The Evaluation of Snow Depth Simulated by Different Land Surface Models in China Based on Station Observations

Abstract

Snow plays an important role in catastrophic weather, climate change, and water recycling. In order to analyze the ability of different land surface models to simulate snow depth in China, we used atmospheric forcing data from the China Meteorological Administration (CMA) Land Data Assimilation System (CLDAS) to drive the CLM3.5 (the Community Land Model version 3.5), Noah (NCEP, OSU, Air Force and Office of Hydrology Land Surface Model), and Noah-MP (the community Noah land surface model with multi-parameterization options) land surface models. We also used 2380 daily snow-depth site observations of CMA to analyze the simulation effects of different models on the snow depth in China and different regions during the periods of snow accumulation and snowmelt from 2015 to 2019. The results show that CLM3.5, Noah, and Noah-MP can simulate the spatial distribution of the snow depth in China, but there are some differences between the models. In particular, the snow depth and snow cover simulated by CLM3.5 are lower than those simulated by Noah and Noah-MP in Northwest China and the Tibetan Plateau. From the overall quantitative assessment results for China, the snow depth simulated by CLM3.5 is underestimated, while that simulated by Noah is overestimated. Noah-MP has the best overall performance; for example, the biases of the three models during the snow-accumulation periods are −0.22 cm, 0.27 cm, and 0.15 cm, respectively. Furthermore, the three models perform differently in the three snowpack regions of Northeast China, Northwest China, and the Tibetan Plateau; Noah-MP has the best snow-depth performance in Northeast China, while CLM3.5 has the best snow-depth performance in the Tibetan Plateau region. Noah-MP performs best in the snow-accumulation period, and Noah performs best in the snowmelt period for Northwest China. In conclusion, no single model can perform optimally for snow simulations in different regions of China and at different times of the year, and the multi-model integration of snow may be an effective way to obtain high-quality snow simulation results. So this study provides some scientific references for the spatiotemporal evolution of snow in the context of climate change, monitoring and analysis of snow, the study of land surface models for snow, and the sustainable development and utilization of snow resources in China and other regions.

1. Introduction

As an important physical quantity in the Earth system, snow, with its high albedo and low thermal-conductivity properties, affects the surface energy balance and thus has an impact on the regional and global weather and climate [1,2,3]. The accumulation and melting of snow also affect the water cycle; the reduction in seasonal snow in particular has a certain impact on the development of agriculture and livestock, crop growth, and natural ecology in arid and semi-arid areas [4,5,6]. Furthermore, the reduction in permanent snow is an indicator of global warming, warning people to take action as soon as possible to mitigate climate change from a sustainable development perspective [7,8]. And in the context of global warming and the increase in extreme weather and climate events, the frequency of blizzards has increased, and the accurate monitoring of snow depth can contribute to improving the level of early warning and prevention of snowstorms [5,9,10]. At present, the snow depth is mainly determined through station observations, satellite remote sensing, and model simulations. There are more than 2400 operational snow-depth sites observed by the China Meteorological Administration (CMA), which are divided into manual and automatic station observations. The site observations can better indicate the snow depth of the location, but the spatial representation is poor [11]. Satellite remote sensing can better capture the distribution of snow than snow-depth station observations; for example, the Moderate-Resolution Imaging Spectroradiometer (MODIS) has the advantages of high spectral resolution and wide spatial coverage, and it produces a global daily snow cover product [12]. But the satellite remote sensing has poor penetration ability for snow depth and is influenced by the inversion algorithm used, the vegetation coverage, and clouds [12,13,14]. For example, although the Feng Yun 4 satellite snow data has a high spatial and temporal resolution, the snow information can only be reflected in non-cloudy areas during the daytime. Therefore, the land-surface model with physical processes, dynamics, and thermodynamics that best approximate the accumulation–melting of snow is a possible way to determine the snow depth in a spatiotemporal continuum [15,16,17,18].

