3.1.

Snow Cover Area

Initially, snow was successfully mapped by visual interpretation using ESSA-3 satellite data.¹⁸ These maps have been continually improved with the addition of new satellite data, such as the Landsat Multispectral Scanning Subsystem, the Landsat Thematic Mapper (TM), the NOAA Advanced Very High Resolution Radiometer (AVHRR), and the Moderate-Resolution Imaging Spectrometer (MODIS).^19–22 Some operational regional or global snow cover binary products have also been issued and can be readily acquired.^23,24 In China, there have been many studies of snow cover mapping since the 1980s. Based on field measurements of snow and ice reflectance values, Zeng et al.²⁵ and Cao et al.²⁶ searched the optimal bands for snow cover retrieval. Ma et al.²⁷ utilized supervised classification (SC) of NOAA-AVHRR data to map the snow cover area in Northern Xinjiang, China. Because these methods are not automated, they can only be used for snow mapping in local regions. Using TM, AVHRR, and MODIS images, Wang^28,29 assessed the accuracy of three methods: training sites’ SC, digital numbers’ statistics, and the Normalized Difference Snow Index (NDSI)—they found the NDSI method to be more accurate. With the development of sensor techniques, various snow cover data have been improved. Xiao et al.^30,31 developed a snow cover product using SPOT-VEGETATION data. Yan³² developed a complex index method to map snow cover and mask cloud cover using NOAA-16 data. In addition, Li et al.³³ produced a daily snow product by combining one day’s worth of multitemporal images at a spatial resolution of $\sim 5\mathrm{km}$, based on FY-2 satellite data. In general, although there are many snow cover products developed using different methods and remote sensors, MODIS snow cover products, which are produced by the NDSI method and derived from the NASA National Snow and Ice Data Center (NSIDC), are more accurate and are widely used in China.³⁴

The NDSI threshold, used for retrieving MODIS snow cover products, obtained from MODIS global snow cover products, was initially developed and validated in the United States, Canada, and Russia. Hao and Wang³⁵ showed that the MODIS snow cover product underestimated snow cover in the Qilian Mountains and that the regional NDSI threshold is much less than 0.4. In addition, a common challenge is how to exclude the cloud pixels in snow cover products. To address this, a series of blending algorithms which provide daily no-cloud snow cover from MODIS data have been developed and include the following: daily Terra-Aqua/MODIS image combination,^36–40 temporal deduction,⁴¹ the snow-line method (SnowL),⁴² and multisensor combinations.⁴³ Hao et al.⁴⁴ developed a blending algorithm which can provide daily no-cloud snow cover from MODIS data by adjusting the NDSI threshold value and combining MODIS and AMSR-E data. This method was used to produce 3-years’ (2008 to 2010) worth of daily no-cloud snow cover products of the Qinghai–Tibet Plateau. Figure 1 shows the workflow of the improved MODIS cloud removal algorithm applied to the Qinghai–Tibet Plateau on January 25, 2008. In the study, there are three steps needed to produce the snow cover products. The output from each step is the input for the next step. Moreover, by combining the previous cloud removal algorithms, Huang et al.⁴⁵ generated a new daily cloud-free snow cover product using MODIS daily snow cover products (MOD10A1, MYD10A1), an AMSR-E daily SWE product, and a digital elevation model. This method is expected to play an important role in improving the accuracy of snow cover monitoring. Considering the inherent deficiency of MODIS data—that patchy snow or snow depth of below 3 cm is often missed—the improved snow products show agreement up to 91.7% with an overall accuracy of 91.9%. Figure 2 shows the MODIS daily cloud-free snow product composite of the Qinghai–Tibet Plateau, China, on February 18, 2011.

Fig. 1

The workflow of the improved MODIS cloud removal algorithm in the Qinghai–Tibet Plateau on January 25, 2008. First, the image fusion of MODIS Terra and Aqua snow cover products which have been improved by adjusting of NDSI. The second step deduces the cloud-covered pixel based on the one day forward and one day backward information of same pixels. Finally, blending MODIS snow cover and improved AMSR-E Snow depth data would eliminate completely remove cloud pixels (Reproduced from Ref. 44).

