-
With the rapid development of economic construction in recent decades, South China has become one of the regions with the highest economic level in China and experienced yearly increased losses caused by meteorological disasters. Especially during the first rainy season (FRS) from April to June (e.g., Huang, 1986; Deng and Wang, 2002; Chi et al., 2005; Ding et al., 2009; Luo, 2017), severe convective weather accompanied by hail, thunderstorms, and tornadoes occurs frequently and has caused many flood disasters, landslides, and debris flows, which endanger lives and properties in South China (e.g., Xu, 1994; Liang, 1997; Hallegatte et al., 2013). For instance, the heavy rainfall that happened in southern China during June 2022 caused a huge loss of 4 billion yuan and 32 lives (e.g., Gan and Deng, 2022). However, in the current Rainy Season Flood Outlook Consultations organized by national operational centers every year, real-time seasonal forecasts have mainly focused on abnormal summer precipitation in the middle and lower reaches of the Yangtze River (MLYR). But, little attention has been paid to the interannual variations of FRS precipitation over South China. Therefore, it is of great importance and socio-economic value to carry out seasonal predictions of precipitation during the FRS in South China. However, owing to the specific geographical location and complex terrain, interactions between large-scale circulation systems originating from the low latitudes and mid-high latitudes (e.g., Zhu et al., 2014; Lu et al., 2021), as well as the local transient mesoscale systems, could result in complicated processes for the precipitation there (e.g., Luo et al., 2013; Jiang et al., 2017; Tang et al., 2021); this leads to great difficulty in predicting FRS precipitation in South China.
FRS precipitation in South China is mainly composed of frontal rainfall and warm-sector rainfall (e.g., Chi et al., 2005; Chang et al., 2006; Zheng et al., 2006; Liu et al., 2019). The FRS can be generally divided into two stages (e.g., Chi et al., 2005): the period before the onset of the South China Sea monsoon, from about April to mid-May (e.g., Xie et al., 1998; Wang et al., 2004), and the post-monsoon-onset period, during which the precipitation increases significantly (e.g., Luo et al., 2013; Li et al., 2020). Previous studies have demonstrated that the areas of water vapor source during the FRS consist of the Indian Ocean, the Pacific Ocean, the South China Sea, and the Chinese mainland (e.g., Chi et al., 2005; Zhong et al., 2019; Chu et al., 2020; Peng et al., 2022). FRS precipitation displays significant interannual variation, which is closely related to tropical sea surface temperatures (SSTs) (e.g., Luo et al., 2020). A positive temperature gradient between the tropical eastern and western Pacific is conducive to suppressed convection in the tropical western Pacific, which could excite Rossby waves and result in an abnormal anticyclone circulation near the Philippines, thus reducing the moisture transport to South China (e.g., Gu et al., 2018).
In addition to tropical Pacific SST, anomalous SST in the tropical Indian Ocean basin may also have great impacts on precipitation in South China (e.g., Chowdary et al., 2011; Yuan et al., 2019). Southwesterly winds associated with the cold SST anomaly in the South Indian Ocean could help transport more water vapor to the subtropical western Pacific and nearby regions and induce significant low-level convergence in South China (e.g., Jia et al., 2021). Model experiment results have suggested that the tropical and North Indian Ocean warming generates Kelvin waves and induces an anticyclone in the northwestern Pacific (e.g., Xie et al., 2009), affecting the moisture transport to South China. Li et al. (2018) reported that the North Atlantic tripole-like SST variation could influence precipitation in South China by stimulating teleconnection wave trains over Eurasia in the mid-high latitudes, and tropical Atlantic SST anomalies could change the Walker Circulation and induce abnormal circulations near the Philippines and precipitation anomalies in South China. In addition, abnormal snow cover on the Tibet Plateau and the Northeast China cold vortex may also influence FRS precipitation (e.g., Wu and Qian, 2003; Miao et al., 2006; Li et al., 2014; Xiao and Duan, 2016).
