et al., 2014) and have shown a moderately better correlation WH zones present new challenges regarding accurate SD re-
with in situ SD with R values of 0.30 and 0.26, respectively trievals using PMW remote sensing. The limited access to
(see Table 3). Further, the topographic conditions can affect in situ SD data, rugged topography, and inclement weather
the reallocation of PMW radiation due to variations in the di- resulted in fewer SD studies over the WH region. In the
rection of polarization and local incidence angles (You et al., mountainous region, the topography parameters, i.e., eleva-
2011), altering the TB values. The higher-elevation regions tion and slope, affect the snow precipitation and its persis-
(i.e., UHZ) of the Western Himalaya experience cold condi- tence.
tions, which aids snow in accumulating. Therefore, the snow- In this study, different regression approaches (i.e., linear,
fall is preserved for most of the winter in higher mountain logarithmic, reciprocal, and power) are used for developing
areas of the MHZ and UHZ, leading to higher SD in these the multifactor SD models using multifrequency AMSR2 TB
regions. The accuracy of PMW SD models varies with the observations and auxiliary parameters (such as terrain (ele-
magnitude of in situ SD, as evident from the current study, vation, slope), location, and SCD) to estimate SD at 500 m
as well as from many previous studies (X. Xiao et al., 2020; spatial resolution in each WH zone. The overall results indi-
L. Xiao et al., 2020; Dai et al., 2018). However, there are cate that power regression performed better than other tested
many other factors (such as land cover, snow wetness, and approaches in all zones. Further, the results of the multi-
grain size) that can affect the accuracy of SD retrievals (Dong factor model from power regression are evaluated by com-
et al., 2005; Foster et al., 2005; Tedesco et al., 2010; Kur- paring the SD estimates with ground SD, other SD prod-
vonen and Hallikainen, 1997; Ansari et al., 2019). Notably, ucts, and PMW models. The results indicate that under deep
the land cover and snow conditions are considerably differ- snow (>25 cm) conditions, the developed multifactor model
ent from one range to another in the WH region. Therefore, has shown higher accuracy than the AMSR2 operational SD
for any given parameter (such as SCD, elevation, and slope), product and other SD models. However, the accuracy of SD
the accuracy trend of multifactor SD model estimates is not from the multifactor model is affected by variations in auxil-
uniform when compared between different zones. The LHZ iary parameters such as SCD and elevation. With an increase
has forest vegetation, higher temperature, and higher mean in SCD, the SD increased in each WH zone. Additionally,
SD (compared to Greater Himalaya). These conditions nega- the RMSE associated with SD also increased alongside SCD

