Decadal Prediction of the Summer Extreme Precipitation over Southern China

Abstract

The decadal variability of the summer extreme precipitation over southern China (EPSC) is remarkable, especially for the significant decadal enhancement after the 1990s. The study documented that the summer sea surface temperature (SST) over the North Atlantic and spring sea ice concentration (SIC) over the East Siberian Sea can significantly affect the EPSC. The summer SST over the North Atlantic influences the low-pressure cyclone in the western Pacific by modulating the SST over the tropical Pacific, thus affecting EPSC. A decrease in the SIC of the East Siberian Sea induces a negative Arctic Oscillation, which induces the increased SST over northwest Pacific and the anomalous cyclone over there, in turn, affecting EPSC. Both predictors have a quasi-period of 10–14 years, which provides useful predictive signals for EPSC. The leading 7-year SST and the leading 5-year SIC are chosen to establish the prediction model based on the decadal increment method, which can well predict the EPSC, especially for the shift in the early 1990s. These results provide a clue to the limited predictability of decadal-scale extreme climate events.

1. Introduction

Extreme precipitation events have attracted increasing attention due to their large impact on both the ecological environment and the social economy [1,2,3]. In China, extreme precipitation events occur more frequently in southern China than in other regions because the abundant rain over there [1,4]. Summer (June–August) precipitation in southern China accounts for more than 45% of the total annual precipitation. Summer also has the highest frequency of extreme precipitation events, which can lead to flooding and tremendous damage to the local economy and ecosystems [5]. For example, in mid-to-late June of 2005, persistent extreme precipitation occurred in South China with daily precipitation exceeding 400 mm, which affected 2.627 million people and caused economic losses of 3.9 billion yuan [6]. Besides the remarkable interannual variability, summer extreme precipitation over southern China has shown a significant increasing trend with a decadal enhancement in the early 1990s [7,8,9]. Hence, the effective and reliable prediction of the summer extreme precipitation over southern China is of great importance.

Numerous studies have revealed possible mechanisms for extreme precipitation over southern China on interannual timescales. Located in the southern coastal region of China, southern China is influenced by tropical and mid- to high-latitude climate systems such as tropical cyclones, the western North Pacific subtropical high, the East Asian subtropical westerly jet, the Asian summer monsoon, and so on [4,8,10,11]. These circulation systems can influence the extreme precipitation over southern China via the transport of water vapor and vertical motion over South China. External forcings have also been suggested to affect the extreme precipitation over southern China. The anomalous sea surface temperature (SST) of the South China Sea and the adjacent western Pacific warm pool are linked to the extreme precipitation over southern China [12]. The variations in the SST of the Indian Ocean not only favor the maintenance of anticyclones around the Philippine Sea [13], but also affect the Pacific–Japan teleconnection pattern [14]. This, in turn, affects anomalous extreme precipitation over southern China [14,15]. The ENSO (El Niño/Southern Oscillation) and the Pacific Decadal Oscillation (PDO) are also found to be related to the decadal variations in the early summer rainfall over southern China [16]. The spring Arctic Sea ice concentration (SIC) also favors the extreme precipitation over southern China via the Eurasian teleconnection [17]. The sea ice loss along the Siberian coast influenced the precipitation anomaly in the summer of 2020 over central-east China through excessive atmospheric blockings [18].

In terms of the decadal timescale, the SST remains an important influencing factor for the decadal variability of extreme precipitation over southern China (EPSC). Zhang et al. linked the EPSC during the pre-rainy season to the decadal variation of the ENSO-like SST, which can influence the transport of water vapor via cyclonic and anticyclonic systems [19]. The anomalous SST in the western Pacific warm pool is also an important factor driving the increase in EPSC [20]. Additionally, anomalies in the latent heat flux over the South China Sea and the sensible heat flux over the Indochina peninsula are also considered important factors in EPSC [7]. However, few studies have focused on the influences of extratropical forcings on the EPSC. Moreover, the prediction of EPSC has rarely been operated. This study focuses on the mechanisms of extreme precipitation over southern China on a decadal timescale and attempts to predict its decadal variability.

Decadal prediction focuses on climate predictions over the next 1 to 10 years, which has attracted widespread attention due to its impacts on economic and social development [21]. Currently, making skillful decadal predictions based on the initialized climate models to predict precipitation over East Asia is full of challenges, especially for extreme precipitation [22,23,24].

Currently, decadal predictions remain challenging, and one of the important scientific issues is that the improvement of initialized climate models requires a lot of human and material resources as well as the innovation and development of related theories [23,25]. Meanwhile, some scholars have used statistical methods for decadal prediction and achieved high predictive skills. Recently, based on the interannual increment method [26], Huang and Wang proposed a new statistical approach called the decadal increment method, which can obtain decadal information by performing a 5-year running smoothing [27,28]. A statistical forecast model is then built with the 3-year decadal increment predictors, and the final prediction results can be obtained by adding the predicted increment to earlier observations. This method can increase the effective samples based on the form of decadal increment and gain useful decadal signals from the previous observations, which has shown high predictive skills in the decadal prediction of precipitation over North China, the PDO, and the East Asian summer monsoon [27,28,29]. Considering the frequent occurrences of extreme precipitation over southern China, this paper will first analyze the causes of the EPSC and then try to build a prediction model based on the decadal increment method. Additionally, a real-time prediction will also be given.