It is believed that the physical basis for the seasonal predictability of precipitation lies in the external forcing, especially the anomalous SST and its related climate signals. Thus, the performance of SST anomaly prediction in climate models is important for the seasonal prediction of precipitation. The global real-time Climate Forecast System version 1.0 of Nanjing University of Information Science and Technology (NUIST-CFS1.0), developed based on the SINTEX-F model of the Japanese Marine Science and Technology Development Agency (JAMSTEC), has shown outstanding performance for predicting major tropical climate signals, such as achieving satisfactory forecasts of ENSO at lead times of up to 1.5–2 years and for the Indian Ocean dipole (IOD) up to 1–2 seasons in advance (Luo et al., 2007, 2008, 2016; He et al., 2020). In addition, NUIST-CFS1.0 also displays reasonably good skill in predicting the features of summer precipitation anomalies in the middle and lower reaches of the Yangtze River (Ying et al., 2022). However, the skill of NUIST-CFS1.0 and an inter-comparison with other international climate forecast systems in predicting the precipitation in the FRS over South China remain to be assessed.
Due to the special geographical location, the climate in East Asia is not only affected by large-scale factors, but also small-scale factors such as regional-scale topographic forcing and mesoscale eddies. The resolutions of most of the current global general circulation models (GCMs) are still too coarse to meet the needs of fine predictions and socioeconomic applications (e.g., Wehner et al., 2010; White et al., 2013; Bao et al., 2015). A dynamical downscaling method using higher-resolution regional climate models (RCMs), in which the initial and boundary conditions are provided by GCMs, is often employed to derive finer predictions and to improve the prediction of regional climate (e.g., Lo et al., 2008; Xu and Yang, 2015; Kitoh et al., 2016; Tang et al., 2016; Dai et al., 2020). The RCMs can not only retain the large-scale characteristics given by GCMs, but also capture more regional climate characteristics (e.g., Dickinson et al., 1989; Giorgi et al., 1994; Fu et al., 2005; Xue et al., 2007; Sato and Xue, 2013; Wang et al., 2013; Sun et al., 2018). Zhang et al. (2015) nested one RCM (i.e., Reg CM4.4) into the fourth version of the Community Climate System Model (CCSM4.0) and found that Reg CM4.4 could improve the forecast skill for precipitation, with the spatial correlation coefficient skill increasing from 0.39 to 0.53. Gao et al. (2008) compared two RCMs nested within the NASA/NCAR global model and found that one RCM can perform better in reproducing the precipitation change in China in terms of both spatial pattern and rainfall amount. The results suggest that higher resolution RCMs help better capture precipitation patterns over China, especially during the monsoon season. In addition, Ratnam et al. (2016) found that the forecast skill for austral summer precipitation over South Africa can be improved by downscaling the SINTEX-F2 global forecast with the Weather Research and Forecasting (WRF) model.
In this study, we selected Guangdong, Guangxi, Fujian, and Hainan provinces as the study area according to the Operational Monitoring of the First Rainy Season of South China (National Climate Center, 2013). The prediction skill of NUIST-CFS1.0 for precipitation in the FRS over South China is evaluated and is compared with nine international climate model forecast systems (see Table 1), and possible sources of the seasonal predictability for precipitation are also explored using the NUIST-CFS1.0 ensemble hindcasts for the period 1982–2020. Furthermore, we use WRF to dynamically downscale the global model forecasts to estimate whether this downscaling method can improve the prediction skill for precipitation in the FRS over South China. Section 2 briefly introduces the forecast model, the observational/reanalysis data, and evaluation methods. Section 3 and section 4 describe the assessments of forecast skill for FRS precipitation in South China based on NUIST-CFS1.0 hindcasts with a brief comparison with nine international models. In section 5, dynamically downscaled results are analyzed. Finally, section 6 gives a brief summary and discussion.
System
nameAtmosphere
ModelOcean
ModelSea ice
ModelHindcast
ensemble sizeECMWF
_SEAS5IFS Cycle 43r1
(Dynamics: TCO319
Physics: O320)NEMO v3.4
(ORCA 0.25L75)LIM2 25 JMA/MRI
-CPS2JMA-GSM
(TL159L60)MRI.COM v3
1°×(3°−5°)Part of
MRI.COM v310 GloSea5
-GC2Met Office UM-
Global Atmosphere
6.0(N216L85)NEMO v3.4
-Global Ocean5.0
(OCRA 0.25L75)CICE v4.1
-Global Sea
-Ice 6.028 METFRA
_System6ARPEGE v6.2
(TL359L91)NEMO v3.4
(OCRA 1L75)GELATO v6 25 METFRA
_System7ARPEGE v6.4
(TL359L91)NEMO v3.6
(OCRA 0.25L75)GELATO v6 25 CMCC
_SPS3CAM5.3
(1°L46)NEMO v3.4
(0.25°L50)CICE 4.0 40 CMCC
_SPS3.5CAM5.3
(0.5°L46)NEMO v3.4
(0.25°L50)CICE 4.0 40 GCFS2.0 ECHAM 6.3.05
(T127L95)MPIOM 1.6.3
(TP04L40)Same as
ocean model30 GCFS2.1 ECHAM 6.3.04
(T127L95)MPIOM 1.6.3
(TP04L40)Same as
ocean model30 Table 1. Nine international climate models used in this study.