[Link] The Cryosphere, 18, 451–474, 2024

470 D. K. Singh et al.: Himalayan snow depth mapping

and SD in each WH zone. The MHZ has stable snow con- Disclaimer. Publisher’s note: Copernicus Publications remains
ditions with relatively less thick snowpack. Therefore, the neutral with regard to jurisdictional claims made in the text, pub-
multifactor SD model in this region has shown improved ac- lished maps, institutional affiliations, or any other geographical rep-
curacy for a given SD class compared to other WH zones. resentation in this paper. While Copernicus Publications makes ev-
Overall, the proposed multifactor SD models for WH zones ery effort to include appropriate place names, the final responsibility
lies with the authors.
have demonstrated substantial improvement in estimating SD
compared to the operational AMSR2 SD product; the her-
itage SD model, i.e., Chang’s model; and previous models
Acknowledgements. This work was sponsored by Defence Geoin-
developed within WH zones.
formatics Research Establishment (DGRE), DRDO, under the Con-
Though the multifactor SD model has outperformed other tract for Acquisition of Research Services (CARS) project. The
models and products, there is still scope for improving PMW authors are grateful to the director of DGRE, the chairperson of
SD estimates in the WH region. The multifactor model is ap- CARS, and the review committee of DGRE for their guidance and
plicable only to dry snow conditions. However, in the WH technical suggestions during the different milestones of the project.
region even during the peak winter a substantial area is cov- The support of the DGRE team is also appreciated during field vis-
ered by wet snow. This constrains the utility of the mul- its. We are also thankful to JAXA for providing AMSR2 data in
tifactor model for these regions. The developed model(s) order to carry out this work.
has shown poor performance compared to AMSR2 prod-
ucts when SD <25 cm. This can be possibly attributed to wet
snow conditions prevailing in the early winter, i.e., when SD Financial support. This research has been supported by the De-
will be shallow. Further, the inclusion of snowpack charac- fence Geoinformatics Research Establishment (DGRE), DRDO
(grant no. DGRE-01/CARS/21-22/24/Buildup), under the Contract
teristics such as snow grain size, wetness, and density data
for Acquisition of Research Services (CARS) project.
during the model development can improve the accuracy of
SD estimates. The available in situ SD observations are very
limited considering the high spatiotemporal variability in SD Review statement. This paper was edited by Alexandre Langlois
in this region. Therefore, there is an immediate need to ex- and reviewed by Sartajvir Singh and two anonymous referees.
pand the in situ network of monitoring stations and field-
based studies to determine first-hand knowledge of snowpack
information in the WH region. The brightness temperature
datasets used in this work are resampled to 500 m. Instead of References
resampling, downscaling the TB can be tested for further im-
provement in the model. It is also worthwhile to investigate Ahmad, S.: Impact of climate change on cryosphere-atmosphere-
how downscaling the TB to different resolutions will impact biosphere interaction over the Garhwal Himalaya, India, Disaster
the model performance. Recently, different machine learn- Adv., 13, 32–38, 2020.
ing models are extensively used for modeling SD in many Amlien, J.,: Remote sensing of snow with passive microwave ra-
studies (Tedesco et al., 2004; Liang et al., 2015; Tanniru and diometers: A review of current algorithms. Norsk Regnesentral,
Ramsankaran, 2023). The potential of such machine learning Report No. 1019, 1–52, 2008.
approaches can be investigated for improving the SD estima- Ansari, H., Marofi, S., and Mohamadi, M.: Topography and Land
Cover Effects on Snow Water Equivalent Estimation Using
tion.
AMSR-E and GLDAS Data, Water Resour. Manag., 33, 1699–
1715, [Link] 2019.
Aschbacher, J.: Land surface studies and atmospheric effects by
Data availability. The in situ dataset used in the study can be col- satellite microwave radiometry, PhD thesis, University of Inns-
lected upon request to and approval from Defence Geoinformatics bruck, Innsbruck, Austria, [Link]
Research Establishment, Chandigarh, India. 6853827 (last access: 29 January 2024), 1989.
Bernier, P. Y.: Microwave Remote Sensing of Snowpack Prop-
erties: Potential and Limitations, Hydrol. Res., 18, 1–20,
Author contributions. DKS and ST have equally contributed for the [Link] 1987.
completion of work and development of the manuscript. DKS, ST, Chandrashekar, G. and Sahin, F.: A survey on feature
and RR designed the working methodology of the study. KKS and selection methods, Comput. Electr. Eng., 40, 16–28,
HSN assisted in curating the in situ data and analysis of the results [Link] 2014.
and discussions. DKS and ST have executed the overall work and Chang, A. T. C., Foster, J. L., and Hall, D. K.: Nimbus-7 SMMR
wrote the manuscript. All the authors were involved in manuscript Derived Global Snow Cover Parameters, Ann. Glaciol., 9, 39–
revisions. RR supervised the work. 44, [Link] 1987.
Chang, A. T. C., Foster, J. L., Hall, D. K., Robinson, D. A., Peiji,
L., and Meisheng, C.: The use of microwave radiometer data for
Competing interests. The contact author has declared that none of characterizing snow storage in western China, Ann. Glaciol., 16,
the authors has any competing interests. 215–219, [Link] 1992.