Fig. 2

MODIS daily cloud free snow product composite process in the Qinghai–Tibet Plateau, China, on February 18, 2011. (a) is the MODIS snow cover product from Terra, (b) is the MODIS snow cover product from Aqua, (c) is the composite of Terra and Aqua snow cover product, (d) is the adjacent temporal composite, (e) is the interpolation using SnowL methods, (f) is the composite of results of (e) and AMSR-E snow water equivalent products. Each step was considered as the input of next step. (Reproduced from Ref. 45).

3.2.

Fraction of Snow Cover

Traditional snow cover mapping methods treat all pixels as uniform. Although the NDSI-based method has improved estimation accuracy as compared with other methods, it is still limited by the mixed pixel problem and is thus unable to meet the precision requirement for snow characteristics at local or basin scales. Thus, a more accurate snow cover fraction (SCF) map is desired over a traditional binary snow map.

At present, three types of approaches have been developed for deriving the SCF: statistical regression, spectral unmixing, and machine learning.

Statistical regression is the most widely used type; it establishes a unified regression model based on the relationship between the SCF and reflectance/NDSI.^46–52 MODIS SCF products from the NSIDC, such as MOD10A1 for Terra and MYD10A1 for Aqua, were produced using statistical regression. In China, Jin et al. modified the parameters of the subpixel SCF model proposed by Salomonson.⁵¹ They retrieved the subpixel SCF from MODIS data using CBERS-2 CCD data as ground truth in Northeast China.⁵³ Zhang et al.⁵⁴ also established a linear relationship model between SCF and NDSI using HJ-1B data as ground truth. The precision of the linear regression using a single value of the NDSI is low, where the SCF is near 1 or 0. To solve this problem, the relationship between the SCF and multiple factors can be defined; Cao and Liu⁵⁵ developed a linear relationship model between SCF, NDSI, and NDVI to improve snow detection in forests. Moreover, several different segmentation methods were also proposed to determine SCF, which were validated by using ETM+ or HJ-1B CCD/IRS images. Findings show that the product generated using the segmentation method is improved, as compared with the MOD10A1/SCF product.^56–58 Additionally, Liu et al.⁵⁹ proposed a nonlinear regression model to retrieve the SCF based on the NDSI. Tang et al.⁶⁰ considered different land cover types and developed a subpixel snow mapping algorithm for the Qinghai–Tibet Plateau using MODIS data and linear regression. The statistical regression model is simple and easy to operate. However, this method is limited and difficult to apply due to large differences in the model parameters, which are caused by factors such as different land cover types and regions and the physical state of the snow.

Spectral unmixing is another widely used method, which creates a hybrid, simulated spectrum and estimates the mixed reflectance for each pixel in each frequency band. The reflectance of each mixed pixel is then calculated as a function of the spectral endmember.^61–67 Linear spectral mixture analysis is the most commonly used spectral unmixing method, which regards the observed reflectance as a linear combination of reflectance from certain surface constituents. Liang et al.⁶⁸ used a linear mixture spectrum disassembling method to study the snow coverage rate and spatial classification for four NOAA satellite images under sunny conditions during two snow disasters in China from 1996 to 1997. Pei et al.⁶⁹ used a linear spectrum mixing model to extract the SCF in the northern Xinjiang Region from MODIS imagery. Chen et al.⁷⁰ extracted snow area at a subpixel scale using a linear spectral mixture model from MOD02HKM imagery. Zhu et al.⁷¹ retrieved the subpixel snow-covered area for the Qinghai–Tibet Plateau from MODIS data by employing a linear spectral mixture method based on multiple-endmember spectral analysis. Zhang⁷² tested one linear mixture model (LMM) and three different nonlinear subpixel analysis methods: fuzzy c-means clustering, back-propagation neural networks, and support vector machine (SVM). They concluded that the LMM and the SVM provided the best estimates of snow cover components. Notably, Hao et al.⁷³ validated the accuracy of four different subpixel analysis methods: linear regression, full-constrained linear mixed-pixel decomposition, sparse regression unmixing, and non-negative matrix unmixing, based on the known proportion of snow in Northern Xinjiang, China. A comparison of the above methods shows that the linear regression method has the lowest accuracy, especially when the snow proportion is less than 50%; the accuracy of the sparse regression algorithm and non-negative matrix factorization were slightly higher than the full-constrained linear mixed-pixel decomposition (Fig. 3). Better results could be obtained using other unmixing inversion algorithms; however, the computation will be very time consuming due to the large amount of remote sensing data. In general, the advantage of linear spectrum unmixing is that it can be easily modeled due to having a clear physical and scientific basis. The previous studies also confirm that the method gives a comparatively accurate estimation of snow. The main drawback of linear spectrum unmixing is the assumption that the same object in a pixel has the same spectral characteristics. However, the same object may yield different spectra; different objects can have the same spectrum. Thus, the selection of reasonable endmembers and reference spectral values can be problematic and may result in large errors. The nonlinear spectral unmixing model is generally more complex. Many of the model parameters are difficult to accurately determine, making it difficult to use in practice. In addition, the computational efficiency of the method is usually problematic for large data volumes.