-
Figure 1 presents the climatological mean precipitation and standard deviation of interannual precipitation anomaly in the FRS during 1982–2020 based on the observations and NUIST-CFS1.0 ensemble mean hindcasts. The observed climatology displays two precipitation centers in north Fujian and central Guangdong province, respectively, with the maximum reaching up to 10 mm d–1. Interestingly, areas with stronger climatological mean precipitation also display larger interannual variations, indicating that these areas are also prone to flooding disasters. NUIST-CFS1.0 could predict the precipitation center in north Fujian province well, albeit with a relatively weak intensity, but the model fails to capture the other precipitation center in central Guangdong province. In addition, compared to the observations, NUIST-CFS1.0 generally underestimates the magnitude of the climatological mean precipitation and the standard deviation of interannual variations in South China. This bias is commonly found in the ensemble mean prediction of precipitation using global climate models. That is, due to the too coarse resolution of the global climate models, small- and medium-scale processes that are important to precipitation are hardly reproduced (e.g., Xu et al., 2009). Figure S1 in the electronic supplementary material (ESM) shows the climatology mean precipitation and interannual standard deviation based on the nine-member forecast initiated from the 1st day of March. The nine-member forecasts show great spread of the skill in predicting the regionally averaged precipitation anomaly. The member with the highest TCC skill of 0.38 adopts a fully coupling scheme to transfer the momentum of the sea surface current directly to the atmosphere and weakest SST nudging coefficient. The exact reasons for the skill spread among the nine-member forecasts are still unknown and warrant future investigations.
Figure 1. The spatial distribution of climatological mean precipitation (shading) and standard deviation of interannual precipitation anomaly (black line) (units: mm d−1) in the first rainy season (FRS, i.e., April–June) during 1982–2020 over South China based on (a) the observations and (b–d) NUIST-CFS1.0 nine-member ensemble mean predictions initiated from the 1st day of March, February, and January, respectively. White dashed lines denote the borders of Guangdong, Guangxi, and Fujian provinces.
To measure the potential predictability of precipitation in the FRS over South China, we calculated the signal-to-noise ratio, which is defined as the ratio of the variance of the ensemble mean predictions to the mean ensemble spread among the nine member forecasts (Ehsan et al., 2017, 2020). The ensemble mean predictions generally represent the atmosphere response to the external forcing (such as SST), while the ensemble spreads correspond to the atmospheric internal variations induced by the strong chaotic dynamics of atmosphere. Thus, the larger the value of the signal-to-noise ratio, the higher the potential predictability of the precipitation anomaly. As shown in Figs. 2a–c, the signal-to-noise ratios of NUIST-CFS1.0 forecasts initiated from the 1st day of January, February, and March are all much smaller than 1 over South China, suggesting that the atmosphere internally driven variation of precipitation during the FRS is greater than the variation forced by the external forcing. In other words, reliable predictions of the precipitation in the FRS over South China remain difficult and fairly challenging.
Figure 2. The spatial distribution of the signal-to-noise ratio (S/N Ratio) (top panels) and the correlation coefficients (bottom panels) of the FRS precipitation anomaly during 1982–2020 over South China based on the NUIST-CFS1.0 ensemble mean predictions initiated from the 1st day of (a) March, (b) February, (c) January. White dashed lines denote the borders of Guangdong, Guangxi, and Fujian provinces.