The Cryosphere, 18, 451–474, 2024 [Link]

D. K. Singh et al.: Himalayan snow depth mapping 471

Chang, A. T. C., Foster, J. L., Hall, D. K., Goodison, B. E., Walker, rithms and early results, Remote Sens. Environ., 83, 287–302,
A. E., Metcalfe, J. R., and Harby, A.: Snow parameters de- [Link] 2002.
rived from microwave measurements during the BOREAS win- Foster, J., Chang, A. T. C., and Hall, D. K.: Comparison of snow
ter field campaign, J. Geophys. Res.-Atmos., 102, 29663–29671, mass estimates from a prototype passive microwave snow algo-
[Link] 1997. rithm, a revised algorithm and a snow depth climatology, Remote
Che, T., Li, X., Jin, R., Armstrong, R., and Zhang, T.: Sens. Environ., 62, 132–142, [Link]
Snow depth derived from passive microwave remote- 4257(97)00085-0, 1997.
sensing data in China, Ann. Glaciol., 49, 145–154, Foster, J. L., Sun, C., Walker, J. P., Kelly, R., Chang, A., Dong,
[Link] 2008. J., and Powell, H.: Quantifying the uncertainty in passive mi-
Che, T., Dai, L., Zheng, X., Li, X., and Zhao, K.: Estimation of crowave snow water equivalent observations, Remote Sens. En-
snow depth from passive microwave brightness temperature data viron., 94, 187–203, [Link]
in forest regions of northeast China, Remote Sens. Environ., 183, 2005.
334–349, [Link] 2016. Ganju, A., Thakur, N. K., and Rana, V.: Characteristics of avalanche
Dai, L., Che, T., Xie, H., and Wu, X.: Estimation of accidents in western Himalayan region, India, in: Proceedings
Snow Depth over the Qinghai-Tibetan Plateau Based on of the International Snow Science Workshop, Penticton, BC,
AMSR-E and MODIS Data, Remote Sensing, 10, 1989, Canada, 29 September–4 October 2002, 200–207, [Link]
[Link] 2018. [Link]/snow-science/objects/[Link] (last
Das, I. and Sarwade, R. N.: Snow depth estimation over north- access: 24 January 2024), 2002.
western Indian Himalaya using AMSR-E, Int. J. Remote Sens., Graf, T., Koike, T., Li, X., Hirai, M., and Tsutsui, H.: Assim-
29, 4237–4248, [Link] ilating passive microwave brightness temperature data into a
2008. land surface model to improve the snow depth predictability,
Derksen, C.: The contribution of AMSR-E 18.7 and 10.7 GHz in: International Geoscience and Remote Sensing Symposium
measurements to improved boreal forest snow water equiv- (IGARSS), Denver, CO, USA, 31 July–4 August, 2006, 710–
alent retrievals, Remote Sens. Environ., 112, 2701–2710, 713, [Link] 2006.
[Link] 2008. Grippa, M., Mognard, N., le Toan, T., and Josberger, E. G.: Siberia
Dietz, A. J., Kuenzer, C., Gessner, U., and Dech, S.: Remote sensing snow depth climatology derived from SSM/I data using a com-
of snow – a review of available methods, Int. J. Remote Sens., bined dynamic and static algorithm, Remote Sens. Environ., 93,
33, 4094–4134, [Link] 30–41, [Link] 2004.
2012. Grody, N. C. and Basist, A. N.: Global identification of snowcover
Dimri, A. P. and Dash, S. K.: Wintertime climatic trends using ssm/i measurements, IEEE T. Geosci. Remote, 34, 237–
in the western Himalayas, Climatic Change, 111, 775–800, 249, [Link] 1996.
[Link] 2012. Gurung, D. R., Giriraj, A., Aung, K. S., Shrestha, B., and
Dong, C.: Remote sensing, hydrological modeling and in situ obser- Kulkarni, A. V.: Snow-cover mapping and monitoring in
vations in snow cover research: A review, J. Hydrol., 561, 573– the Hindu Kush-Himalayas, International Centre for In-
583, [Link] 2018. tegrated Mountain Development (ICIMOD), Patan, Nepal,
Dong, J., Walker, J. P., and Houser, P. R.: Factors affecting re- [Link] 2011.
motely sensed snow water equivalent uncertainty, Remote Sens. Gusain, H. S., Mishra, V. D., Arora, M. K., Mamgain,
Environ., 97, 68–82, [Link] S., and Singh, D. K.: Operational algorithm for genera-
2005. tion of snow depth maps from discrete data in Indian
Estilow, T. W., Young, A. H., and Robinson, D. A.: A long-term Western Himalaya, Cold Reg. Sci. Technol., 126, 22–29,
Northern Hemisphere snow cover extent data record for cli- [Link] 2016.
mate studies and monitoring, Earth Syst. Sci. Data, 7, 137–142, Imaoka, K., Maeda, T., Kachi, M., Kasahara, M., Ito, N., and
[Link] 2015. Nakagawa, K.: Status of AMSR2 instrument on GCOM-
Farr, T. G., Rosen, P. A., Caro, E., Crippen, R., Duren, R., W1, in: Earth observing missions and sensors, Develop-
Hensley, S., Kobrick, M., Paller, M., Rodriguez, E., Roth, ment, implementation, and characterization II, 8528, 201–206,
L., Seal, D., Shaffer, S., Shimada, J., Umland, J., Werner, [Link] 2011.
M., Oskin, M., Burbank, D., and Alsdorf, D. E.: The Shut- Jarvis, A., Reuter, H., Nelson, A., and Guevara, E.: Hole-filled
tle Radar Topography Mission, Rev. Geophys., 45, RG2004, SRTM for the globe Version 4, available from the CGIAR-
[Link] 2007. CSI SRTM 90m Database, [Link] (last access:
Ferraro, R. R., Weng, F., Grody, N. C., and Basist, A.: An Eight- 24 January 2024), 2008.
Year (1987–1994) Time Series of Rainfall, Clouds, Water Vapor, Jiang, L. M., Wang, P., Zhang, L. X., Yang, H., and Yang,
Snow Cover, and Sea Ice Derived from SSM/I Measurements, B. J. T.: Improvement of snow depth retrieval for FY3B-
Am. Meteorol. Soc., 77, 891–905, [Link] MWRI in China, Sci. China Earth Sci., 57, 1278–1292,
0477(1996)077<0891:AEYTSO>[Link];2, 1996. [Link] 2014.
Friedl, M. A., McIver, D. K., Hodges, J. C. F., Zhang, X. Joshi, J. C., Tankeshwar, K., and Srivastava, S.: Hidden Markov
Y., Muchoney, D., Strahler, A. H., Woodcock, C. E., Gopal, Model for quantitative prediction of snowfall and analysis of haz-
S., Schneider, A., Cooper, A., Baccini, A., Gao, F., and ardous snowfall events over Indian Himalaya, J. Earth Syst. Sci.,
Schaaf, C.: Global land cover mapping from Modis: Algo- 126, 33, [Link] 2017.

