Accurate subseasonal-to-seasonal (S2S) atmospheric forecasts and hydrological forecasts have considerable socioeconomic value. This study conducts a multimodel comparison of the Tibetan Plateau snow cover (TPSC) prediction skill using three models (ECMWF, NCEP and CMA) selected from the S2S project database to understand their performance in capturing TPSC variability during wintertime. S2S models can skillfully forecast TPSC within a lead time of 2 weeks but show limited skill beyond 3 weeks. Compared with the observational snow cover analysis, all three models tend to overestimate the area of TPSC. Another remarkable issue regarding the TPSC forecast is the increasing TPSC with forecast lead time, which further increases the systematic positive biases of TPSC in the S2S models at longer forecast lead times. All three S2S models consistently exaggerate the precipitation over the Tibetan Plateau. The exaggeration of precipitation is prominent and always exists throughout the model integration. Systematic bias of TPSC therefore occurs and accumulates with the model integration time. Such systematic biases of TPSC influence the forecasted surface air temperature in the S2S models. The surface air temperature over the Tibetan Plateau becomes colder with increasing forecast lead time in the S2S models. Numerical experiments further confirm the causality.
Anomalous weather- and climate-related natural disasters are among the most common disasters and are associated with severe socioeconomic consequences. Reliable forecasts of such weather and climate anomalies with sufficient lead time have significant benefits for decision-makers (White et al., 2017). Traditionally, weather forecasts cover a time range of up to 2 weeks, while climate forecasts begin at the seasonal timescale and extend outward. Demands are growing rapidly in operational forecasts in the subseasonal-to-seasonal (S2S) range (from 2 weeks to a season). The primary basis for longer lead forecasts beyond 2 weeks is the interaction of the atmosphere with other, more slowly varying earth system components, such as the ocean or land, that evolve over timescales of weeks and months rather than days as in the atmosphere (Mariotti et al., 2018). Land–atmosphere coupling is one of the key physical processes for S2S prediction but is not well simulated and may reduce S2S prediction skill (Robertson et al., 2014; Dirmeyer et al., 2019).
Snow cover is a crucial component in both the climate system and the cryosphere. The radiative and thermal properties of snow cover significantly influence the ground thermal regime (Zhang, 2005). As the lower boundary condition of the atmosphere, snow cover forces the regional and global atmosphere and can serve as an indicator of the atmosphere (Barnett et al., 1989; Bamzai and Shukla, 1999; Wu and Kirtman, 2007; Henderson et al., 2018). Snow cover can vary rapidly within a season over discontinuous or sporadic permafrost zones (Wang et al., 2015; Suriano and Leathers, 2018; Song et al., 2019; Li et al., 2020a) and rapidly influence the atmosphere (Clark and Serreze, 2000; Zhang et al., 2019). Snow cover may provide a potential source of S2S predictability via its variability and atmospheric effects at the subseasonal timescale (F. Li et al., 2019; Diro and Lin, 2020).
The Tibetan Plateau is the highest plateau in the world and is known as the “third pole”. Due to its high elevation and cold climate, the Tibetan Plateau has much more snow cover than the other regions at the same latitude. Tibetan Plateau snow cover (TPSC) is a key component of the climate system. TPSC influences land surface thermal conditions (Chen et al., 2017; Li et al., 2018) and thus influences atmospheric circulations and monsoons over Asia and beyond (Wu and Qian, 2003; Lin and Wu, 2011; Xiao and Duan, 2016; Wang et al., 2017; You et al., 2020). TPSC shows variations at multiple timescales, including the subseasonal scale (Li et al., 2016; Song and Wu, 2019; Li et al., 2020a). The subseasonal variations in TPSC influence the atmosphere over East Asia (Li et al., 2018; Li et al., 2020b). A better TPSC simulation and forecast may favor a better forecast for weather and climate at the S2S timescale.