The TCC, PCC, and regionally averaged FRS precipitation anomaly over South China (i.e., the area of Guangdong, Guangxi, Fujian, and Hainan provinces as defined above) are calculated to evaluate the forecast skill (Figs. 2d–f and 3). The highest TCC score of up to 0.4 appears in Fujian, followed by 0.3 in Guangxi, while the lowest score of –0.2 appears in part of Hainan province. Although the TCC skill is somewhat dependent on lead time in general, forecasts from the three different initial months show a similar distribution of the prediction skill in South China (Figs. 2d–f). Interestingly, the forecasts initiated from 1 March (i.e., at one-month lead) do not outperform those initiated from 1 February and 1 January (i.e., at two- and three-month leads). This is in contrast to the forecasts of SST that usually display better skill at shorter lead times (e.g., Luo et al., 2016). This may be related to the fact that FRS precipitation is influenced by strong atmosphere internal variability, which can induce a random-type change in the prediction skill at different lead times. In addition, biases and uncertainties in model physics and initial conditions may also play a role. Another possibility is that impacts of preceding signals such as winter SST anomalies on FRS precipitation (e.g., Feng et al., 2011; Feng and Li, 2011; Cheng and Jia, 2014) may be better captured in the predictions initiated from January and February, compared to that from March. The rapidly growing monsoon and the attenuated ENSO signal in springtime may also increase uncertainty in predicting the tropical air–sea interactions in a coupled ocean–atmosphere forecast model (Webster and Yang, 1992; Yang et al., 2018). The complexity regarding FRS precipitation prediction over South China warrants thorough investigation in future studies.
The PCC score acts to measure how well the spatial distribution of the precipitation anomalies over South China are predicted. NUIST-CFS1.0 performs well in some years such as 2004, 2010, and 2011, with the highest PCC value reaching above 0.8 (Fig. 3a). However, the PCC score can also be very poor in some other years, for instance, below –0.8 in 1991. This means that NUIST-CFS1.0 may predict almost an opposite spatial pattern of precipitation anomalies compared to the observations in some years. The large variation of the PCC from year to year indicates that the precipitation anomalies in the FRS over South China can be driven by fairly distinctive mechanisms in different cases. Figure 3b shows the interannual variation of the regionally averaged precipitation anomaly in South China during 1982–2020. The magnitudes of the predicted anomalies are generally weaker than the observations, partly because the ensemble mean acts to smooth out the atmospheric internal signals. In the years with strong El Niño events (like 1998), the predicted magnitudes at the three different lead times are close to the observations. In contrast, when the precipitation over South China during the FRS is mostly resulted from medium- and small-scale processes, such as in 2008 (e.g., Wang et al., 2011), it is difficult to be predicted by global climate models, although some members of the model ensemble forecasts could provide relatively better results. In short summary, the above results suggest that the seasonal forecast skill for FRS precipitation in South China assessed based on NUIST-CFS1.0 ensemble forecasts for 1982–2020 is relatively low, which is consistent with the low signal-to-noise ratio.
Figure 3. (a) The spatial correlation coefficients of the FRS precipitation anomaly during 1982–2020 in South China between the observations and NUIST-CFS1.0 ensemble mean predictions initiated from the 1st day of March (pink bar), February (blue bar), January (gray bar). (b) The regionally averaged precipitation anomaly (units: mm d−1) in South China based on the observations (black line), NUIST-CFS1.0 ensemble mean predictions initiated from the 1st day of March (pink bar), February (blue bar), January (gray bar). Error bars indicate the minimum-to-maximum range of the nine-member forecasts. Note that South China is defined as the area of Guangdong, Guangxi, Fujian, and Hainan provinces in this study.
To better understand the predictability of the FRS precipitation anomaly in South China, we further assessed nine international climate models’ forecast skills. The nine models’ ensemble hindcast data was downloaded from Copernicus Climate Change Service (as shown in Table 1), and many of the models were developed by operational centers for issuing real-time forecasts. We selected their hindcast outputs for 1993–2016, which is available for all the models. Here, we briefly compare the prediction skills based on the hindcasts initiated from 1 March (i.e., one-month lead). Figure 4 shows the TCC skill in predicting FRS precipitation anomalies over South China based on nine international climate models’ hindcasts initiated from 1 March from 1993 to 2016. The results show great variation in prediction skill among the models. Some of them produce positive skill over a large part of South China, reaching as high as 0.6 in local regions (e.g., CMCC_SPS3.5). In contrast, some of the models produce weak and even negative skill in many parts of South China (e.g., JMA/MRI_CPS2). The regionally averaged TCC skill, the 24-year mean PCC skill, and the TCC skill in predicting the regionally averaged precipitation anomaly over South China for 1993–2016 based on NUIST-CFS1.0 are better than the skills of most of the nine international models (Table 2). As was discussed in Luo et al. (2008), a good dynamical model forecast system must meet the following four elements: good performance in simulating the phenomenon/predictand, realistic and well-balanced initial conditions that help reduce initial shock, good ensemble scheme that can measure uncertainties/errors in model physics and initial states, and realistic representation of external radiative forcing. The difference in the prediction skill among the different model forecast systems is generally a result of their different performance on the above four elements. Exploring detailed reasons for these results requires more data, sensitivity experiments, and future investigations. It is worth noting that the prediction skill of almost all the models is lower than 0.5, indicating the great challenge in predicting FRS precipitation in South China. In addition, the skill of NUIST-CFS1.0 for 1993–2016 is higher than that during 1982–2020, suggesting a potential decadal change in FRS precipitation predictability.