[Link] The Cryosphere, 18, 451–474, 2024

472 D. K. Singh et al.: Himalayan snow depth mapping

Kelly, R.: The AMSR-E Snow Depth Algorithm: Description and High Mountain Asia (2002–2019), PANGAEA [data set],
Initial Results, Journal of The Remote Sensing Society of Japan, [Link] 2020.
29, 307–317, [Link] 2009. Muhammad, S. and Thapa, A.: An improved Terra–Aqua
Kelly, R. E., Chang, A. T., Tsang, L., and Foster, J. MODIS snow cover and Randolph Glacier Inventory 6.0 com-
L.: A prototype AMSR-E global snow area and snow bined product (MOYDGL06*) for high-mountain Asia be-
depth algorithm, IEEE T. Geosci. Remote, 41, 230–242, tween 2002 and 2018, Earth Syst. Sci. Data, 12, 345–356,
[Link] 2003. [Link] 2020.
Kelly, R. E. J., Chang, A. T. C., Tsang, L., and Chen, C. Mukherji, A., Sinisalo, A., Nüsser, M., Garrard, R., and Eriksson,
T.: Parameterization of snowpack grain size for global M.: Contributions of the cryosphere to mountain communities
satellite microwave estimates of snow depth, in: Inter- in the Hindu Kush Himalaya: a review, Reg. Environ. Change,
national Geoscience and Remote Sensing Symposium 19, 1311–1326, [Link]
(IGARSS), Toronto, ON, Canada, 24–28 June 2002, 686– 2019.
688, [Link] 2002. Mukul, M., Srivastava, V., Jade, S., and Mukul, M.: Uncertainties in
Kelly, R. E. J., Chang, A. T. C., Foster, J. L., and Hall, D. K.: the Shuttle Radar Topography Mission (SRTM) heights: Insights
Using Remote Sensing and Spatial Models to Monitor Snow from the Indian himalaya and Peninsula, Scientific Reports, 7,
Depth and Snow Water Equivalent, in: Spatial Modelling of 41672, [Link] 2017.
the Terrestrial Environment, edited by: Kelly, R. E. J., Drake, Negi, H. S., Kanda, N., and Ganju, A.: Recent wintertime climatic
N. A., and Barr, S. L., John Wiley & Sons, Ltd., 35–57, variability over the North West Himalayan cryosphere, Curr. Sci.,
[Link] 2005. 114, 760–770, 2018.
Kinar, N. J. and Pomeroy, J. W.: Measurement of the physi- Negi, H. S., Ganju, A., Kanda, N., and Gusain, H. S.: Climate
cal properties of the snowpack, Rev. Geophys., 53, 481–544, Change and Cryospheric Response Over North-West and Central
[Link] 2015. Himalaya, India, in: Himalayan Weather and Climate and their
Kumar, P., Saharwardi, M. S., Banerjee, A., Azam, M. F., Dubey, A. Impact on the Environment, edited by: Dimri, A. P., Bookhagen,
K., and Murtugudde, R.: Snowfall Variability Dictates Glacier B., Stoffel, M., and Yasunari, T., Springer Nature, Switzerland,
Mass Balance Variability in Himalaya-Karakoram, Scientific 309–330, [Link] 2020.
Reports, 9, 18192, [Link] Nüsser, M., Dame, J., Parveen, S., Kraus, B., Baghel, R.,
2019. and Schmidt, S.: Cryosphere-Fed Irrigation Networks in the
Kurvonen, L. and Hallikainen, M.: Influence of land-cover category Northwestern Himalaya: Precarious Livelihoods and Adapta-
on brightness temperature of snow, IEEE T. Geosci. Remote, 35, tion Strategies Under the Impact of Climate Change, Mt. Res.