Fig. 3

The comparison of results of fractional snow cover between in situ measurements and the results from four spectral unmixed models from the Fukang and Fuyun study areas (Reproduced from Ref. 73).

Based on intelligent computing, a machine learning approach can be used to estimate the SCF as an alternative to regression and spectral unmixing. Dobreva and Klein^74,75 first investigated the applicability of three-layered feed-forward error back propagation artificial neural networks (ANN) to estimate MODIS SCF of flat areas in the Northern Hemisphere. Inspired by this, Hou and Huang⁷⁶ developed a three-layer feed-forward ANN for mountainous fractional snow cover (FSC) mapping using MODIS products [reflectance at seven bands, NDSI, land surface temperature (LST), and SCF], ETM+, and auxiliary topographic (elevation, slope, and aspect) data. They designed eight experimental schemes according to the different combinations of input components. Images from three time periods were chosen to train and validate the ANN; the proposed method was tested on three different test sets (nonindependent, temporal independent, and spatial-temporal independent) at the upper reaches of the Heihe River Basin. The SCF maps, derived from Landsat ETM+ reference values, MODIS SCF (MOD10), eight ANN schemes on nonindependent Test1 and spatial-temporal independent Test4, are shown in Fig. 4. The figure illustrates that the ANN can easily incorporate auxiliary information to improve the accuracy of mountainous FSC mapping with considerable generalization. However, a machine learning-based approach requires adequate representation of the training samples, which can be difficult to achieve. Additionally, due to the massive number of pixels in remote sensing images, the training set is very large, which often leads to unstable performance.

Fig. 4

Snow cover fraction (SCF) maps from (a) ground-truth derived from Landsat ETM+, (b) MOD10, and (c)–(j) results derived from different artificial neural network schemes on nonindependent Test1 (Reproduced from Ref. 76).

3.3.

Snow Grain Size

Snow grain size is an important parameter that can indicate snow albedo change. It is also an important input factor of many snow and climate models. Satellite remote sensing of snow is critical for monitoring snow age, pollution levels, and grain sizes over difficult-to-access mountainous regions. Fundamental to all remote sensing algorithms is the relationship between the measured reflectance and the to-be-retrieved microphysics characteristics.

Over the past 20 years, many researchers have made use of remote sensing satellite images, such as TM and AVHRR, to retrieve snow grain size.^77–79 Nolin and Dozier⁸⁰ described a method for remotely sensing grain size from airborne visible / infrared imaging spectrometer (AVIRIS) data using the snow reflectance at 1.03 μm generated by the discrete ordinates radiative transfer (DISORT) program for a multilayered plane-parallel medium model and the wavelength of prominent ice absorption features. This method was sensitive to sensor noise and required knowledge of the solar and viewing geometries. Subsequently, a more robust algorithm⁸¹ that integrated the absorption bands near the 1.03-μm wavelength (0.96 to 1.08 μm) minimized many effects of noise on grain size retrievals. Painter et al.^82,83 used a linear spectral unmixing method to map relative grain size at subpixel resolution from AVIRIS data. Then, they developed MODSCAG (MODIS snow-covered area and grain size), which maps snow and its grain size simultaneously using spectral mixture analysis coupled with a radiative transfer model.