Current research on model-simulated snow depths includes improvements to the model parameters, atmospheric forcing data, and the evaluation of different model simulations or the reanalysis of snow datasets and their variable characteristics across China. One way to enhance the model parameterization is to improve the land parameters and the physical scheme. For example, Zhang et al. [19] found that the Noah-MP adds a parameterization for soil organic matter, which can improve the soil temperature and heat fluxes, thus alleviating the cold bias of snow depth at high latitudes. Xie et al. [20] analyzed the effect of snow distribution and the surface energy balance on the Tibetan Plateau using two snow schemes (NY07 and SL12) in the CLM, showing that both snow schemes capture the distribution of the maximum snow depth well but show a large positive bias in the mean values over all periods. Yang et al. [21] used the real-time updated leaf area index, a green vegetation fraction, and land cover based on WRF/Noah-MP for snow simulation in the Tianshan Mountains, and the results show that more realistic vegetation parameters could improve the performance of snow simulation, especially in forested areas. Improving the atmospheric forcing data or comparing simulations under different atmospheric forcing datasets can potentially improve snow simulations. For example, Zhang et al. [18] used the improved before-and-after CLDAS data to drive the Noah model to simulate the snow in China and found that improvements in the winter precipitation could better enhance the simulation effect of snow. Gao et al. [22] used CMFD and ERA-Interim data to drive the Noah-MP model to simulate the snow in the Ordos River basin, and the results showed that the snow depth simulated under CMFD was better than that under ERA-Interim. In their evaluation of snow simulated by different models or the reanalysis of datasets and their variability characteristics, Liu et al. simulated a snowfall event on the Tibetan Plateau based on WRF, where the CLM, Noah, and Noah-MP models were selected for sensitivity analysis for the land surface process scenario; the results showed that CLM and Noah-MP were more accurate for the snow water equivalent [23]. Li et al. [16] evaluated the ERA5, ERA5_land, microwave inversion snow depth, and WRF-simulated snow depth in the Tianshan region using site observations, and they showed that there was less bias in the WRF-simulated snow depth. Orsolini et al. [24] evaluated the ERA5 reanalysis, ERA-Interim reanalysis, JRA reanalysis, and MERRA2 reanalysis in the Tibetan Plateau using snow-depth site observations and showed that the reanalysis values were overestimated, with MERRA-2 and JRA-55 having the best relative results.

It can be seen that the main influences on the snow-simulation results of the land surface model include the parameterization schemes, land parameters, atmospheric forcing data, and different land surface model types [15,25,26]. However, there are few comparative analyses of the effects of different land surface models on the simulation of snow depth in China, and most studies have focused on parameterization schemes [20,22,27], land surface parameters [19,21,25], and atmospheric forcing data [18,25,28,29]. Therefore, we use CLDAS atmospheric forcing data to drive the CLM3.5, Noah, and Noah-MP land-surface models to simulate the snow depth in China at a spatiotemporal resolution of 6.25 km/d from 2015 to 2019. The snow depths simulated using different land surface models are evaluated based on CMA observational snow-depth data, and the models’ performance in different regions is analyzed to develop the snow simulation and improve the analysis of the spatiotemporal evolution of snow in the context of climate change in China.

2. Data and Methods

2.1. The Snow-Depth Observational Data

The snow-depth observations of the CMA are calculated as the vertical depth from the snow surface to the ground, which mainly includes manual observations and automatic station observations, where the snow depth of the manual observations mainly adopts the snow ruler or common ruler, and the snow depth of the automatic station observations mainly adopts an ultrasonic snow-depth detector. More than 2400 snow depth observations are currently available from the China Meteorological Administration, and we selected 2380 stations’ snow depth observations (Figure 1) from 2015 to 2019 that are relatively stable and under strict quality control for evaluation. These data are available from the China Meteorological Data Service Centre (http://data.cma.cn/ (accessed on 9 April 2023)).

2.2. Atmospheric Forcing Data

The atmospheric forcing data for the land surface model use the air temperature, pressure, wind speed, precipitation, humidity, and solar shortwave radiation of CLDAS with a time resolution of 1 h and a spatial resolution of 6.25 km [30]. The CLDAS datasets are developed by National Meteorological Information Center of CMA. The CLDAS air temperature, air pressure, specific humidity, and wind speed are based on the numerical forecast product of the European Center for Medium-Range Weather Forecasts (ECMWF) as the background field. A multi-grid variational analysis method was used to integrate nearly 50,000 observations of temperature, pressure, humidity, and wind from the China Meteorological Administration with the ECMWF background field. The CLDAS solar radiation data are inversed by the Feng Yun II satellite based on the DISORT radiative transfer model. The CLDAS precipitation data are obtained by fusing CMORPH or MERRA2 precipitation with more than 60,000 hourly precipitation observations of CMA using the multi-grid variational analysis method and the time-downscaling method for solid precipitation. CLDAS precipitation fully takes into account the information on solid precipitation from the artificial observation sites and the information on solid precipitation from the reanalysis data, effectively solving the problem of solid precipitation in winter, and it has better quality than the international similar products GLDAS and CMORPH [31]. CLDAS has been widely used in land-surface simulation [18,19,25,31,32]. Therefore, we choose CLDAS atmospheric forcing data for the simulation of snow in China. These datasets are available from the China Meteorological Data Service Centre (http://data.cma.cn/ (accessed on 9 April 2023)).

2.3. Land-Surface Model

In this study, we chose the CLM3.5, Noah, and Noah-MP land surface models. CLM incorporates the advantages of LSM, BATS, IAPS, and other models, not only incorporating parameterization such as surface runoff, vegetation dynamics, and biophysical–chemical processes, but also involving parameterized processes such as vegetation, soil, permafrost, moisture, and lakes. In particular, the snow parameterization is mainly based on Anderson et al. [33], Jordan et al. [34], and Dai and Zeng [35]. When the snow depth was greater than 0.01 m, the snow was divided into a maximum of five layers (−4, −3, −2, −1, 0), with layer 0 being near the soil surface and the number of snow layers varying with the snow depth. The state variables of snow include the water mass, the ice mass, the snow thickness, and the snow temperature, where the water’s phase change is neglected [36].