Snow cover also affects the hydrologic cycle. The accumulation of precipitation in the form of snow and its release through snowmelt runoff is an important component of the hydrologic cycle (Jeelani et al., 2012; Fayad et al., 2017). TPSC plays an important role in hydrological systems, providing a reservoir of water and acting as a buffer that controls river discharge. Rivers including the Yangtze River, Yellow River, Yarlung Zangbo River and Mekong River have headwaters over the Tibetan Plateau. Studies on the variability in TPSC are critical for water management in downstream regions (Immerzeel et al., 2009; Zhang et al., 2012, 2013). Skillful predictions of TPSC with sufficient lead time are thus of great societal importance for hydrologic prediction.
Since the implementation of the S2S prediction project database (Vitart et al., 2016), many studies have evaluated the skill of S2S models for atmospheric elements and variables, such as the Madden–Julian Oscillation (Vitart, 2017), surface air temperature (Yang et al., 2018; Wulff and Domeisen, 2019) and precipitation (de Andrade et al., 2019). Some works also focus on the skill of S2S models for hydrological elements (W. Li et al., 2019; Schmitt Quedi and Mainardi Fan, 2020). However, we still know little about the skill of S2S models for TPSC. Understanding the forecasting skills of the S2S model on the TPSC is the first step to applying the S2S model to hydrological forecasts over the Tibetan Plateau. Moreover, considering the influence of TPSC on the atmosphere, clarifying the issue of the S2S model for TPSC helps improve the ability of the S2S model for atmospheric forecasting.
This study conducts a multimodel comparison of the TPSC prediction skill using selected models from the S2S project database to learn about their performance in capturing TPSC variability. Our main goal is to use the state-of-the-art S2S prediction systems of these operational centers to demonstrate why models exhibit systematic biases of TPSC and whether such systematic biases influence the regional air temperature forecasted in S2S models. The remainder of this paper is organized as follows. Details on the dataset and method used in this study are described in Sect. 2. The systematic bias of TPSC in S2S models and its effect on local temperature during wintertime are presented in Sects. 3 and 4, respectively. The conclusions and a discussion are presented in Sect. 5.
The reforecasts considered for this study are taken from three operational forecast systems that are part of the S2S project database: the European Centre for Medium Range Weather Forecasts (ECMWF), the US National Centers for Environmental Prediction (NCEP) and the China Meteorological Administration (CMA). These models share a common reforecast period of 1999–2010 with a reforecast initialized frequency that is equal to or greater than once a week. This study only used reforecasts produced by the control forecast (using a single unperturbed initial condition). Details of the S2S database can be found in Vitart et al. (2016). Daily reforecast data were averaged for each 7 d period starting every 1 January to create a total of 52 weeks per year (31 December was excluded). The reforecasts that initialized on the first day of these weeks were selected. Forecast lead times were defined here as 1 week (1–7 d), 2 weeks (8–14 d), 3 weeks (15–21 d), 4 weeks (22–28 d) and 5 weeks (29–35 d).
For the ECMWF model, the reforecasts initialization is based on ERA-Interim
and ERA-Interim/Land datasets. The daily Interactive Multisensor Snow and
Ice Mapping System (IMS) snow cover product has been used to constrain the
ERA-Interim snow analysis (Dee et al., 2011). The NCEP model also initialized realistic snow in the forecasts. The snow initialization comes from the
Climate Forecast System Reanalysis snow analysis using IMS and the Air Force
Weather Agency snow depth analysis. Snow in the CMA model was not directly
initialized in the forecasts. The initial conditions of the snow in the CMA
model are from a balanced state produced by long-term air–sea initialization
integration. See the details on snow initialization in the S2S models at
The land surface models used for ECMWF, NCEP and CMA are the Hydrology Tiled ECMWF Scheme for Surface Exchanges over Land (HTESSEL; Balsamo et al., 2009), Noah (Ek et al., 2003) and BCC_AVIM2 (Wu et al., 2014), respectively. All these land surface models contain snow schemes. According to the snow scheme in each land surface model, we obtain the snow cover fraction, which is a diagnostic variable in this study.
