HTML
-
Recently, BCC-CSM has been upgraded to its second generation (Wu et al., 2020). The medium-resolution version (BCC-CSM2-MR) has carried out most of the CMIP6 DECK and MIP experiments (Eyring et al., 2016; Xin et al., 2019). Its previous generation (BCC-CSM1.1m; Wu et al., 2013, 2014) participated in CMIP5 (Taylor et al., 2012; Xin et al., 2012). Among the experiments endorsed by CMIP5 and CMIP6, historical simulation is one of the entry cards for models to participate in the project. The historical period is defined as beginning in 1850 and extends to near to the present day (2012 for CMIP5 and 2014 for CMIP6). The CMIP historical simulation provides a good opportunity to assess model ability in simulating climatic variability and trends. In this study, monthly surface (0–10 cm) SM data from the historical simulation of BCC-CSM2-MR are evaluated by comparing with 13 other CMIP6 models (Table 1) and BCC-CSM1.1m in CMIP5. The monthly precipitation data of the BCC-CSMs are also used, to analyze the coupling between SM and precipitation.
Model name Model center (or group) Spatial resolution (lat × lon) BCC-CSM2-MR Beijing Climate Center, China Meteorological Administration 160 × 320 CanESM5 Canadian Centre for Climate Modelling and Analysis 64 × 128 CESM2 Community Earth System Model Contributors 192 × 288 CESM2-WACCM Community Earth System Model Contributors 192 × 288 E3SM-1-0 U.S. Department of Energy’s Office of Biological and Environmental Research 180 × 360 EC-Earth3-Veg EC-Earth consortium (27 institutions in Europe) 256 × 512 FGOALS-f3-L LASG, Institute of Atmospheric Physics, Chinese Academy of Sciences 192 × 288 FIO-ESM-2-0 First Institute of Oceanography, Ministry of Natural Resources of China 192 × 288 GFDL-CM4 NOAA Geophysical Fluid Dynamics Laboratory 180 × 288 IPSL-CM6A-LR Institut Pierre-Simon Laplace 143 × 144 MIROC6 Atmosphere and Ocean Research Institute (The University of Tokyo) 128 × 256 MRI-ESM2-0 Meteorological Research Institute 160 × 320 NorESM2-LM Norwegian Climate Centre 96 × 144 SAM0-UNICON The National Renewable Energy Laboratory 192 × 288 Table 1. Details of the 14 CMIP6 models used in this study.
-
Representative of assimilated SM data products, GLDAS, which comprises ingested satellite- and ground-based observational data products and uses advanced land surface modeling and data assimilation techniques (Rodell et al., 2004), is widely used in studies of land–atmosphere interaction (Cheng et al., 2013; Zhang et al., 2016b). Therefore, we select the monthly data of near-surface (0–10 cm) SM from GLDAS v2.0 and v2.1 as the reference data in this study. In addition, we also use a station-based observational SM dataset in China produced by Wang and Shi (2019), to make the best use of the observations available at present. The SM measurements at a total of 1471 stations for the period January 1992 to September 2013 are from the National Meteorological Information Center of the China Meteorological Administration. Among these, 732 stations with good spatial and temporal continuity from January 1992 to December 2012 are ultimately selected following quality control processes. We use the data of the first layer (0–10 cm) and transform the units into kg m−2 for comparison. In the analysis of the coupling between SM and precipitation, we use the monthly data of precipitation from the CPC (Climate Prediction Center of NOAA) Merged Analysis of Precipitation (CMAP; Huffman et al., 1997).
-
The study region in this paper is Eurasia (10°–80°N, 30°–160°E) and the study period ranges from 1979 to 2012, in which we use the annual mean and the four seasons defined as December–January–February (DJF) for winter, March–April–May (MAM) for spring, June–July–August (JJA) for summer, and September–October–November (SON) for autumn. For convenience of comparison, both the simulation data and reference data are regrided to a uniform resolution of 1° × 1° using bilinear interpolation. Although the dataset of the 732 stations is quality controlled, there are still plenty of missing values in the spatial and temporal range. Therefore, it is inaccurate and inconvenient to interpolate the station dataset to the grid. Accordingly, when using the station dataset from 1992 to 2012 as the reference to make comparisons, we interpolate the simulation data in China to the positions of the 732 stations. Pearson correlation (r), root-mean-square error (RMSE), standard deviation and linear regression, the most commonly used metrics (Legates and McCabe, 1999), are used to quantify the agreement between reference data and model simulations. Besides, empirical orthogonal function (EOF) analysis and spectral analysis are performed to identify the major modes of the surface SM anomalies in Eurasia and its periodic characteristics.