The Noah Land Surface Model was developed by OSU-LSM and was officially named “Noah” in 2000. The Noah snow parameterization scheme is based on the NCEP climate model snow cover and the permafrost parameterization scheme, which portrays the snow accumulation and melting process, taking into account the snow depth and density change process during snow aging, liquid water re-condensation, and snow compaction. Noah uses precipitation-forcing data to initialize the snow water equivalent, and the snow depth is influenced by the snow water equivalent, snow density, snow surface temperature, and soil surface temperature, which are calculated from the snow water equivalent and snow density [37]. The Noah model has been refined over the years and has been widely used in integrated simulations of land-surface processes and in many land-data-assimilation systems, such as the US Global Land Surface Data Assimilation System (GLDAS) and the North American Land Surface Data Assimilation System (NLDAS).

The Noah-MP land surface model is further developed from the Noah land surface model. Compared with the Noah model, Noah-MP has adjusted the overall framework of the model by separating the vegetation from the surface and improving the energy balance of the vegetation cover, snow accumulation, permafrost and infiltration, and soil moisture–groundwater interaction. The model can also provide thousands of combinations of parameterization schemes for multiple options of physical processes such as dynamic vegetation, runoff, and groundwater, and users can configure the parameterization schemes according to their needs; its default scheme is used in this study. For the simulations of snow, the Noah model looks at the soil and canopy as a whole. When the snow is thick, a large amount of energy is stored on the snow surface because the snow in the Noah model only considers a layer of snow and does not take into account the energy exchange between snow interiors, which decreases the snowmelt, surface temperature, and soil temperature for Noah [38]; in contrast, Noah-MP develops a new multi-layer physical snow-accumulation scheme and snow-interception model for this problem, which divides the snow into three layers. When the snow depth is less than 0.045 m, the snow layer does not exist, and the snow is combined with the topsoil layer; when it is greater than 0.045 m, it is layered, and the model has zero to three layers, depending on the snow depth, with a threshold of 0.05 m for the first layer and 0.18 m for the second layer, where the snow temperature is also introduced, thus representing the processes of snow infiltration, retention and refreezing, and energy transfer [38].

2.4. Experimental Design

In this study, we used CLDAS atmospheric forcing data to drive the CLM3.5, Noah, and Noah-MP land-surface models to simulate the snow in China at a spatial resolution of 6.25 km/day from 2015 to 2019. The results of the three sets of simulations were evaluated using the snow-depth stations of the CMA. The bias, root mean square error (RMSE), correlation coefficient (CORR), false alarm rate (FAR), probability of detection (POD), and threat score (TS) were calculated throughout China including in Northeast China, Northwest China, and the Tibetan Plateau during the snow-accumulation period (November, December, and January each year) and the snowmelt period (February, March, and April each year).

$$\mathrm{B}\mathrm{i}\mathrm{a}\mathrm{s}=\frac{1}{\mathrm{N}}\sum _{\mathrm{i}=1}^{\mathrm{N}}({\mathrm{S}\mathrm{i}\mathrm{m}}_{\mathrm{i}}-{\mathrm{O}\mathrm{b}\mathrm{s}}_{\mathrm{i}})$$

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{\sum _{\mathrm{i}=1}^{\mathrm{N}}{({\mathrm{S}\mathrm{i}\mathrm{m}}_{\mathrm{i}}-{\mathrm{O}\mathrm{b}\mathrm{s}}_{\mathrm{i}})}^2}{\mathrm{N}}}$$

$$\mathrm{C}\mathrm{o}\mathrm{r}\mathrm{r}=\frac{\sum _{\mathrm{i}=1}^{\mathrm{N}}\left({\mathrm{S}\mathrm{i}\mathrm{m}}_{\mathrm{i}}-\overline{\mathrm{S}\mathrm{i}\mathrm{m}}\right)\left({\mathrm{O}\mathrm{b}\mathrm{s}}_{\mathrm{i}}-\overline{\mathrm{O}\mathrm{b}\mathrm{s}}\right)}{\sqrt{\sum _{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{S}\mathrm{i}\mathrm{m}}_{\mathrm{i}}-\overline{\mathrm{S}\mathrm{i}\mathrm{m}}\right)}^2}\sqrt{\sum _{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{O}\mathrm{b}\mathrm{s}}_{\mathrm{i}}-\overline{\mathrm{O}\mathrm{b}\mathrm{s}}\right)}^2}}$$

where $\mathrm{N}$ is the total number of samples, ${\mathrm{S}\mathrm{i}\mathrm{m}}_{\mathrm{i}}$ is the simulated value, ${\mathrm{O}\mathrm{b}\mathrm{s}}_{\mathrm{i}}$ the observed value, and $\overline{\mathrm{S}\mathrm{i}\mathrm{m}}$ and $\overline{\mathrm{O}\mathrm{b}\mathrm{s}}$ are the mean of the simulated and observed values, respectively.