The snow cover fraction (
The
The
The surface air temperature (SAT) in these S2S models is also used. All
variables are at a 1
The Tibetan Plateau area of focus in this study is the region within
26–41
The location and topography of the Tibetan Plateau. Shading shows
topography (unit: m). The black rectangle shows the region within
26–41
The reforecasts in the S2S models are verified against observational daily
snow cover and SAT in the reanalysis. Observational daily snow cover data
are obtained at a 24 km resolution from the Interactive Multisensor Snow and
Ice Mapping System (IMS) snow cover analysis (Helfrich et al., 2007)
provided by the National Oceanic and Atmospheric Administration. The IMS
examines satellite images and other sources of data on snow cover and
generates maps of snow cover distribution. The IMS analysis over the Tibetan
Plateau corresponds well with ground-based measurements and can capture the
general subseasonal variability in TPSC (Yang et al., 2015; Li et al.,
2018). The original 24 km resolution IMS analysis is interpolated into the
1
Two precipitation datasets, the Global Precipitation Analysis Products of the Global Precipitation Climatology Centre (GPCC; Schneider et al., 2011) and the Tropical Rainfall Measuring Mission (TRMM; Huffman et al., 2007), are used to evaluate the wintertime mean precipitation. The GPCC precipitation dataset is from built rain gauges that were GTS based. The TRMM precipitation dataset is based on satellite observations. The precipitation used in this study spans 11 winters (from 1999/2000 to 2009/2010).
To quantify the forecast ability of S2S models, three common statistical measures, i.e., the temporal correlation coefficient (TCC), the root-mean-square error (RMSE) and the mean bias, are calculated in this study. A composite analysis is performed to investigate the different performances on predicting the snow cover for increasing cases and decreasing cases (details are described in Sect. 3.2).
To reveal the causality of the systematic bias of the TPSC-induced regional
SAT bias, numerical experiments are performed. Numerical experiments are
performed using the Advanced Weather Research and Forecasting Model
(WRF-ARW, version 4.1.3), which was developed by the National Center for
Atmospheric Research (NCAR). WRF-ARW has been applied to climate research,
including studies of land–atmosphere interactions. The land surface
parameterization scheme used in this study is the Noah land surface model
(Ek et al., 2003). Important physics options include the WRF single-moment
6-class microphysics scheme (Hong and Lim, 2006), the NCAR Community
Atmosphere Model (CAM 3.0) spectral-band shortwave and longwave radiation
schemes (Collins et al., 2006), the Yonsei University planetary boundary
layer scheme (Hong et al., 2006) and the Kain–Fritsch convective
parameterization scheme (Kain, 2004). The WRF is driven by atmospheric and
surface forcing data extracted from the National Centers for Environmental
Prediction (NCEP) FNL (Final) Operational Model Global Tropospheric Analyses. The
simulation domain is in a cylindrical equidistant projection with a
horizontal resolution of 1
Two ensemble experiments are performed: control (CTL) runs and sensitive experimental (EXP) runs. All these runs have the same initial times as the forecasts in the S2S models that we used in this study for each winter. But the experiments were run for 20 winters (from 2000/2001 to 2019/2020), and both runs contain 340 cases. Each member ran continuously for 22 d. The first day in each run is for spin-up, and the results are discarded. The CTL runs are integrated freely without any modification. Because both the NCEP S2S model and our numerical experiment use Noah as the land surface model, the TPSCs in CTL runs are expected to show unreal increases with integration time, which is similar to that in the NCEP S2S model (will be revealed in Sect. 3). The EXP run is designed to eliminate such bias in TPSC. The FNL analyses are from the Global Data Assimilation System (GDAS), which continuously collects observational data from the GTS and other sources for many analyses. GDAS incorporates daily snow data from IMS analyses and the Air Force Weather Agency Snow Depth Analysis Model. We replace the forecasted TPSC in the WRF model with TPSC in the FNL analyses every 6 h. Because FNL analyses assimilate the observed TPSC, the TPSC in the EXP run is expected to show a small bias that increases with integration time. We averaged all 340 cases in CTL runs and EXP runs respectively. Ensemble mean results between the CTL and EXP runs are compared with each other.