2.1. Data
2.1.1. BCC-CSM historical simulations in CMIP5 and CMIP6
2.1.2. Reference data
2.2. Research region and diagnostic method
-
The climatological mean surface SM in terms of annual mean and the four seasons of GLDAS and BCC-CSM simulations are firstly presented (Fig. 1). Basically, except the less pronounced gradient in model simulations, both BCC-CSM2-MR and BCC-CSM1.1m are able to capture the spatial patterns of the SM of GLDAS, but BCC-CSM2-MR is closer to GLDAS in terms of the higher pattern correlation. GLDAS shows a triple pattern in the annual mean, whereby the surface soil with higher soil water content is typically located in central and western Siberia and South China, while that with lower soil water content is mainly distributed in central Asia and the Arabian Peninsula (Fig. 1a). From northwest to southeast, the SM in Eurasia exhibits a wet–dry–wet distribution, which is consistent with previous research (Nijssen et al., 2001; Guo et al., 2003; Erdenebat and Sato, 2018). This pattern is also evident in the BCC-CSM2-MR and BCC-CSM1.1m model simulations (Figs. 1b and 1c). The pattern correlation coefficient between BCC-CSM2-MR and GLDAS is 0.97 (at the 0.01 significance level), which is higher than that between BCC-CSM1.1m and GLDAS (r = 0.93). In different seasons, the pattern correlation coefficients of BCC-CSM2-MR and GLDAS range from 0.96 to 0.97, which are higher than those of BCC-CSM1.1m (between 0.92 and 0.94). The average pattern correlation coefficient between BCC-CSM2-MR and GLDAS is 0.97, which is a rise of 4.3% over the previous version. Therefore, BCC-CSM2-MR is more capable than BCC-CSM1.1m in its reproduction of the distribution of the climate mean surface SM.
Figure 1. Climate mean surface soil moisture (units: kg m−2) over Eurasia (1979–2012) on annual and seasonal time scales: (a, d, g, j, m) GLDAS as the reference data; (b, e, h, k, n) BCC-CSM2-MR and the pattern correlation coefficients with GLDAS; (c, f, i, l, o) BCC-CSM1.1m and the pattern correlation coefficients with GLDAS.
The difference in patterns of the climate mean SM between the BCC model simulations and GLDAS in terms of their annual mean and the different seasons are shown in Fig. 2. As the first two columns of the figure show, BCC-CSM2-MR and BCC-CSM1.1m both tend to reproduce soil that is drier than in GLDAS over the whole of Eurasia, with a large coverage of negative values, especially in southern China. Underestimations of SM in climate models have been demonstrated in many previous studies. For example, Ramillien et al. (2003) indicated that Land Dynamics hydrological model tends to underestimate the absolute water storage in the soil and provide smoother values than in-situ measurements. However, in several regions (Siberia, Northeast China, the Yangtze–Huaihe River basin, etc.), the results of the BCC-CSMs tend to be significantly wetter than those of GLDAS. Comparing Figs. 2a and b, we can see that the areas and absolute values of most regions with large differences have reduced from BCC-CSM1.1m to BCC-CSM2-MR in terms of their annual mean. Taking the positive value in Siberia as an example, the SM difference between BCC-CSM1.1m and GLDAS is greater than 20 kg m−2. However, it decreases to around 5 kg m−2 in BCC-CSM2-MR, with obvious improvement. This is also the case in the four seasons. Furthermore, compared to the previous-generation model, the RMSE values of BCC-CSM2-MR are lower both in terms of their annual mean and in the four seasons. The mean RMSE of the climate mean state declines by 7.5% from 9.37 (BCC-CSM1.1m) to 8.67 (BCC-CSM2-MR), which demonstrates that BCC-CSM2-MR is more capable of describing the actual distribution of the climate mean surface SM, with fewer biases. The differences between BCC-CSM2-MR and BCC-CSM1.1m are shown in the rightmost panel of Fig. 2. Clear improvements can be seen in Siberia, Northeast China, and the Yangtze–Huaihe River basin, which are basically the same as the areas of large BCC model–GLDAS differences mentioned above.