$$\mathrm{P}\mathrm{O}\mathrm{D}=\frac{\mathrm{h}\mathrm{i}\mathrm{t}\mathrm{s}}{\mathrm{h}\mathrm{i}\mathrm{t}\mathrm{s}+\mathrm{m}\mathrm{i}\mathrm{s}\mathrm{s}}$$

$$\mathrm{F}\mathrm{A}\mathrm{R}=\frac{\mathrm{f}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}}{\mathrm{h}\mathrm{i}\mathrm{t}\mathrm{s}+\mathrm{f}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}}$$

$$\mathrm{T}\mathrm{S}=\frac{\mathrm{h}\mathrm{i}\mathrm{t}\mathrm{s}}{\mathrm{h}\mathrm{i}\mathrm{t}\mathrm{s}+\mathrm{m}\mathrm{i}\mathrm{s}\mathrm{s}+\mathrm{f}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}}$$

where hits is the number of samples hit, miss is the number of samples missed, and false is the number of samples misreported.

3. Results

3.1. Spatial Distribution of the Snow Depth

To compare the simulation effects of different models on the spatial distribution of the snow depth, we calculated the average values of the snow depths for the CMA observations, CLM3.5, Noah, and Noah-MP during periods of snow accumulation and snowmelt from 2015 to 2019 (Figure 2). Given the spatial distribution of the snow depth during the snow-accumulation period (Figure 2a–d), it can be seen that CLM3.5, Noah, and Noah-MP properly reflect the spatial distribution of the snow depth in Northeast China, Northwest China, and the Tibetan Plateau, among which Noah and Noah-MP have a more consistent spatial distribution of snow depth, while CLM3.5 has less snow cover in the Tibetan Plateau compared with the other models. In terms of the magnitude of snow depth during the snow-accumulation period, CLM3.5, Noah, and Noah-MP all overestimated the snow depth in the northern part of Northeast China relative to the observations, especially for the Noah model. As can be seen in the spatial distribution of the snowmelt period (Figure 2e–h), CLM3.5, Noah, and Noah-MP better reflect the spatial distribution of the snow depth during the snowmelt period in Northeast China, Northwest China, and the Tibetan Plateau, among which Noah and Noah-MP are more consistent in their snow cover values, and CLM3.5 has a relatively lower amount of snow cover in the Tibetan Plateau. In terms of the magnitude of the snow depth during the snowmelt period, Noah has a higher snow depth than CLM3.5 and Noah-MP in Northwest China.

3.2. Snow-Depth Error Time Series

The snow depths of CLM3.5, Noah, and Noah-MP were quantified using quality-controlled CMA observations from 2015 to 2019 in the snow accumulation and snowmelt periods; the bias, RMSE, CORR, FAR, POD, and TS values were calculated throughout China.

Firstly, we calculated the time series of the bias, RMSE, CORR, FAR, POD, and TS in the snow-accumulation periods (January, November, and December), among which the bias, RMSE, and CORR properly reflect the simulated snow-depth magnitude error for the three models. It can be seen from the bias time series (Figure 3a) that the bias of CLM3.5, Noah, and Noah-MP are all between −2 and 1 cm, among which CLM3.5 presents a negative bias, with a maximum negative bias of −1.98 cm and an average value of −0.22 cm. The Noah and Noah-MP models present positive biases, with an average of 0.27 cm and a maximum of 1.13 cm for Noah, and an average of 0.15 cm and a maximum of 1.37 cm for Noah-MP. Thus, Noah-MP has the smallest bias among the three models.

It can be seen from the RMSE time series (Figure 3b) that the RMSE of the snow depth simulated using the three models had a certain monthly variation, with that in November being smaller than that in December and January; the minimum value of the RMSE was 0.24 cm, the average value was 1.96 cm, and the maximum value was 5.37 cm for CLM3.5. The minimum RMSE value was 0.13 cm (0.13 cm), the average value was 1.85 cm (1.75 cm), and the maximum value was 3.72 cm (4.62 cm) for Noah (Noah-MP). Therefore, Noah-MP had the smallest RMSE among the three models.

From the time series of the correlation coefficients (Figure 3c), the CORR of all three models was low in early November each year, with the lowest CORR being 0.009, the average value being 0.73, and the maximum value being 0.94 for CLM3.5. The lowest CORR value was 0.161 (0.08), the average CORR value was 0.83 (0.82), and the maximum CORR value was 0.95 (0.94) for Noah (Noah-MP). Therefore, it can be seen from bias that CLM3.5 underestimates the snow depth, and Noah and Noah-MP overestimate the snow depth. Overall, Noah-MP is relatively better than both CLM3.5 and Noah for snow-depth simulations in the snow-accumulation periods.