367–377, [Link] 1997. Dev., 39, 1–11, [Link] (last ac-
Kwon, Y., Yang, Z. L., Hoar, T. J., and Toure, A. M.: Improving the cess: 27 January 2024), 2019.
Radiance Assimilation Performance in Estimating Snow Water Obaid, H. S., Dheyab, S. A., and Sabry, S. S.: The impact of data
Storage across Snow and Land-Cover Types in North America, pre-processing techniques and dimensionality reduction on the
J. Hydrometeorol., 18, 651–668, [Link] accuracy of machine learning, in: IEMECON 2019 – 9th Annual
D-16-0102.1, 2017. Information Technology, Electromechanical Engineering and
Liang, J., Liu, X., Huang, K., Li, X., Shi, X., Chen, Microelectronics Conference, Jaipur, India, 13–15 March 2019,
Y., and Li, J.: Improved snow depth retrieval by in- 279–283, [Link]
tegrating microwave brightness temperature and visible/in- 2019.
frared reflectance, Remote Sens. Environ., 156, 500–509, Rango, A., Chang, A. T. C., and Foster, J. L.: Utiliza-
[Link] 2015. tion of Spaceborne Microwave Radiometers for Moni-
Lemke, P., Ren, J., Alley, R. B., Allison, I., Carrasco, J., Flato, G., toring Snowpack Properties, Nord. Hydrol., 10, 25–40,
Fujii, Y., Kaser, G., Mote, P., Thomas, R. H., and Zhang, T.: AR4 [Link] 1979.
– Changes in snow, ice and frozen ground, Climate Change 2007: Saraf, A. K., Tarafdar, S., Foster, J. L., and Singh, P.: Passive
The Physical Science Basis. Contribution of Working Group microwave data for snow-depth and snow-extent estimations
I to the Fourth Assessment Report of the Intergovernmental in the Himalayan mountains, Int. J. Remote Sens., 20, 83–95,
Panel on Climate Change, [Link] [Link] 1999.
observations-changes-in-snow-ice-and-frozen-ground/ (last ac- Saydi, M. and Ding, J.: Impacts of topographic factors on regional
cess: 27 January 2024), 2007. snow cover characteristics, Water Science and Engineering, 13,
Luojus, K., Pulliainen, J., Takala, M., Lemmetyinen, J., Mortimer, 171–180, [Link] 2020.
C., Derksen, C., Mudryk, L., Moisander, M., Hiltunen, M., Sharma, S. S. and Ganju, A.: Complexities of avalanche forecasting
Smolander, T., Ikonen, J., Cohen, J., Salminen, M., Norberg, J., in Western Himalaya – an overview, Cold Reg. Sci. Technol., 31,
Veijola, K., and Venäläinen, P.: Globsnow v3.0 Northern Hemi- 95–102, [Link] 2000.
sphere Snow Water Equivalent Dataset, Scientific Data, 8, 163, Sharma, V., Mishra, V. D., and Joshi, P. K.: Topographic con-
[Link] 2021. trols on spatio-temporal snow cover distribution in North-
McClung, D. M.: Avalanche character and fatalities in the west Himalaya, Int. J. Remote Sens., 35, 3036–3056,
high mountains of Asia, Ann. Glaciol., 57, 114–118, [Link] 2014.
[Link] 2016. Singh, D., Juyal, V., and Sharma, V.: Consistent seasonal
Muhammad, S.: Improved daily MODIS TERRA/AQUA snow cover depth and duration variability over the West-
Snow and Randolph Glacier Inventory (RGI6.0) data for