Figure 2. Differences in climate mean surface soil moisture (units: kg m−2) between GLDAS and BCC simulations on annual and seasonal time scales: (a, d, g, j, m) differences and RMSEs between GLDAS and BCC-CSM2-MR; (b, e, h, k, n) differences and RMSEs between GLDAS and BCC-CSM1.1m; (c, f, i, l, o) differences between BCC-CSM2-MR and BCC-CSM1.1m.
Figure 3 compares the standard deviations of the climate mean surface SM over Eurasia in terms of their annual mean and in the different seasons between the BCC model simulations and the GLDAS data. GLDAS shows that the larger standard deviations tend to be found at high latitudes and in part of central Asia, and clearly during winter and spring (Figs. 3d and g). This illustrates that the surface SM over these regions varies greatly and is spread out over a wider range in winter and spring. Conversely, the variations of surface SM in the low–middle latitudes are quite small, as illustrated by the relatively lower standard deviations over these areas. Gu et al. (2019) indicated that significantly lower SM is generally found in Russia and northeastern Asia, which are similar to the areas with large variations in Figure 3. It appears that both BCC-CSM2-MR and BCC-CSM1.1m are able to capture the spatial patterns, except that the amplitudes of variation are relatively weaker than in GLDAS. For instance, the standard deviation over western Siberia during winter is supposed to exceed 6 kg m−2 (Fig. 3d), but the simulated values in the BCC models range between 4 kg m−2 and 5 kg m−2. Despite the underestimations to a certain extent, BCC-CSM2-MR has made progress in terms of the standard deviation distributions, with higher pattern correlation coefficients than BCC-CSM1.1m in the annual mean and most seasons. The average pattern correlation coefficient increases by 4%, from 0.81 to 0.84. Therefore, BCC-CSM2-MR is better at describing the variations of surface SM.
Figure 3. Standard deviations of climate mean surface soil moisture on annual and seasonal time scale: (a, d, g, j, m) GLDAS as the reference data; (b, e, h, k, n) BCC-CSM2-MR and the pattern correlation coefficients of standard deviations with GLDAS; (c, f, i, l, o) BCC-CSM1.1m and the pattern correlation coefficients of standard deviations with GLDAS.
We also calculate the differences of the aforementioned standard deviations over Eurasia. Although the difference patterns between BCC-CSM2-MR and GLDAS are similar to those between BCC-CSM1.1m and GLDAS in terms of their annual mean (Figs. 4a and b), the discrepancies of BCC-CSM2-MR are less significant than those of BCC-CSM1.1m in the four seasons, especially in western Siberia and central Asia during spring, summer and winter, because the areas and absolute negative values decrease. That is to say, there is a certain underestimation in BCC-CSM1.1m when describing the standard deviations of SM. In addition, BCC-CSM2-MR is more skillful than BCC-CSM1.1m in terms of RMSE. The RMSE values of BCC-CSM2-MR are distinctly lower than those of BCC-CSM1.1m in terms of their annual mean and all seasons. The average RMSE reduces by 15.6%, from 1.56 to 1.35. This demonstrates that BCC-CSM2-MR is more capable of describing the spatial distributions of surface SM standard deviations, with fewer biases. From the rightmost panel of Fig. 4, we can see that the reason for the better performances in BCC-CSM2-MR is that the standard deviations described in BCC-CSM2-MR are systematically greater than those in BCC-CSM1.1m. This overcomes the defect of the underestimation in the previous generation, hence allowing BCC-CSM2-MR to present the degree of surface SM variation more accurately, especially in the middle and high latitudes.
Figure 4. Differences in standard deviations of mean surface soil moisture between GLDAS and model simulations on annual and seasonal times scales: (a, d, g, j, m) differences and RMSEs between GLDAS and BCC-CSM2-MR; (b, e, h, k, n) differences and RMSEs between GLDAS and BCC-CSM1.1m; (c, f, i, l, o) differences between BCC-CSM2-MR and BCC-CSM1.1m.