The FAR, POD, and TS reflect the error in the simulated snow cover for the three models. From the time series of FAR (Figure 3d), the false alarm rates of all three models in early November were larger, in which the average was 0.17 and the maximum was 0.83 for CLM3.5, the average was 0.35 and the maximum was 0.87 for Noah, and the average was 0.28 and the maximum was 0.73 for Noah-MP. The FAR of CLM3.5 was the lowest, followed by Noah-MP and then Noah. It can be seen from the time series of POD (Figure 3c) that all three models had lower POD values in early November, with the average being 0.72 and the maximum being 0.97 for CLM3.5, the average being 0.95 and the maximum reaching 1 for Noah, and the average being 0.87 and the maximum reaching 1 for Noah-MP. Noah had the highest POD, followed by Noah-MP and then CLM3.5. It can be seen from the time series of TS (Figure 3f) that all three models had lower TS scores in early November, with an average of 0.62 and a maximum of 0.90 for CLM3.5, an average of 0.646 and a maximum of 0.90 for Noah, and an average of 0.65 and a maximum of 0.93 for Noah-MP; Noah-MP had the highest TS scores. Therefore, CLM3.5 underestimates snow cover and Noah overestimates snow cover Noah, and Noah-MP is relatively better than both Noah and CLM3.5.

Secondly, we calculated the time series of the bias, RMSE, CORR, FAR, POD, and TS in the snowmelt periods (February, March, and April). It can be found (Figure 4a) from the bias time series that the bias values of CLM3.5, Noah, and Noah-MP were all between −1.62 and 1.2 cm, among which CLM3.5 presented a negative bias, with a maximum negative bias of −1.62 cm and an average value of −0.2 cm. The biases of Noah and Noah-MP were around 0, with an average of 0.06 cm (−0.02 cm) and a maximum of 1.18 cm (1.3 cm) for Noah (Noah-MP). The bias for the three models in the snowmelt period was smaller than in the snowpack period, and Noah-MP had the smallest bias among the three models. Figure 4b shows that the three models had a specific monthly variation, with the RMSE in February being larger than that in March and April. The minimum value was 0.02 cm, the average was 1.96 cm, and the maximum was 4.65 cm for CLM3.5; the minimum was 0.06 cm, the average was 1.75 cm, and the maximum was 3.81 cm for Noah; and the minimum was 0.02 cm, the average was 1.64 cm, and the maximum was 4.09 cm for Noah-MP. Noah-MP had the smallest RMSE among the three models, which was consistent with the RMSE of the snow depth during the snow-accumulation period. All three models show low CORR from mid-March to April each year from the time series of CORR (Figure 4c), which may be related to the snow-melting parameterization scheme of the models. For CLM3.5, the lowest CORR was 0, the average was 0.60, and the maximum was 0.99; for Noah, the lowest CORR was 0.1, the average was 0.68, and the maximum was 0.94; and for Noah-MP, the lowest CORR was 0.02, the average was 0.67, and the maximum was 0.96. The Noah model had the best CORR, which was consistent with the snow depth CORR in the snow-accumulation period. Therefore, it can be seen that the errors in the three models are generally smaller in the snowmelt period than in the snowpack period, with CLM3.5 underestimating the snowpack depth and Noah and Noah-MP slightly overestimating the snow depth. Noah-MP was also relatively better than CLM3.5 and Noah for snow-depth simulations in the snowmelt period.

It can be seen from the time series of FAR that the three models had larger FAR values from mid-March to April (Figure 4d), in which the average was 0.23 and the maximum could reach 1 for CLM3.5; the average was 0.43, and the maximum could reach 1 for Noah, and the average was 0.3 for Noah-MP. All three models seriously overestimated the snow cover during the snowmelt period, with CLM3.5 having the lowest FAR, followed by Noah-MP; the FAR of Noah was the largest. Compared with the time series of POD (Figure 4c), the three models had lower PODs from mid-March to April, where the average was 0.51 and the maximum was 1 for CLM3.5, the average was 0.83 and the maximum could reach 1 for Noah, and the average was 0.60 and the maximum was 1 for Noah-MP. The POD in the snowmelt period was, thus, lower than that in the snow-accumulation periods, with Noah having the highest POD, followed by Noah-MP. All three models had lower TS scores from mid-March to April (Figure 4f), with an average of 0.45 and a maximum of 1 for CLM3.5, an average of 0.51 and a maximum of 0.93 for Noah, and an average of 0.47 and a maximum of 0.91 for Noah-MP. Therefore, the snow-cover simulations of the three models all produced overestimates, especially from mid-March to April; Noah-MP performed relatively better than Noah and CLM3.5.

3.3. Assessment of the Snow Depth in Three Major Snow Areas

Quantitative assessments of CLM3.5, Noah, and Noah-MP were performed using CMA snow-depth observations from 2015 to 2019 in Northeast China, Northwest China, and the Tibetan Plateau, and the RMSE values were calculated for the simulated snow depth of the three models during the snow-accumulation and snowmelt periods.