Before we present the systematic bias of TPSC in the S2S models, the overall forecast skill of TPSC is evaluated. Here, we focus on the variation in snow-covered area over the entire Tibetan Plateau, which can be measured by a Tibetan Plateau snow cover index (TPSCI). The TPSC index represents the percentage of grid points covered by snow in the analysis or models over the entire Tibetan Plateau. The unit of the TPSC index is percent (%). The prediction skill of the TPSC index has been investigated through the TCC and RMSE between the TPSC index in the predictions and that in the observations during wintertime (Fig. 2). A skillful prediction is generally defined as a TCC greater than 0.5. All three models show good prediction skills at lead times of 1–2 weeks with a TCC greater than 0.5 (Fig. 2a). At lead times of 1–2 weeks, the TCC for the ECMWF model is largest among the three models. The NCEP model has the lowest TCC among the three models at a lead time of 1 week. However, the TCC for NCEP falls the most slowly at lead times of 2 weeks or more. The NCEP model has a larger TCC than the CMA model at lead times of 2 weeks or more. The TCC values decrease with the increase in the forecast lead time and decline below 0.5 at and after lead times of 3 weeks for all three models. RMSEs increase with the forecast lead time (Fig. 2b). The RMSE for ECMWF is the smallest among the three models. Additionally, CMA has the largest RMSEs. These results indicate that the S2S models can skillfully forecast TPSC variations within a lead time of 2 weeks during wintertime but show limited skill at a lead time of 3 weeks or more.
Prediction skill of the Tibetan Plateau snow cover (TPSC) index in
the S2S models during wintertime.
The above results also indicate that the ECMWF model is shown to have a better TPSC forecasting skill than the other two models. Even so, the ECMWF model shows nonnegligible RMSEs with a TPSC index of more than 15 % (Fig. 2b). The other two models, especially the CMA model, show even more significant RMSEs up to more than 25 %. These large errors in the forecasting of the TPSC are induced by systematic bias of the TPSC, as shown by the following. The multiyear wintertime mean biases of the TPSC index in forecasts against that in the IMS snow cover analysis for all three models show positive values, which indicates that all of the models tend to overestimate the TPSC during winter (Fig. 3a). The TPSC index in the ECMWF is higher than the observed TPSC index by approximately 20 %–30 %. NCEP has a larger TPSC index than that in the observation by approximately 5 %–20 %. The CMA shows largest biases of approximately 25 %–40 %.
Another remarkable issue regarding the forecast of TPSC is the increasing TPSC with forecast lead time, which further increases the overestimation of TPSC in models at longer forecast lead times. These increasing biases can be detected from the multiyear winter mean biases (Fig. 3a). To highlight such increasing biases, we further present differences in the multiyear winter biases for the TPSC index between forecasts for leads of 2–5 weeks and forecasts for leads of 1 week in three modes (Fig. 3b). Such differences are obtained by subtracting the multiyear winter mean of the TPSC index at a lead time of 1 week from that at forecast lead times of 2–5 weeks. The differences in the three models show common features: the differences in all three models are all positive and increase with increasing forecast lead time. The positive biases of TPSC with the longest forecast lead time (5 weeks) are largest among all forecasts. The increases in the differences in the ECMWF model are the smallest, while the CMA model has the largest increases in the differences. Taking the differences between the forecasts with a lead of 4 weeks and the forecasts with a lead of 1 week as an example, the spatial patterns of these increases in the biases in the three models show some similarities (Fig. 4). Although the spatial patterns of the differences in the three models show some small discrepancies, the differences are mainly positive in the three models, especially over parts of central and eastern Tibetan Plateau. These indicate that the increasing TPSC with the forecasting lead time occurs at a regional scale.