To make the results of the comparison more robust, we also employ a station-based observational SM dataset in China as the reference data. Due to the relatively greater number of missing values in the cold seasons, Fig. 5 only shows the difference patterns of the climate mean SM and the distributions of the standard deviations in summer; the results of the annual mean and other seasons are listed in Tables 2 and 3. From Figs. 5a and b, we can see that the SM over the Yangtze–Huaihe River basin and Northeast China simulated by BCC-CSM1.1m is obviously wetter than observed, but this situation improves in BCC-CSM2-MR. Figure 5c shows that through revising the overestimations over the regions mentioned above in the previous generation, the biases of the climate mean SM relative to observations are less severe in BCC-CSM2-MR. The reduction in RMSE (from 10.92 to 9.62) also demonstrates that the ability of the model simulation has been improved. It is worth noting that the patterns shown in Figs. 5a–c are consistent with the patterns in Fig. 2, where the reference data are from GLDAS. Figures 5d and e show that both BCC-CSM1.1m and BCC-CSM2-MR can capture the pattern of variation in SM over North China as being larger than in other areas, but the amplitudes are smaller than those of the observations to different degrees. However, BCC-CSM2-MR (standard deviation: 2.04) still performs better than BCC-CSM2-MR (standard deviation: 1.63), based on the closer average standard deviation to that of the observations (standard deviation: 3.22). The simulations of BCC-CSM2-MR in terms of the annual mean and the other seasons are also improved (Tables 2 and 3).
Annual DJF MAM JJA SON BCC-CSM2-MR 11.04 13.04 9.21 9.62 10.71 BCC-CSM1.1m 13.21 14.63 11.16 10.92 12.53 Table 2. RMSEs between observations and the BCC models in terms of climate-mean SM.
Annual DJF MAM JJA SON Observations 1.99 3.40 3.17 3.22 3.44 BCC-CSM2-MR 0.97 1.77 1.59 2.04 1.84 BCC-CSM1.1m 0.77 1.39 1.13 1.63 1.44 Table 3. Mean standard deviations of the observed and BCC-modeled SM.
-
The spatial features of surface SM anomalies over Eurasia from 1979 to 2012 are analyzed by using the EOF method. Figure 6 displays the first two principle components of the EOF analysis of GLDAS and the BCC model simulations. The EOF1 pattern in GLDAS, which explains 34.1% of the variance (Fig. 6a), shows that there are significant negative signals in the western Siberia, central Asia and the Kamchatka region, while other areas in Eurasia are covered by weak positive signals. This pattern suggests that the surface soil tends to getting drier in western Siberia, central Asia and Kamchatka region, but slightly wetter in other areas. In the EOF2 pattern of GLDAS, with an explained variance of 8.9% (Fig. 6b), there is a large-scale drying trend at high latitudes and mild wetting trend in other areas. The results are consistent with previous studies (Dong et al., 2007; Cheng et al., 2015; Gu et al., 2019). In BCC-CSM2-MR, the explained variance of the first two principle components reaches 20.0% in total (Figs. 6d and e). The EOF1 pattern of BCC-CSM2-MR that can capture negative values in western Siberia and central Asia has relatively high similarity with the EOF1 pattern of GLDAS. However, there are still some disagreements between BCC-CSM2-MR and GLDAS in the first principle component over the Arctic coastal region. This area is covered by positive values in BCC-CSM2-MR, which means the surface SM there tends to be wet. This trend is opposite to that of GLDAS. Therefore, BCC-CSM2-MR is able to capture the EOF1 spatial pattern of GLDAS, except at high latitudes in Arctic coastal regions. The EOF2 pattern of BCC-CSM2-MR, which shows that large-scale negative values cover the high latitudes, is similar to that of GLDAS, except that the gradient is not as obvious as shown in GLDAS. As for BCC-CSM1.1m, EOF1 (Fig. 6g) is unable to capture the spatial pattern of drying at high latitudes, as shown in the EOF1 of GLDAS and BCC-CSM2-MR. Neither western Siberia nor the Kamchatka region has the signs of a drying trend of surface SM. In the EOF2 pattern of BCC-CSM1.1m (Fig. 6h), the high latitudes are covered by negative values, but their area and gradient are relatively smaller. In conclusion, compared to the previous-generation model, BCC-CSM2-MR is more skillful in describing the first two principle components of the EOF analysis of the surface SM anomaly over Eurasia from 1979 to 2012. The periodogram estimates of the spectra of the first principal component time series (PC1) of GLDAS and the BCC model simulations are shown in the rightmost column of Fig. 6. The spectra above the red line are approved by the Markov “red noise” test. As shown in the panel, the corresponding period of GLDAS PC1 is 16.6 years. For the PC1s of BCC-CSM2-MR and BCC-CSM1.1m, the periods are 13.3 years and 8.3 years, respectively. Obviously, the period of PC1 in BCC-CSM2-MR is much closer to that in GLDAS. Therefore, BCC-CSM2-MR is more capable of capturing the periodicity characteristics of the SM variation.