In China, there are also studies on retrieving snow grain size from remote sensing data. Hao et al.⁸⁴ retrieved snow grain size using three methods applied to HYPERION data over Qilian Mountain, China. The results of the three methods were compared with the in situ data of snow grain size, which were measured synchronously as the HYPERION sensor passed over the snowpack field.⁸⁵ Jiang et al.⁸⁶ also validated these methods using spectrum reflection data of snow in different grain sizes measured over the Binggou Basin. However, their methods generally used either the WW (Wiscombe–Warren) or DISORT model. They assume snow particles to be an equivalent sphere; Mie scattering (MIE) scattering theory is then used to model the optical characters of a single spherical snow particle. Nevertheless, as suggested by numerous experimental studies of microphysics snow properties, the snow pack should be considered as a multiple scattering, closely packed medium consisting of snow particles with irregular shapes.^87–90 Kokhanovsky et al.^91,92 developed an asymptotic radiative transfer (ART) model to model the snow spectra for different snow shapes and grain sizes. It simplified the traditional radiative transfer model, but without the loss of model accuracy. Hao et al.⁹³ developed an improved snow retrieval method using the ART model and the in situ spectral reflectance of different snow grain sizes. Finally, the method was validated using in situ snow grain size, which was used to calculate the equivalent snow grain size by the improved digital snow particle properties method. Figure 5 shows the validation results of the improved snow grain retrieval algorithm.

Fig. 5

Scatterplot showing the correlation between measured snow grain size and grain size derived from the asymptotic radiative transfer method (Reproduced from Ref. 93).

3.4.

Microwave Remote Sensing of Snow

Seasonal snow cover plays an important role in hydrological cycles and climate processes.^94,95 The observations of sparse weather stations cannot meet the demand; thus, it is important to develop a retrieval method of snow depth or SWE from remote sensing data. Passive microwave remote sensing is the most efficient way to retrieve snow depth and SWE.

Snow depth retrieval can be influenced by the condition of snowpacks, such as snow crystal,^96–98 snow density,^99,100 and vegetation.¹⁰¹ Inaccurate snow properties will lead to large errors in the retrieved snow depths. The global passive microwave snow depth retrieval algorithms did not consider these characteristics and overestimated snow depth in western China.^101,102

Snow depth retrieval from passive microwave dates back to 1993 in China. Cao and Li¹⁰³ used scanning multichannel microwave radiometer (SMMR) data to develop a snow depth retrieval algorithm for high mountains, plateaus, low mountains, hills, and basins in western China. In order to meet the demands of climate and hydrological researchers, Che et al.¹⁰⁴ developed a snow depth time series. Che et al. modified the Chang algorithm¹⁰⁵ and developed a relationship between the observed snow depth and passive microwave brightness temperature from SMMR and SSM/I, respectively. In this algorithm, the influence of forest on snow depth retrieval was removed by integrating the forest fraction into the algorithm. Because precipitation, cold deserts, and frozen ground have similar volume scattering with snowpack, they are initially discerned from snowpack using a decision tree.¹⁰⁶ This long-term daily snow depth data product was obtained via the WESTDC (the Environmental and Ecological Science Data Center for West China, http://westdc.westgis.ac.cn/). The data were used to analyze the interannual variations of snow depth in China and three stable snow cover regions [Northwest, Northeast, and Qinghai-Xizang (Tibet) Plateau] from 1978 to 2007;¹⁰⁴ updated results (2012) are shown in Fig. 6.

Fig. 6

Interannual variation of snow depth in China and three stable snows cover regions (Reproduced from Ref. 104).

Since 2010, the microwave radiation instrument, carried on the FY-3B satellite, has provided brightness temperature data at the same frequencies as AMSR-E. Based on these data, a statistical relationship between snow depth and brightness temperature was developed for bare soil, grassland, and forest.¹⁰⁷ This relationship is used in the operational algorithm for developing snow depth and snow-water equivalent products at the Chinese Meteorological Administration.

Because snow properties vary in both space and time, static algorithms tend to underestimate snow depth in the early snow season and overestimate in the late snow season. Therefore, a dynamic algorithm is needed to accurately monitor snow depth variation in a snow season and requires seasonal information of snow properties. In China, Che et al.¹⁰⁴ calculated the effects on snow grain size and density due to seasonal variations and used an empirical offset to modify their respective influences. The snow depth data product (at the WESTDC) was further updated using this dynamic algorithm. It has been downloaded more than 550 times (one file contains one year of snow depth data) by 337 registered users and has been used in many fields of research, such as spatiotemporal distribution of snow cover, climate change, ecologic change, cold region water cycle, and hydrological modeling.^108–110

Detailed snow information was considered in the algorithm developed by Dai et al.¹⁰² The authors investigated various properties of snow during different periods of a snow season and analyzed the seasonal snow depth and density variation based on 10 years of station data; finally, they obtained a priori snow characteristics (stratigraphy, snow grain size, density, and snow temperature) for Northwest China. The microwave emission model of layered snowpacks (MEMLS) model was used to simulate the snow brightness temperature based on the a priori snow characteristics; reference tables were then established showing the relationship between brightness temperature and snow depth. By considering detailed snow characteristics, the retrieval results were more accurate than other equivalent snow depth and snow water products.