The RMSE time series and their statistical values for Northeast China, Northwest China, and the Tibetan Plateau in the snow-accumulation periods are given in Figure 5. Firstly, the RMSE for the three models showed a gradual increase, with January being higher than December and December being higher than November in Northeast China (Figure 5a,b), which was related to the gradual increase in snow depth. The lowest RMSE was 0.11 cm, the average was 3.95 cm, the median was 4 cm, and the standard deviation was 2.20 cm for CLM3.5. The lowest RMSE was 0.15 cm, the average was 4.06 cm, the median was 4.19 cm, and the standard deviation was 1.99 cm for Noah, while the lowest RMSE was 0.16 cm, the average was 3.94 cm, the median was 3.92 cm, and the standard deviation was 1.66 cm for Noah-MP. In comparison, the snow depth simulated by Noah-MP and CLM3.5 in Northeast China are basically similar and are better than those of Noah.

Secondly, the RMSE of the snow depth in Northwest China also showed a gradual increase, with January being higher than December and December being higher than November, which is related to the gradual increase in snow depth (Figure 5c,d). For CLM3.5, the lowest RMSE was 0.01 cm, the average was 6.91 cm, the median was 7.10 cm, and the standard deviation was 2.62 cm. For Noah, the lowest RMSE was 0.12 cm, the average was 5.27 cm, the median was 5.16 cm, and the standard deviation was 2.56 cm, while for Noah-MP, the lowest was 0.07 cm, the average was 4.69 cm, the median was 4.75 cm, and the standard deviation was 2.25 cm. In comparison, the effects of Noah-MP and CLM3.5 in Northeast China were similar to and better than those of Noah. The RMSE of the three models in Northwest China was greater than that in Northeast China, with Noah-MP having the smallest RMSE, followed by the Noah model and then CLM3.5.

Thirdly, the RMSE values of the three models in the Tibetan Plateau are smaller than those in Northeast China and Northwest China (Figure 5e,f). The average RMSE was 1.06 cm, the median was 0.72 cm, and the standard deviation was 1.17 cm for CLM3.5. The average RMSE was 1.56 cm, the median was 0.7 cm, and the standard deviation was 2.02 cm for Noah, while the average was 1.1 cm, the median was 0.8 cm, and the standard deviation was 1.07 cm for Noah-MP. In comparison, the snow depth of CLM3.5 is the most accurate, followed by Noah-MP and then Noah, which performed the worst in this region.

Therefore, the models are the best in the Qinghai–Tibet Plateau followed by Northeast China, and the worst in Northwest China. For different models in the same region, Noah-MP performed best for snow depths in Northeast China and Northwest China, and CLM3.5 performed best in the Tibetan Plateau region.

The RMSE time series and their statistical values for Northeast China, Northwest China, and the Tibetan Plateau in the snowmelt periods are given in Figure 6. Firstly, it can be seen that the RMSE of the three models showed a gradually decreasing trend, with April being lower than March and March being lower than February, which was related to the gradual melting of snow in Northeast China (Figure 6a,b). The average RMSE was 3.43 cm, the median was 2.2 cm, and the standard deviation was 3.28 cm for CLM3.5. The average RMSE was 3.02 cm, the median was 2.18 cm, and the standard deviation was 2.62 cm for Noah, while the average was 2.9 cm, the median was 2.58 cm, and the standard deviation was 2.45 cm for Noah-MP. In comparison, the snow depth of Noah-MP was the best, followed by that of the Noah model and then CLM3.5.

Secondly, the RMSE of the snow depth in Northwest China showed a gradually decreasing trend, with April being lower than March and March being lower than February, which is related to the gradual melting of snow (Figure 6c,d). For CLM3.5, the average RMSE was 6.57 cm, the median was 6.11 cm, and the standard deviation was 3.93 cm. For Noah, the average RMSE was 5.29 cm, the median was 4.63 cm, and the standard deviation was 3.52 cm, while for Noah-MP, the average was 5.54 cm, the median was 5.43 cm, and the standard deviation was 3.92 cm. In comparison, the RMSE of the snow depth simulated by the three models was greater in Northwest China than in Northeast China. Noah had the best effect, followed by the Noah-MP model and then CLM3.5.

Thirdly, the RMSE of the three models in the Tibetan Plateau is smaller than that in Northeast China and Northwest China (Figure 6e,f) from 2015 to 2018, but the RMSE suddenly increased in 2019, which may be related to the quality of the CLDAS atmospheric forcing data. For CLM3.5, the average RMSE was 1.94 cm, the median was 0.79 cm, and the standard deviation was 2.66 cm. For Noah, the average RMSE was 2.73 cm, the median was 1.04 cm, and the standard deviation was 3.59 cm, while for Noah-MP, the average was 2.27 cm, the median was 0.91 cm, and the standard deviation was 2.87 cm. In comparison, the snow depth of CLM3.5 was the best, followed by Noah-MP and then Noah.

Therefore, for the same model in different regions, the snow depths of the three models are better in the Tibetan Plateau than in Northeast China, and the snow depths were better in Northeast China than in Northwest China, with the models performing in the same manner as they did during the snow-accumulation period. For the different models in the same region, the three models had their own advantages in different regions in the snowmelt period; for the snow depths, Noah-MP performed best in Northeast China, Noah performed best in Northwest China, and CLM3.5 performed best in the Tibetan Plateau.