Figure 6. The first two EOFs of annual mean soil moisture and periodogram estimates of the spectra of the PC1s: (a, d, g) EOF1 for GLDAS, BCC-CSM2-MR and BCC-CSM1.1m, respectively; (b, e, h) EOF2 for GLDAS, BCC-CSM2-MR and BCC-CSM1.1m, respectively; (c, f, i) periodogram estimates of the spectra of the PC1s of GLDAS, BCC-CSM2-MR and BCC-CSM1.1m (red lines: Markov “red noise” spectrum; green lines: upper confidence bound for Markov; blue lines: lower confidence bound for Markov).
-
We also compare the performance of BCC-CSM2-MR over Eurasia and China on annual and seasonal time scales with the other models participating in CMIP6, by using centered pattern correlation (computing anomalies from a central mean) and “amplitude of variations” (relative standard deviation). To make the comparison more comprehensive, BCC-CSM1.1m, the previous-generation model of BCC-CSM2-MR that participated in CMIP5, is also taken into account. Taylor diagrams (Taylor, 2001) are used to provide a visual representation of the aforementioned metrics.
Centered pattern correlations between the surface SM simulated by the models and GLDAS are indicated by the azimuthal position of each dot in the Taylor diagrams. For the simulated annual mean surface SM (Fig. 7a), the dots are relatively scattered. Correlations generally fall between 0.6 and 0.9 and tend to be clustered around 0.7. Compared with other CMIP6 models, BCC-CSM2-MR (r = 0.77), second only to EC-Earth3-Veg (r = 0.83) and MRI-ESM2-0 (r = 0.78), performs at a high level. BCC-CSM2-MR also shows marked progress in correlation compared with BCC-CSM1.1m (r = 0.62), with a rate of increase of 24.2%. In different seasons (Figs. 7b–e), although the correlations between the CMIP6 models and GLDAS are more variable, BCC-CSM2-MR is still well ahead of most models. Especially in winter (Fig. 7b), BCC-CSM2-MR has the highest correlation coefficient with GLDAS (r = 0.77). Also, the coefficients of BCC-CSM2-MR are obviously higher than those of BCC-CSM1.1m in all seasons. The radial distance from the origin represents the standard deviation of the model simulation relative to the standard deviation of GLDAS (σ sim / σ obs). The closer to 1 the ratio is, the fewer biases in simulating SM variations the models have. In terms of the annual mean (Fig. 7a), most models overestimate the standard deviation of surface SM considerably, with ratios above 1.25. BCC-CSM2-MR (σ sim / σ obs = 0.83) is one of the models able to give a comparatively accurate representation of the standard deviation. Meanwhile, compared with BCC-CSM1.1m (σ sim / σ obs = 1.24), BCC-CSM2-MR also does a better job. During different seasons (Figs. 7b–e), the ratios of BCC-CSM2-MR are between 0.77 and 0.88 (the average ratio is 0.81), which places BCC-CSM2-MR at a better level among the CMIP6 models (the average ratio is 1.48) in representing similar standard deviations to GLDAS. However, compared to the previous-generation model, with slight overestimation (average ratio of 1.15), the improvements of BCC-CSM2-MR are not that obvious.
Figure 7. Taylor diagrams for model-simulated surface soil moisture based on GLDAS over Eurasia: (a) annual mean; (b) winter; (c) spring; (d) summer; (e) autumn.
For the area of China, the dots in Fig. 8 are more clustered than in Fig. 7, which means that the models perform relatively consistently over China. For the annual mean surface SM (Fig. 8a), the correlation coefficients in the model simulations fall between 0.4 and 0.8. BCC-CSM2-MR (r = 0.61) is at the mid-upper level among the CMIP6 models. Compared with BCC-CSM1.1m (r = 0.41), BCC-CSM2-MR has made an evident improvement. In the four seasons (Figs. 8b–e), the correlation coefficients of BCC-CSM2-MR always fall within the range of 0.5–0.8. The average is 0.64, which is higher than that of the CMIP6 models (r = 0.61), showing BCC-CSM2-MR to be more skillful in accurately and consistently representing the distribution of surface SM compared to most of the CMIP6 models. Meanwhile, BCC-CSM2-MR takes the lead in comparison with BCC-CSM1.1m (average correlation coefficient is 0.40). As for the ratios between the standard deviations of the model simulations and those of the reference data, the advantages of BCC-CSM2-MR are highlighted. The ratios of BCC-CSM2-MR are generally around 0.7–0.8 which are the closest to 1 among all the CMIP6 models. This indicates that BCC-CSM2-MR can capture the variability of the surface SM properly and maintain a minimum bias relative to the reference data. Compared with BCC-CSM1.1m, BCC-CSM2-MR carries forward the advantage of a similar standard deviation to that of GLDAS, and a slight improvement to the overestimation of the previous-generation model.