The Cryosphere, 18, 451–474, 2024 [Link]

D. K. Singh et al.: Himalayan snow depth mapping 473

ern Himalayas (WH), J. Earth Syst. Sci., 125, 1451–1461, Tedesco, M., Reichle, R., Löw, A., Markus, T., and Foster, J. L.:
[Link] 2016. Dynamic approaches for snow depth retrieval from spaceborne
Singh, D. K., Singh, K. K., Mishra, V. D., and Sharma, J. K.: microwave brightness temperature, IEEE T. Geosci. Remote, 48,
Formulation of Snow Depth Algorithms for Different regions 1955–1967, [Link] 2010.
of NW Himalaya using Passive Microwave Satellite Data, Tedesco, M., Derksen, C., Deems, J. S., and Foster, J. L.:
International Journal of Engineering Research & Technology, 1, Remote sensing of snow depth and snow water equiva-
1–9, [Link] lent, in: Remote Sensing of the Cryosphere, edited by:
algorithms-for-different-regions-of-nw-himalaya-using-passive- Tedesco, M., John Wiley & Sons, Ltd, Chichester, UK,
[Link] (last access: [Link] 2015.
27 January 2024), 2012. Thakur, P. K., Garg, V., Nikam, B. R., and Aggarwal, S. P.:
Singh, D. K., Gusain, H. S., Mishra, V., and Gupta, N.: Snow cover Cryosphere Studies in Northwest Himalaya, Remote Sensing
variability in North-West Himalaya during last decade, Arab. of Northwest Himalayan Ecosystems, Springer Verlag, Singa-
J. Geosci., 11, 579, [Link] pore, Singapore., 69–107, [Link]
2018. 2128-3_5, 2019.
Singh, K. K. and Mishra, V. D.: Snow cover study of north- Trujillo, E., Ramírez, J. A., and Elder, K. J.: Topographic, meteo-
west Himalaya using passive microwave remote sensing data, rologic, and canopy controls on the scaling characteristics of the
in: Proceedings of SPIE Asia pacific remote sensing con- spatial distribution of snow depth fields, Water Resour. Res., 43,
ference, Microwave Remote Sensing of the Atmosphere and W07409, [Link] 2007.
Environment V, Goa, India, 13–17 November 2006, 641014, Velliangiri, S., Alagumuthukrishnan, S., and Thankumar, J. S.
[Link] 2006. I.: A review of dimensionality reduction techniques for ef-
Singh, K. K., Mishra, V. D., and Negi, H. S.: Evaluation of snow ficient computation, Procedia Comput. Sci., 165, 104–111,
parameters using passive microwave remote sensing, Def. Sci. [Link] 2019.
J., 57, 271–278, 2007. Vishwakarma, B. D., Ramsankaran, R. A. A. J., Azam, M. F.,
Singh, K. K., Kumar, A., Kulkarni, A. V., Datt, P., Dewali, S. K., Bolch,T., Mandal, A., Srivastava, S., Kumar, P., Sahu, R.,
Kumar, V., and Chauhan, R.: Snow depth estimation in the Indian Navinkumar, P. J., Tanniru, S. R., Javed, A., Soheb, M., Dimri,
Himalaya using multi-channel passive microwave radiometer, A. P., Yadav, M., Devaraju, B., Chinnasamy, P., Reddy, M.
Curr. Sci., 108, 942–953, [Link] J., Murugesan, G. P., Arora, M., Jain, S. K., Ojha, C. S. P.,
(last access: 27 January 2024), 2015 Harrison, S., and Bamber, J.: Challenges in Understanding the