Figure 7 gives the statistical values of FAR, POD, and TS for the different models during the snow-accumulation period and snowmelt period in the three major snow areas of China. The FAR values of the three models in Northeast China and Northwest China were smaller than those in the Tibetan Plateau, and the POD and TS in Northeast China and Northwest China were better than those in the Tibetan Plateau during the snow-accumulation and snowmelt periods. In terms of the performance of different models during the snow-accumulation period (Figure 7a–c) in Northeast China, for Noah, Noah-MP, and CLM3.5, respectively, the FAR values were 0.20, 0.19, and 0.16; the POD values were 0.99, 0.98, and 0.90; and the TS values were 0.79, 0.78, and 0.77. In Northwest China, CLM3.5 had the smallest FAR, but its POD and TS were lower than in the other two models, and the POD, FAR, and TS values for Noah were higher than those for Noah-MP, indicating that the snow cover was better simulated in Noah than in Noah-MP and CLM3.5. In the Tibetan Plateau, CLM3.5 has the best FAR, but its POD and TS were lower than in the other two models. The FAR values of Noah and Noah-MP were similar, while the POD and TS values of Noah were higher than those of Noah-MP, indicating that the snow cover simulated by Noah was better than that of the other two models.

From the performance of the different models during the snowmelt periods (Figure 7d–f), Noah had the largest FAR in both Northeast China and Northwest China, but its POD and TS were best in those regions, and the FAR (POD and TS) of Noah-MP (CLM3.5) was greater than that of CLM3.5 (Noah-MP) in both Northeast China and Northwest China. For the Tibetan Plateau, the FAR and POD of the three models indicated that Noah had the highest POD and the largest FAR, and the FAR of Noah-MP was higher than that of CLM3.5, while its POD was better than that of CLM3.5. The TS of CLM3.5 (0.376) was better than that of Noah (0.366) and Noah-MP (0.354).

Therefore, for the FAR, POD, and TS, the CLM 3.5 model was the lowest and the Noah model was the highest; according to the comparison of the snow cover in different regions, the three models had the best simulation effect in Northeast China and Southwest China and the worst in the Tibetan Plateau.

4. Discussion

In the context of global warming, the frequency of local extreme snowstorms is increasing, which seriously affects people’s lives, transportation, agriculture, animal husbandry, etc. On the other hand, as the temperature increases due to global warming, the snow cover, snow days, and snow depth in some regions of China are also decreasing [12,39,40]; snow changes have a definite impact on the local climate, such as in the Tibetan Plateau. Therefore, the accurate simulation of snow is crucial to prevent and respond to snow disasters and improve the detection and prediction of climate change and the impact assessment for the sustainable development of people, nature, and society [10,41,42,43]. Now, snow simulations mainly suffer from errors in the model parameterizations and atmospheric forcing data [29].

Firstly, we analyzed the parameterizations of the land model. From the bias, RMSE, CORR, POD, FAR, and TS of snow in different regions, we can find that three models have advantages and disadvantages in simulating snow depth and snow cover in different regions and different periods in China; these advantages are mainly related to the models’ parameterization schemes. The snow parameterization in the models has evolved from a simple bulk layer to multiple layers to accommodate more physical processes, and the multi-layered snow facilitates the exchange of surface heat fluxes, the retention of melted liquid water, and infiltration and refreezing processes within the snowpack [38,44,45]. The snow in CLM3.5 was divided into a maximum of five layers, with layer 0 being near the soil surface and the number of snow layers varying with the snow depth. The snow in Noah-MP had a three-layer snow model and a snow-interception model, and the three-layer model represents the percolation, retention, and refreezing of meltwater within the snowpack. The snow in Noah only considers a bulk layer, and the vegetation canopy and snow surface are considered as a whole in the model’s structure. This “bulk layer” version does not account for liquid water within the snowpack, so it predicts more snow mass and longer snow seasons [44]. This is one of the reasons why the snow depth simulated using Noah in this study shows positive bias and high FAR, which are consistent with the conclusions of the Noah model’s snow-depth simulation by Yang et al. [44]. Niu et al. [38] found that the improvements in snow simulations using Noah-MP may be attributed to changes in the model layers more so than those using Noah. In addition, the surface albedo plays a controlling role in the surface energy budget, especially in snow-covered areas that will be involved in the ablation of snow, so the snow-albedo-parameterization scheme is critical for snow simulation. The snow surface albedo in CLM3.5 is taken from BATS [36]. The Noah-MP includes both the BATS and the CLASS snow surface albedo options, and we selected the default CLASS scheme [38]. The snow surface albedo in Noah is calculated using a linear weighting between the nonsnow background albedo on the snow-free fraction and a blended snow-vegetation albedo on the snow-covered fraction [46]. Xie et al. [23] found that the advanced albedo-parameterization scheme in CLM seems to be a potentially important factor in improving snow simulation. Niu et al. [38] found that the Noah-MP captures the surface albedo peaks and recessions by using the CLASS scheme, which greatly improves the simulation of snow depth compared to Noah. Liu et al. [47] found that the improved snow albedo parameterization scheme in Noah can reduce the RMSE of snow depth by 21% in the Tibetan Plateau. Therefore, the multi-model integration of snow can be considered in the analysis of the spatiotemporal evolution of long-series snow, reducing the uncertainty in the climatic analysis of snow. There have been more advances in the integration of multi-land-surface models for soil moisture and evapotranspiration, such as the Global Soil Wetness Project (GSWP) [48], the International Land Model Bench-marking (ILAMB) Project [49], and multi-model integration methods, such as Bayesian and machine learning, which have been explored to improve simulation accuracy [25]. Therefore, subsequent studies will fully consider the error of different models for snow depth in different regions and periods, and different weights will be set for each model in different regions for multi-model snow integration to reduce uncertainty [50].