When the dynamic algorithm developed in Northwest China was applied to Northeast China, the snow depth was underestimated due to the presence of large forested areas. The attenuation due to the presence of forest, and the associated upwelling radiation, can decrease the scattering signal from snow cover, causing an underestimation of snow depth in forested areas.^111–113 In order to eliminate the influence of forest on snow depth retrieval and improve retrieval accuracy in Northeast China, an optimal iterative method was used. Forest transmissivities at 18 and 36 GHz based on the microwave radiative transfer models of forest and snow were calculated, as well as snow properties measured from field experiments.¹¹⁴ The results indicated that the transmissivities in forested areas in Northeast China are 0.656 and 0.895 at 18 and 36 GHz, respectively. The reference tables linking brightness temperature and snow depth in forested areas were built, which allowed for the estimation of the snow depths.

Topography is another factor that influences snow depth retrieval using passive microwave data. Jiang combined the Dense Media Radiative Transfer Theory and Advanced Integral Equation Model (AIEM) to simulate brightness temperatures of snowpacks.¹¹⁵ Interface roughness was considered in the AIEM, which was then used to simulate the reflectivity on the soil/snow interface.

Since the electromagnetic wave absorption properties of water, the volume scattering of the snowpack is removed, and wet snow cannot be monitored by passive microwave. The strong absorption properties of wet snow are utilized by the synthetic aperture radar (SAR) to efficiently discern wet snow from other types of snow.¹¹⁶ In the mid-1990s, Li et al.¹¹⁷ examined the capability of snow-cover mapping with multifrequency and multipolarized SAR images from SIR-C.¹¹⁷ Sun et al.¹¹⁸ used multitemporal ENVSAT-ASAR images to characterize the seasonal variations of snow-covered landscapes in the upper reaches of the Heihe River Basin and retrieved the wet snow cover with a $-2\mathrm{dB}$ threshold. Wen et al.¹¹⁹ developed a snow depth retrieval algorithm based on ENVISAT-ASAR data by considering the microwave radiative transfer and the interaction between microwave radiation and the snow/soil interface. Moreover, the launch of interferometric SAR allowed for the mapping of snow cover using repeat pass SAR images.¹²⁰ However, most of the current methods can only retrieve wet snow cover area; a method to map total snow cover area has yet to be developed.

3.5.

Snow Data Assimilation

Snow data assimilation provides a method to integrate all available observations related to snow variables into a land surface model to improve the estimation of variables, such as snow cover area, snow depth, and snow water equivalence. Because microwave remote sensing data can derive snow depth/SWE and optical remote sensing data can accurately obtain snow cover area, snow data assimilation usually focuses on how to assimilate microwave data on the basis of a radiative transfer model of snow and how to assimilate satellite-based snow cover area on the basis of an empirical relationship between snow cover area and snow depth. In China, the study of snow data assimilation is very limited and mainly conducted at CAREERI, CAS. With respect to snow radiance assimilation, the MEMLS has been integrated into the Chinese Land Data Assimilation System (CLDAS).¹²¹ The CLDAS is capable of assimilating passive microwave brightness temperatures (i.e., AMSR-E and SSM/I) to improve snow depth and SWE estimates on the basis of ensemble Kalman filtering (EnKF). Huang et al.¹²² analyzed the impacts of snow grain size and stratigraphy on snow radiance assimilation. They then developed an ensemble-based snow radiance assimilation method using an ensemble batch smoother to simultaneously characterize the SWE and soil freeze-thaw state by assimilating multifrequency active/passive microwave data.¹²³ Che et al.¹²⁴ developed a snow data assimilation scheme that directly assimilates passive microwave brightness temperature data into a snow process model. Results show that the data assimilation system can improve the estimation of snow depth in the accumulation period, but not in the ablation period. To better assimilate satellite-based snow cover products, Huang¹²⁵ developed and validated a technique to assimilate daily MODIS SCF products with an ensemble Kalman filter. Applied in the Xinjiang Province, the assimilation of MODIS SCF with an ensemble Kalman filter gave reasonable SCF values and improved snow depth estimation. Li et al.¹²⁶ assimilated MODIS snow cover area data and microwave snow depth data into a physically based snow model, using ensemble Kalman filtering to simulate snow distribution in regions where snow is patchy and shallow. Results show this method is applicable to other ungauged regions.