Singh, K. K., Kumar, R., Singh, D. K., Negi, H. S., Dewali, S. K., Variability of the Cryosphere in the Himalaya and Its Im-
and Kedia, J.: Retrieving snow cover information from AMSR-2 pact on Regional Water Resources, Front. Water, 4, 909246,
satellite data for North-West Himalaya, India, Geocarto Int., 35, [Link] 2022.
1783–1799 , [Link] Wang, J., Huang, X., Wang, Y., and Liang, T.: Retrieving
2020. Snow Depth Information from AMSR2 Data for Qinghai-
Singh, S. K., Rathore, B. P., Bahuguna, I. M., and Ajai: Snow cover Tibet Plateau, IEEE J. Sel. Top. Appl., 13, 752–768,
variability in the Himalayan–Tibetan region, Int. J. Climatol., 34, [Link] 2020.
446–452, [Link] 2014. Wang, P., Jiang, L., Zhang, L., and Guo, Y.: Impact of
Stiles, W. H. and Ulaby, F. T.: The active and pas- terrain topography on retrieval of snow water equiva-
sive microwave response to snow parameters: 1. lence using passive microwave remote sensing, in: In-
Wetness, J. Geophys. Res.-Oceans, 85, 1037–1044, ternational Geoscience and Remote Sensing Symposium
[Link] 1980. (IGARSS), Honolulu, HI, USA, 25–30 July 2010, 1757–1760,
Takala, M., Luojus, K., Pulliainen, J., Derksen, C., Lemmetyi- [Link] 2010.
nen, J., Kärnä, J. P., Koskinen, J., and Bojkov, B.: Estimating Wang, Y., Huang, X., Deng, J., Ma, X., and Liang, T.: Development
northern hemisphere snow water equivalent for climate research and validation for daily cloud-free snow products in middle-and-
through assimilation of space-borne radiometer data and ground- high latitude areas in Eurasia, Remote Sensing Technology and
based measurements, Remote Sens. Environ., 115, 3517–3529, Application, 31, 1013–1021, [Link]
[Link] 2011. V31/I5/1013 (last access: 24 January 2024), 2016.
Tanniru, S. and Ramsankaran, R. A. A. J.: Passive Mi- Wang, Y., Huang, X., Wang, J., Zhou, M., and Liang, T.: AMSR2
crowave Remote Sensing of Snow Depth: Techniques, Chal- snow depth downscaling algorithm based on a multifactor ap-
lenges and Future Directions, Remote Sens., 15, 1052, proach over the Tibetan Plateau, China, Remote Sens. Environ.,
[Link] 2023. 231, 111268, [Link] 2019.
Tedesco, M. and Narvekar, P. S.: Assessment of the NASA Webb, G. I., Sammut, C., Perlich, C., Horváth, T., Wrobel, S., Korb,
AMSR-E SWE Product, IEEE J. Sel. Top. Appl., 3, 141–159, K. B., Noble, W. S., Leslie, C., Lagoudakis, M. G., Quadrianto,
[Link] 2010. N., Buntine, W. L., Quadrianto, N., Buntine, W. L., Getoor, L.,
Tedesco, M., Pulliainen, J., Takala, M., Hallikainen, M., and Pam- Namata, G., Getoor, L., Han, X. J. J., Ting, J.-A., Vijayaku-
paloni, P.: Artificial neural network-based techniques for the re- mar, S., Schaal, S., and de Raedt, L.: Leave-One-Out Cross-
trieval of SWE and snow depth from SSM/I data, Remote Sens. Validation, in: Encyclopedia of Machine Learning, Springer
Environ., 90, 76–85, [Link] US, 600–601, [Link]
2004. 2011.