Figure 8. Taylor diagrams for model-simulated surface soil moisture based on GLDAS over China: (a) annual mean; (b) winter; (c) spring; (d) summer; (e) autumn.
In general, compared with other CMIP6 models and BCC-CSM1.1m, BCC-CSM2-MR is more competent in describing the distributions and variations of annual and seasonal SM, as shown by the relatively higher centered pattern correlation and standard deviation that is closer to that of the reference data, either in Eurasia or China.
-
The above analysis shows that BCC-CSM2-MR has made certain progress in the simulation of surface SM compared to BCC-CSM1.1m. This might benefit from improvements to parameterization schemes in the component models, especially the land model implemented in BCC-CSM2-MR (Wu et al., 2019), such as the inclusion of a variable temperature threshold to determine soil water freeze–thaw, rather than a fixed temperature of 0 °C, a better calculation of snow cover fraction, and so on (Li et al., 2019). In the actual process of land–atmosphere coupling, there are many elements that have effects on the variations of SM, such as precipitation, temperature, wind, etc. Among these elements, precipitation is well known as the most vital and has thus been widely studied. The wetting of soil by precipitation, identified as the first part of the land–atmosphere feedback of water, is straightforward and intuitive and indisputably occurs in nature (Koster et al., 2003, 2004; Tawfik and Steiner, 2011). Numerous studies have shown that precipitation is the most direct and vital among the factors affecting SM. Zhang et al. (2008) used analysis data of precipitation and SM from GLDAS and pointed out that the strong land–atmosphere coupling lies mainly in semi-humid forest to grassland transition zones or in arid to semi-arid transition zones, including central Eurasia, northern China, etc. Research has also shown that the interaction between precipitation and SM exists in atmospheric general circulation models (Oglesby and Erickson III, 1989; Dirmeyer, 2000).
The SM procedure adopted in BCC-CSM is almost the same as that in the NCAR Community Land Model (Oleson et al., 2004). SM is governed by infiltration, surface and sub-surface runoff, gradient diffusion, gravity, and root extraction through canopy transpiration. For one-dimensional vertical water flow in soils, the conservation of mass is stated as
where θ is the volumetric soil water content, t is time and z is height above some datum in the soil column, q is the soil water flux, and e is the evapotranspiration loss. In the coupling between the land model and atmospheric model, liquid and solid precipitation from the atmospheric model will have an important effect on q and subsequently influence the simulation of SM. Therefore, the improvements in SM simulation in BCC-CSM2-MR may be attributable to a better simulation of precipitation. Influences of precipitation on SM in the models are thus discussed as follows.
Differences in the climate mean of the surface SM (c.f. Fig. 2) have shown that there are three regions where the difference values are significantly improved: Siberia (55°–64°N, 60°–87°E), Northeast China (40°–50°N, 120°–135°E), and the Yangtze–Huaihe River basin (28°–33°N, 110°–121°E). The difference values in these three regions decrease by 32.7%, 30.0% and 20.6%, respectively. Hence, we select these three subregions (see the black sectors in Figs. 2a–c) and the whole of Eurasia as target areas to investigate the possible reasons behind the improvement in SM simulation, which has been linked with precipitation.