The intraseasonal variability in TPSC leads to obvious rapid variations in TPSC with a period shorter than a season, making TPSC exhibit a distinct lack of persistence within one season (Li et al., 2020a). Both accumulation and dissipation of snow cover occur within a season over the Tibetan Plateau. The increase in TPSC with forecast lead time in the models may be induced by overestimation of snow cover accumulation or underestimation of snow cover dissipation. To support this hypothesis, we analyzed the frequency of weekly TPSC accumulation and dissipation in the observation and forecast models in winter (Table 1). Here, the increasing (decreasing) weeks means that the TPSC index is greater (less) than that in the preceding week. The TPSC indexes in the S2S models are compared with the TPSC indexes in the preceding week, which are initialized at the same time, but with different forecast lead times.
The proportion of increasing (decreasing) weeks in the observations and forecast models with different lead times (in weeks).
The proportions of increasing and decreasing weeks in the observations are 50.3 % and 49.7 %, respectively, which is fairly even (Table 1). However, this kind of balance does not exist in the models. In the models, the proportion of increasing weeks is mostly more than 2 times as large as the proportion of decreasing weeks. The proportion of decreasing weeks is low compared with that in the observations. Specifically, decreasing weeks occupy only 23.0 %–31.0 % of the total forecasts by ECMWF. NCEP shows similar results, except for forecasting at a lead time of 5 weeks. This underestimation of the proportion of decreasing weeks is more severe in CMA. Moreover, the most severe underestimations of the proportion of decreasing weeks are the forecasts with a lead time of 2 or 3 weeks for all models.
The above results indicate that the models underestimate the frequency of TPSC dissipation, whereas they overestimate the frequency of TPSC accumulation, which leads to a systematic TPSC bias. To highlight increases in the overall TPSC biases, as well as changes in biases in successive weeks, a composite analysis is performed for all TPSC reforecasts during winter (Fig. 5a), increasing TPSC cases (Fig. 5b) and decreasing TPSC cases (Fig. 5c). All reforecasts initialized in winter are taken into account for the composite of all cases shown in Fig. 5a. The sample numbers of all cases are 187. Among all cases, we further select the increasing TPSC cases and decreasing TPSC cases. If the TPSC index continues to increase (decrease) for 3 weeks, this case is regarded as an increasing (decreasing) TPSC case. There are 46 increasing TPSC cases and 53 decreasing TPSC cases. We average the 46 (53) cases for different lead times. To focus on the increase in biases, values with a lead time of 1 week are removed for forecasting at all lead times.
Differences in the multiyear wintertime mean
Tibetan Plateau snow cover fraction (unit: %) between forecasts with a
lead of 4 weeks and forecasts with a lead of 1 week in
Composites of the Tibetan Plateau snow cover index (unit: %)
for
On a seasonal average, the growth of the TPSC index in winter is only
1.3 % over 2 weeks in the observation (black line in Fig. 5a). However,
the models tend to exaggerate the growth of the TPSC index (color lines in
Fig. 5a). The growth of the TPSC index over the 2 weeks in the models
ranged from 4.9 % (ECMWF) to 9.8 % (CMA). The TPSC index in the forecast
shows distinct differences between the increasing TPSC cases and decreasing
TPSC cases (Fig. 5b and c). The growth of the TPSC index in the increasing
TPSC cases is 14.1 % over 2 weeks in the observation (black line in Fig. 5b). The growth of the TPSC index over 2 weeks in NCEP and CMA is close to
that in the observation, while there is some underestimation of such growth
in the ECMWF (color lines in Fig. 5b). Although there are some differences
between the TPSC index in the models and that in the observation, all models
can forecast the increasing trend in the TPSC index. However, the situation
for the decreasing TPSC cases is quite different. The reduction of the TPSC
index in the decreasing TPSC cases is
Studies have shown that current state-of-the-art atmospheric general
circulation models (GCMs) tend to strongly overestimate the precipitation