4.1.

Hydrological Remote Sensing in Cold Regions: a Ground-Based Synchronous Observation Experiment in the Upper Reaches of the Heihe River

This experiment was one of three components of the Watershed Allied Telemetry Experimental Research (WATER) project. WATER was a multiscale land surface/hydrological experiment conducted in a cold and arid region in China. It was a simultaneous airborne, satellite-borne, and ground-based remote sensing experiment aiming to improve the observation ability, understanding, and predictability of hydrological and related ecological processes at a catchment scale.¹²⁸ In cold regions, hydrological remote sensing and a ground-based synchronous observation experiment in the upper reaches of the Heihe River were implemented from July 2007 to December 2009. The aim of the experiment was to understand hydrological processes and enhance the level of quantitative remote sensing in cold regions. Based on airborne remote sensing and ground-based tests, satellite-based remote sensing methods were developed and improved to extract hydrological process parameters for cold regions. The experiment was carried out in three areas (Binggou watershed, Arou grassland, and Biandukou) where land cover ages differed. Quantifying snow cover and frozen soil variables and parameters was the primary goal of the experiment (Table 1). Synchronous experiments occurred in four areas of different scales: watershed, sub-basin, hydrological unit, and pixel. In these areas, dense ground synchronous observation was jointly carried out with observations of fluxes and meteorological and hydrological elements. Aviation sensors containing a microwave radiometer, a hyperspectral imagery camera, a thermal infrared imagery camera, and a multispectral CCD camera collected a wealth of visible/near infrared, thermal infrared, active and passive microwave satellite data over the experimental areas (Tables 2 and 3). Through the experiments, an air-satellite-ground-based integrated dataset was built for the upper reaches of the river, which can be used to improve and validate land surface/hydrological process models in cold regions.

Table 1

The ground observation items in the simultaneous observation experiment.

Table 2

The information of airborne remote sensing in the simultaneous observation experiment.

Table 3

The information of satellite remote sensing in the simultaneous observation experiment.

This experiment is the first remote sensing experiment in China to focus on snow and frozen soil, and it has had wide participation from snow remote sensing scientists in China. Twenty-six ground-based remote sensing sensors, with more than 100 supplementary instruments, were used to observe and monitor the land surface properties of snow and its variations.

4.2.

Cooperative Observation Series for Snow

The snowfield measurement experiments were initially designed to investigate snow properties of typical snow areas in China, such as the North Xinjiang in China, Qinghai–Tibet Plateau, and the Northeast of China. The experiments are composed of a series of long-term snow surveys, primarily conducted every year, with joint collaboration from different universities and institutes. Thus, the snowfield measurement experiments were termed the COSS. The first experiment was carried out in the daytime at two snow fields, located in Fukang and Fuyun Counties of the North Xinjiang from December 1, 2010, to December 15, 2010. The purpose of this experiment was to develop retrieval methods of snow properties based on the quantitative understanding of measurements. First, blending methods of optical and passive microwave remote sensing techniques on snow were explored using multiple observations coincident in time and location. Second, the passive microwave signal responding to heterogeneous snow depth was measured, and the relationship was studied. In addition, quantitative observations of interactions between the snow optical spectrum, snow grain size, snow albedo, snow dust, and black carbon (BC) were made. As we know, optical and microwave signal responses from remote sensors are influenced by factors such as snow depth, density, wetness, crystal size and shape, and ice crusts, among others. To better understand these factors, we measured snow layer properties in snow pits, including snow depth, snow density, wetness, snow grain size, snow temperature, snow albedo, snow spectral, and snow microwave brightness temperatures. Snow samples were collected to analyze the content of snow dust and snow BC (Table 4). All of the airborne datasets described here and detailed information for data collection and processing are available online ( http://westdc.westgis.ac.cn/).

Table 4

Snow properties, measurement methods and objectives on Cooperative Observation Series for Snow (COSS).

In order to quantify the influence of forested areas on the estimation of snow depth, fieldwork was carried out in northeast China in January and March 2012 and in January 2013. Each time, snow transects through both forested and nonforested regions were designed, and snow properties along the transects were measured, including snow depth, snow density, grain size, temperature, and stratigraphy.