Figure 4. The correlation coefficients of the FRS precipitation anomaly during 1993–2016 over South China between the observations and ensemble mean hindcasts initiated from the 1st day of March based on (a) ECMWF_SEAS5, (b) JMA/MRI_CPS2, (c) GloSea5_GC2, (d) METFRA_System6, (e) CMCC_SPS3, (f) GCFS2.1, (g) METFRA_System7, (h) CMCC_SPS35, and (i) GCFS2.0. White dashed lines denote the borders of Guangdong, Guangxi, and Fujian provinces.
Model TCC PCC Regionally averaged anomaly The anticyclone index NUIST-CFS1.0 0.29 0.24 0.38 0.67 WRF 0.11 0.07 0.14 ECMWF_SEAS5 0.22 0.06 0.32 0.68 JMA/MRI-CPS2 −0.08 −0.03 −0.24 0.34 GloSea5-GC2 0.30 0.23 0.41 0.32 METFRA_System6 0.12 0.08 0.26 0.41 METFRA_System7 0.15 0.14 0.21 0.56 CMCC_SPS3 0.18 0.26 0.49 0.73 CMCC_SPS3.5 0.31 0.27 0.51 0.77 GCFS2.0 0.13 0.07 0.21 0.63 GCFS2.1 0.13 0.04 0.22 0.63 Table 2. List of the regionally averaged temporal correlation coefficient (TCC), and the 24-year mean pattern correlation coefficient (PCC) and the correlation coefficient of regionally averaged precipitation anomaly in the FRS over South China and the anomalous anticyclone index in the tropical northwestern Pacific between 10 global models’ forecasts and the observations during 1993–2016. The skill of WRF downscaling predictions are also listed. Note that South China is defined as the area of Guangdong, Guangxi, Fujian, and Hainan provinces in this study.
System name | Atmosphere Model | Ocean Model | Sea ice Model | Hindcast ensemble size |
ECMWF _SEAS5 | IFS Cycle 43r1 (Dynamics: TCO319 Physics: O320) | NEMO v3.4 (ORCA 0.25L75) | LIM2 | 25 |
JMA/MRI -CPS2 | JMA-GSM (TL159L60) | MRI.COM v3 1°×(3°−5°) | Part of MRI.COM v3 | 10 |
GloSea5 -GC2 | Met Office UM- Global Atmosphere 6.0(N216L85) | NEMO v3.4 -Global Ocean5.0 (OCRA 0.25L75) | CICE v4.1 -Global Sea -Ice 6.0 | 28 |
METFRA _System6 | ARPEGE v6.2 (TL359L91) | NEMO v3.4 (OCRA 1L75) | GELATO v6 | 25 |
METFRA _System7 | ARPEGE v6.4 (TL359L91) | NEMO v3.6 (OCRA 0.25L75) | GELATO v6 | 25 |
CMCC _SPS3 | CAM5.3 (1°L46) | NEMO v3.4 (0.25°L50) | CICE 4.0 | 40 |
CMCC _SPS3.5 | CAM5.3 (0.5°L46) | NEMO v3.4 (0.25°L50) | CICE 4.0 | 40 |
GCFS2.0 | ECHAM 6.3.05 (T127L95) | MPIOM 1.6.3 (TP04L40) | Same as ocean model | 30 |
GCFS2.1 | ECHAM 6.3.04 (T127L95) | MPIOM 1.6.3 (TP04L40) | Same as ocean model | 30 |