Furthermore, the quality of atmospheric forcing data is also crucial [51]. Numerous previous studies have shown that precipitation and temperature are important factors influencing snow accumulation and melting [25,29,31]. Sun et al. [31] optimized the CLDAS precipitation and found that the improvement in CLDAS precipitation forcing data in winter could effectively improve the simulation of snow depth, so we selected the optimized CLDAS precipitation to simulate the snow in this study. Wang et al. [29] used different atmospheric forcing datasets to drive the CAS-LSM model for snow and found that the snow depth and the snow water equivalent were most affected by the quality of atmospheric forcing data; the actual uncertainties in precipitation are much greater and should be strengthened to improve the quality of precipitation data. Liu et al. [25] analyzed the effect of different model-parameterization schemes, different forcing data, and different soil parameters of snow depth simulation, and found that the quality of atmospheric forcing is most important for snow depth, followed by soil texture, which can also enhance the simulation effect of snow depth. Furthermore, the solar radiation, longwave radiation, and wind speed from atmospheric forcing data are also important for snow accumulation, melting, and sublimation [15,45,52]. For example, solar radiation affected the rate of snow melting [15], and longwave radiation affected the time of snow melting [52]. Raleigh et al. [52] found that the availability of longwave radiation (i.e., observed vs. empirically estimated) caused maximum SWE differences up to 234 mm and up to 32 days difference in the timing of snow disappearance. Furthermore, the wind speed affects snow sublimation, and windblown snow is also present in areas with high wind speeds [45]. Zhang et al. [53] found that the uneven distribution of snow became more pronounced as the wind speed increased.

This study does have limitations in its evaluation of snow, such as the comparison between the observed snow depth at the site and the gridded snow depth of the model, as well as the comparison between the presence or absence of snow at the site and the snow cover in the model; these concerns have definite differences in spatial scale, and a comparative analysis between the snow cover produced via satellite inversion and model simulation can be considered in the future. With the development of satellite remote sensing, site observations, and other technologies, one of the most effective ways of obtaining high-quality snow datasets is making full use of the advantages of satellites, models, and site observations to carry out snow–land-data assimilation research. This can be achieved, for example, by assimilating snow cover to solve the problem of model simulations with or without snow and also by assimilating site observations to address problems with snow-depth simulation [15,18,54].

5. Conclusions

We used the CLDAS atmospheric forcing data to drive the CLM3.5, Noah, and Noah-MP land surface models for snow-depth simulations in China. And the spatial distribution, time series, and three major snow regions of the snow depth simulated by the three models were analyzed based on the 2380 CMA snow-depth station observations from 2015 to 2019. The main findings are as follows.

(1)

The CLM3.5, Noah, and Noah-MP models were able to simulate the spatial distribution of snow in China, but there were some differences in the magnitude of the simulated snow depth. In particular, the snow depth and snow cover simulated by CLM3.5 were lower than those of Noah and Noah-MP in Northwest China and the Tibetan Plateau, which was mainly related to the parameterization schemes of the models themselves.

(2)

From the overall evaluation across China, there was an underestimation of snow cover by CLM3.5 and an overestimation of snow cover by Noah in the snow-accumulation period. And the snow cover simulations of the three models all produced overestimates, especially from mid-March to April in the snowmelt period. Overall, the snow depth simulated by Noah-MP was better than that of CLM3.5 and Noah in China.

(3)

From the evaluation of regions, the snow depths simulated by the three models were better in the Tibetan Plateau than in Northeast China, and the snow depth simulated in Northeast China was better than in Northwest China. The snow depth simulated by Noah-MP was best in Northeast China, and the snow depth simulated by CLM3.5 was best in the Tibetan Plateau in the snow-accumulation and snowmelt periods. For Northwest China, Noah-MP simulated snow depth best in the snow-accumulation period, and the Noah model had the best snow-depth performance in the snowmelt period.

Therefore, different land surface models have their own advantages and disadvantages for snow simulation in different regions and time periods, which may be related to the snow-parameterization scheme and snowpack albedo scheme. So conducting multi-model integration studies of snow may be one of the effective ways to obtain high-quality snow datasets.