[Link] The Cryosphere, 18, 451–474, 2024

474 D. K. Singh et al.: Himalayan snow depth mapping

Wei, Y., Li, X., Gu, L., Zheng, X., Jiang, T., Li, X., Yang, J. W., Jiang, L. M., Lemmetyinen, J., Pan, J. M., Luojus, K.,
and Wan, X.: A Dynamic Snow Depth Inversion Algo- and Takala, M.: Improving snow depth estimation by coupling
rithm Derived from AMSR2 Passive Microwave Bright- HUT-optimized effective snow grain size parameters with the
ness Temperature Data and Snow Characteristics in North- random forest approach, Remote Sens. Environ., 264, 112630,
east China, IEEE J. Sel. Top. Appl., 14, 5123–5136, [Link] 2021.
[Link] 2021. You, Q., Kang, S., Ren, G., Fraedrich, K., Pepin, N., Yan, Y., and
Xiao, L., Che, T., and Dai, L.: Evaluation of Remote Sens- Ma, L.: Observed changes in snow depth and number of snow
ing and Reanalysis Snow Depth Datasets over the Northern days in the eastern and central Tibetan Plateau, Clim. Res., 46,
Hemisphere during 1980–2016, Remote Sensing, 12, 3253, 171–183, [Link] 2011.
[Link] 2020. Yu, H., Zhang, X., Liang, T., Xie, H., Wang, X., Feng, Q., and
Xiao, X., Zhang, T., Zhong, X., Shao, W., and Li, X.: Support vec- Chen, Q.: A new approach of dynamic monitoring of 5-day
tor regression snow-depth retrieval algorithm using passive mi- snow cover extent and snow depth based on Modis and amsr-e
crowave remote sensing data, Remote Sens. Environ., 210, 48– data from Northern Xinjiang region, Hydrol. Process., 26, 3052–
64, [Link] 2018. 3061, [Link] 2012.
Xiao, X., Zhang, T., Zhong, X., and Li, X.: Spatiotemporal Variation Zhang, R., Liang, T., Feng, Q., Huang, X., Wang, W.,
of Snow Depth in the Northern Hemisphere from 1992 to 2016, Xie, H., and Guo, J.: Evaluation and adjustment of the
Remote Sensing, 12, 2728, [Link] amsr2 snow depth algorithm for the northern Xinjiang re-
2020. gion, China, IEEE J. Sel. Top. Appl., 10, 3892–3903,
Yang, J., Jiang, L., Luojus, K., Pan, J., Lemmetyinen, J., Takala, [Link] 2017.
M., and Wu, S.: Snow depth estimation and historical data
reconstruction over China based on a random forest ma-
chine learning approach, The Cryosphere, 14, 1763–1778,
[Link] 2020.

The Cryosphere, 18, 451–474, 2024 [Link]