over the Tibetan Plateau (e.g., Su et al., 2013; Chen and Frauenfeld, 2014;
Zhang and Li, 2016; Zhang et al., 2019). For example, Su et al. (2013)
evaluated 24 GCMs that were available in the fifth phase of the Coupled
Model Intercomparison Project (CMIP5) over the eastern Tibetan Plateau by
comparing the model outputs with ground observations, and they found that
all of the models consistently overestimated the observed precipitation for
all seasons. Zhang et al., (2019) found similar results, in that all climate
models they evaluated exaggerated the daily precipitation in the Tibetan
Plateau during winter compared with the observed values. Here, we also found
that the S2S models tended to overestimate the precipitation over the
Tibetan Plateau. We compared the precipitation in the S2S models with both
the gauge-based GPCC precipitation dataset and the satellite-based TRMM
precipitation dataset (Fig. 6). The regional averaging wintertime mean
precipitation values over the Tibetan Plateau in the GPCC and TRMM models are 0.27 and 0.32 mm d
The multiyear wintertime mean precipitation over the Tibetan
Plateau (unit: mm d
In this section, it was found that S2S models underestimate the frequency of TPSC dissipation and have some difficulties forecasting TPSC dissipation with an observed rate. Exaggerations of the precipitation were found in all three models, which directly lead to accumulated overestimation of TPSC. As a result, systematic bias of TPSC occurs and increases with the model integration time.
The local SAT over the Tibetan Plateau is highly correlated with simultaneous TPSC at a subseasonal timescale (Li et al., 2020a). Local snow–temperature relationships in S2S models were examined. We took a similar approach as in F. Li et al. (2019) and Diro and Lin (2020). The temporal correlation between the snow cover fraction and SAT with a lead of 1 week and 4 weeks for each grid point in the three models was computed to identify the extent and nature of the relationship (Fig. 7). Almost all of the regions exhibit a significant negative correlation in all of these three models. Additionally, such a relationship in all three models did not weaken with the forecasting lead time (compare Fig. 7a–c and Fig. 7d–f), even if the forecasting skill on the TPSC declined over time. The reason is that the relationship between the snow cover fraction and the SAT is embedded in the land surface model.
Spatial pattern of correlations between the snow cover
fraction and the surface air temperature with a lead of 1 week in
The skill of predicting the TPSC will further influence the skill of
predicting the SAT. As shown in Sect. 3, the TPSC in the S2S models during
the cold season increases with increasing forecast lead time. Such
systematic biases of TPSC may influence the forecasted SAT in the S2S
models. To test this hypothesis, we performed an analysis on SAT over the
Tibetan Plateau similar to our analysis on TPSC. The SAT over the Tibetan
Plateau is derived by averaging the SAT over the Tibetan Plateau region as
defined in Sect. 2.2. Differences in the multiyear winter mean SAT over
the Tibetan Plateau between forecasts with leads of 2–5 weeks and forecasts
with leads of 1 week in the three models, which were obtained by subtracting
the multiyear winter mean with a lead time of 1 week from that for forecast
lead times of 2–5 weeks, are examined (Fig. 8). The differences in the
three models show some common features. The differences in all three models
are all negative. By comparing values at different lead times, we also find
that such negative differences increase with increasing lead time, except for the value at a lead of 3 week in the CMA model. The negative differences of SAT
with the longest forecast lead time (5 weeks) are largest among all
forecasts. The differences in SAT between the forecast for a lead of 5 weeks and
the forecast for a lead of 1 week can be up to 1.9
Differences in the multiyear wintertime mean surface air
temperature over the Tibetan Plateau (unit:
Differences in the multiyear wintertime mean
surface air temperature over the Tibetan Plateau (unit:
The above results indicate that the SAT over the Tibetan Plateau becomes colder with increasing forecast lead time in the S2S models. Considering the results we obtained in Sect. 3, it can be concluded that the increasing TPSC is accompanied by decreasing SAT with forecast lead time.