Figure 9 shows the correlation coefficients between the time series of the annual mean surface SM and precipitation in the reference data (GLDAS for SM, CMAP for precipitation) and the BCC model simulations. In GLDAS/CMAP, the correlation coefficients range between 0.35 and 0.60, which are statistically significant at the 95% confidence level according to the Student’s t-test. As for the BCC model simulations, the correlation coefficients are generally significant in most areas except the whole of Eurasia. For Eurasia, which of course covers a wide range of longitudes and latitudes, the factors that influence SM in are more complex and diverse. Therefore, this complicated land–atmosphere interaction makes the proportion of the influence of precipitation on SM smaller and the correlation coefficients between them to be reduced further. Coupled climate models may not properly capture the complicated relationships between SM and factors of influence, and perhaps may even weaken the influence of precipitation excessively. Hence, the correlation coefficients in the BCC model simulations are quite low and even non-significant. As for the three subregions, the correlation coefficients of BCC-CSM2-MR in Siberia, Northeast China and the Yangtze–Huaihe River basin are 0.49, 0.67 and 0.87, respectively, while those of BCC-CSM1.1m are 0.60, 0.78 and 0.89, respectively. The former is apparently lower than the latter and much closer to the coefficients in GLDAS\CMAP (r = 0.52, 0.59 and 0.49, respectively). That is to say, the proportion of influence of precipitation on surface SM in BCC-CSM2-MR is not as much as in BCC-CSM1.1m. As seen on the annual time scale, the coefficients of BCC-CSM2-MR are more consistent with those of the reference data on the seasonal scale (not shown here). In other words, compared with the previous-generation model, BCC-CSM2-MR is able to represent, relatively realistically, the relationship between precipitation and surface SM in these three subregions.
Figure 9. Correlation coefficients between the annual mean time series of soil moisture and precipitation of BCC-CSM2-MR, BCC-CSM1.1m and reference data (GLDAS/CMAP) over Eurasia (EA), Siberia (SIB), Northeast China (NEC) and the Yangtze–Huaihe River basin (YH).
The anomalies of surface SM in GLDAS and the BCC models, in conjunction with corresponding anomalies of precipitation in CMAP and the BCC models, over Eurasia and the three subregions, are shown in Fig. 10. In Eurasia, the SM responses in GLDAS are positively correlated with the precipitation variations in CMAP (Fig. 10a). By analyzing the slope of the regression line, we can conclude that BCC-CSM2-MR is able to represent the correlation with a positive linear regression coefficient, but BCC-CSM1.1m shows an opposite correlation with a negative linear regression coefficient. In the three subregions (Figs. 10b–d), the dots are not as spread as they are in Eurasia. This means that precipitation anomalies play a more obvious and important role in increasing SM. In general, the regression lines in BCC-CSM2-MR are closer to those in the reference data, as indicated by the similar spacing and trends of the dots. Taking Northeast China as an example, the linear regression coefficient of GLDAS\CMAP is 0.58, while the coefficients of BCC-CSM2-MR and BCC-CSM1.1m are 0.67 and 0.77, respectively. Obviously, the former is closer than the latter to the reference data. Therefore, compared to the previous version, BCC-CSM2-MR is more skillful in describing the relationship between precipitation and surface SM variations.
Figure 10. Scatterplots showing the anomalies of soil moisture in the reference data and BCC models, in conjunction with corresponding changes in precipitation, over (a) Eurasia (EA), (b) Siberia (SIB), (c) Northeast China (NEC) and (e) the Yangtze–Huaihe River basin (YH).
We also evaluate the RMSE values of the mean precipitation between CMAP and the BCC model simulations on the annual time scale (Table 4). The RMSEs of BCC-CSM2-MR are lower than those of BCC-CSM1.1m to some extent. For Eurasia, the RMSE of the annual mean decreases by 9.6%, from 1.14 (BCC-CSM1.1m) to 1.03 (BCC-CSM2-MR). In the three subregions, the RMSEs of BCC-CSM2-MR tend to be obviously lower. The Yangtze–Huaihe River basin is the region with the most obvious improvement, where the RMSE is reduced by 40.4%. The progress is also significant on the seasonal time scale (not shown here). Hence, BCC-CSM2-MR is generally more skillful than BCC-CSM1.1m in realistically representing the variation of precipitation, as shown by the smaller deviation between the simulation and CMAP.
Eurasia (10°–80°N, 30°–160°E) Siberia (55°–64°N, 60°–87°E) Northeast China (40°–50°N, 120°–135°E) Yangtze–Huaihe River basin (25°–33°N, 110°–121°E) BCC-CSM2-MR 1.03 0.24 0.36 0.81 BCC-CSM1-1-m 1.14 0.28 0.47 1.36 Table 4. RMSEs between the BCC models and CMAP data of annual mean precipitation over Eurasia and four subregions.
From the above analyses, BCC-CSM2-MR is better able to properly describe the correlation between the surface SM and precipitation, the response of the surface SM variation to precipitation anomalies, and the variation of precipitation. These qualities may contribute to the better simulation of surface SM in BCC-CSM2-MR.