Section 3.2 reveals that models show different performances on snow cover accumulation and dissipation. We also found that there are some difficulties for the models in forecasting the dissipation of TPSC. To learn whether such different performances influence the SAT forecast and to examine the sensitivity of SAT to TPSC in the S2S models, we investigated the changes in SAT in the S2S models over the Tibetan Plateau during winter (Fig. 10a), as well as the increasing TPSC cases (Fig. 10b) and decreasing TPSC cases (Fig. 10c). To provide a SAT reference in the models, a composite was performed on SAT in the ERA-Interim reanalysis. We performed the same composite method as that is used in Sect. 3.2 on TPSC but for SAT over the Tibetan Plateau.
Composites of surface air temperature over the Tibetan Plateau
(unit:
On a seasonal average, the change in SAT over the Tibetan Plateau in the
reanalysis during winter is less than 0.1
The change in SAT should be closely connected to the variations in TPSC. The
change in SAT in the increasing TPSC cases is
Here, we further find that such biases lead to biases in SAT. SAT increases
by 1.4
Through the results in Sect. 4.1 and 4.2, we find that the local SAT over the Tibetan Plateau becomes colder with increasing forecast lead time. We assumed that the cold SAT biases are induced by the overestimation of TPSC. However, the relationship between snow cover and the atmosphere is a two-way coupling connection (Henderson et al., 2018). The assumption should be tested by numerical experiments (see Sect. 2.2 for details about the numerical model and experimental design). Otherwise, one may suspect that the cold SAT induces an increasing TPSC other than the TPSC influence on SAT. Therefore, we used the predicted TPSC as a boundary condition in CTL runs (with overestimated TPSC), while observational TPSC in GDAS was used as a boundary condition in the EXP runs (without overestimated TPSC). The difference between the CTL and EXP runs is considered to represent the response or the sensitivity of the SAT to the overestimated TPSC.
We averaged snow cover and SAT over the Tibetan Plateau in all simulations
for the CTL and EXP runs to obtain a composite for all reforecasts of TPSC during
winter in the numerical experiment (Fig. 11a–b). As we discussed in Sect. 3.2, the growth of the TPSC index in winter is only 1.3 % for 2 weeks in
the observations, while the S2S models tend to exaggerate the growth of the
TPSC index (Fig. 5a). In the numerical experiment, CTL also exaggerates the
growth of the TPSC index (blue line in Fig. 11a). Because both the NCEP S2S
model and our numerical experiment use Noah as the land surface model, such
biases may be attributed to the land surface model. Compared with the CTL run, the EXP run
shows smaller cumulative biases (red line in Fig. 11a), which is because
TPSC in the EXP run is replaced by TPSC in the FNL analyses every 6 h. The SAT
becomes colder with increasing forecast lead time in CTL (blue line in Fig. 11b). However, such a decrease in SAT is much smaller in the EXP run (red line in
Fig. 11b). By checking the land surface energy fluxes over the Tibetan
Plateau between the CTL run and the EXP run (Fig. 11c), we found that the overestimated TPSC
strongly increases the upward-reflected shortwave radiation (negative value
indicates enhanced upward radiation) due to the snow-albedo effect. This
difference in the solar surface energy leads to a decrease in the absorbed
solar radiation. Thus, the net shortwave radiation is decreased (
Sensitivity of SAT and surface energy balance to TPSC
biases in the numerical experiments.
Accurate subseasonal-to-seasonal (S2S) atmospheric forecasts and hydrological forecasts have considerable socioeconomic value. This study evaluates the Tibetan Plateau snow cover (TPSC) prediction capabilities of three S2S forecast models (ECMWF, NCEP and CMA) during wintertime. These three S2S models can skillfully forecast TPSC variations within a lead time of 2 weeks during wintertime with temporal correlation coefficients greater than 0.5. ECMWF better captures TPSC variations compared with NCEP and CMA at a lead time of 1–2 weeks. All models show limited skill in forecasting TPSC at a lead time of 3 weeks or more. Compared with the IMS snow cover analysis, all three models tend to overestimate the area of TPSC. Another remarkable issue regarding the TPSC forecast is the increasing TPSC with forecast lead time, which makes the systematic positive biases of TPSC in models further increase at longer forecast lead times.
S2S models underestimate the frequency of TPSC dissipation, whereas they
overestimate the frequency of TPSC accumulation. The accumulation and
dissipation of wintertime TPSC occurs evenly in the observations. However,
this kind of balance does not exist in the S2S models. In the models, the
proportion of TPSC accumulation is mostly more than 2 times as large as the
dissipation proportion. The most severe underestimations of the dissipation
proportions are the forecasts at a lead time of 2 or 3 weeks for all models.
The models also have some difficulties forecasting the TPSC dissipation at
an observed rate. The growth of TPSC in the decreasing TPSC cases is
All of the three S2S models consistently exaggerate the precipitation over the Tibetan Plateau compared to the observations. The exaggeration of the precipitation is prominent and always exists throughout the model integration. Systematic bias in the TPSC therefore occurs and accumulates with the model integration time due to exaggeration of the precipitation in the models.
The increasing TPSC is accompanied by decreasing surface air temperature
(SAT) with forecast lead time. The SAT over the Tibetan Plateau becomes
colder with increasing forecast lead time in the S2S models. The differences
in SATs between the forecast for a lead of 5 weeks and the forecast for a
lead of 1 week can be up to 1.9
Land–atmosphere coupling is one of the key physical processes for S2S prediction but is not well simulated and may reduce S2S prediction skill (Robertson et al., 2014; Dirmeyer et al., 2019). Studies have shown that better snow cover initialization improves subseasonal and seasonal forecasts/simulations (Jeong et al., 2013; Orsolini et al., 2013; Senan et al., 2016; Lin et al., 2016; Kolstad, 2017; F. Li et al., 2019). This study indicates that in addition to snow cover initialization, a better model skill for snow cover prediction may also improves S2S prediction skill. More work is necessary and valuable to improve the prediction ability of models for snow cover.
Previous studies have shown that current state-of-the-art GCMs tend to strongly overestimate the precipitation over the Tibetan Plateau (e.g., Su et al., 2013; Chen and Frauenfeld, 2014; Zhang and Li, 2016; Zhang et al., 2019). It is worthwhile to note that the S2S models also significantly overestimate the precipitation over the Tibetan Plateau and further cause other biases (e.g., TPSC biases and SAT biases). It is of great significance to reduce the biases of the precipitation over the Tibetan Plateau in the GCMs. Surface winds and snow sublimation could also play a role in causing the snow ablation. Identifying the relative contributions of these factors to the biased snow prediction needs more detailed and careful diagnoses. Note that the current study analyzed the data during common reforecast period of 1999–2010 for ECMWF, NCEP and CMA models. All these three operational models provide real-time forecasts since 2015 based on the improved prediction systems. It could be valuable to carry out evaluation works based on the up-to-date forecast results. Future studies on these issues are potentially valuable.
The data and model used in this study are free to the public. The S2S datasets and ERA-Interim data are available at
WL led the overall scientific questions and designed the research. SH and WL analyzed the data and drafted the manuscript for initial submission. WL analyzed the data for the revised manuscript. WL, PH, WG and JW made substantial contributions to revise the manuscript and prepare the responses to the referees.
The authors declare that they have no conflict of interest.
This research has been supported by the National Key Research and Development Program of China (grant no. 2018YFC1505804), the Natural Science Foundation of China (grant no. 41905074), and the Natural Science Foundation of Jiangsu Province (grant no. BK20190782).
This paper was edited by Mark Flanner and reviewed by two anonymous referees.