2. Data and Methods

2.1. Data

The daily precipitation data are from the CN05.1 gridded dataset based on 2416 station observations over China with a horizontal resolution of (0.25° × 0.25°) and cover the time period 1961–2021 [30]; the data are used to calculate the extreme precipitation index. The monthly mean SST data are obtained from the National Oceanic and Atmospheric Administration (NOAA) Extended Reconstructed SST V5 dataset with a horizontal resolution of (2° × 2°) and cover the time period 1954–2021 [31]. The monthly sea ice concentration (SIC) data are from the UK Met Office Hadley Centre with a horizontal resolution of (1° × 1°) and cover the time period 1954–2021 [32]. Both the SST data and SIC data are used for determining the predictors. The monthly sea-level pressure (SLP), meridional and zonal winds in 850/500/10 hPa level, and integrated water vapor flux data are obtained from the ERA5 reanalysis dataset of the European Centre for Medium-Range Weather Forecasts with a horizontal resolution of (1° × 1°) and cover the time period 1950–2021 [33]; they are used to analyze the mechanisms of SST and SIC affecting the EPSC. The monthly AMO data during 1946–2021 and PDO data during 1946–2019 are obtained from the NCAR Climate Data and are used to analyze their potential correlation with the EPSC.

2.2. Definition of the Extreme Precipitation Index

According to the recommendation of the Expert Team on Climate Change Detection and Indices (ETCCDI, http://cccma.seos.uvic.ca/ETCCDI/indices.shtml, accessed on 15 March 2021), the summer extreme precipitation index (R95p) is defined as the total amount of precipitation on days with daily precipitation (daily precipitation ≥ 0.1 mm) larger than the 95th percentile of the summer (June–August) daily precipitation.

2.3. Methods

The decadal variability of extreme precipitation is obtained by performing a 5-year running mean of regional averaged extreme precipitation. The 5-year running mean of regional averaged extreme precipitation over southern China is defined as the decadal variability of the southern China extreme precipitation (EPSC) index. The 3-year decadal increment of the EPSC (DI_EPSC) is the predictand, which is defined as the EPSC of the current year minus the EPSC at 3 years ago (Equation (1)). The final predicted EPSC can be obtained by adding the predicted DI_EPSC to the observed EPSC from 3 years ago (Equation (2)). The main predictors are determined by the stepwise regression with the F-test at the 95% confidence level, which must be well-correlated with the predictand and independent of each other. The multiple linear regression method is applied to build the statistical model with the determined predictors in the form of a 3-year decadal increment. To prevent the prediction model from using information from any prediction period, predictors are at least 3 years ahead of the predictand.

$$DI\text{\_}EPS{C}_i=EPS{C}_i-EPS{C}_{i-3}$$

(1)

$$EPS{C}_i=EPS{C}_{i-3}^{obs}+DI\text{\_}EPS{C}_i^{s-model}$$

(2)

To verify the prediction skills of the statistical model, cross-validation and independent hindcast are applied. In the cross-validation, the observational data of the EPSC during 1966–2018 are selected to predict the DI_EPSC for the target year. Noteworthily, the statistical model is built by leaving 5 years out (from 2 years before to 2 years after the target year for the DI_EPSC and the leading 5 years for the predictors). The same observational data are also used in the independent hindcast. The statistical model is established to predict the DI_EPSC for the target year with the data from 43 years to 3 years leading the target year. This process is repeated from 2009 to 2018. For instance, the predictand during 1966–2006 and the predictors during 1963–2003 are chosen to build the statistical model to predict the DI_EPSC for 2009.

Empirical orthogonal function (EOF) analysis is a linear data transformation defined in terms of the eigenvectors of their covariance matrix [34]. In this study, it is applied to investigate the decadal variation of extreme precipitation over southern China.

The significance of the correlation coefficient is examined by the Student’s t-test. The effective sample size N* is calculated as follows [35]:

$$N^{*}=N\frac{1-r_1{r}_2}{1+r_1{r}_2}$$

(3)

where N is the number of available time steps, and r₁ and r₂ are the autocorrelation coefficients of the two variables lagged by one step.

The mean square skill score (MSSS) is used to measure the predictive skill of the statistical model [36] and is calculated as follows:

$$MSSS=1-\frac{MSE}{MS{E}_c}$$

(4)

$$MSE=\frac{1}{n}{\displaystyle \sum }_{i=1}^n{\left(f_i-o_i\right)}^2$$

(5)

$$MS{E}_c=\frac{1}{n}{\displaystyle \sum }_{i=1}^n{\left(o_i-\overline{o}\right)}^2$$

(6)

where $f_i$ and $o_i$ represent the time series of the observations and forecasts, respectively. $MSSS$ is the percentage reduction of the mean square error (MSE) predicted by the statistical model, and $MS{E}_c$ is the MSE of the climatological forecast. A positive skill means that the statistical model forecast is better than the climatology forecast.

3. Results

3.1. Decadal Variation

Figure 1 shows the spatial pattern of the first EOF mode of the decadal variation of summer extreme precipitation in East China during 1963–2018. The spatial pattern of the first EOF mode can explain 21.6% of the total variance of summer extreme precipitation in East China. There is abundant extreme precipitation occurred over southern China compared to other regions (Figure 1a), which has been noticed by previous studies [1]. Since there is a uniform spatial pattern over southern China (22°–30° N, 107°–120° E), the 5-year running mean of regional averaged summer extreme precipitation over southern China is defined as the decadal variability of the extreme precipitation over southern China (EPSC) index (Figure 1b). The EPSC has an increasing trend with a significant shift after the 1990s [7,8,9]. The time series of the leading EOF mode (PC1) correlates well with the EPSC with a correlation coefficient of 0.94, indicating that the EPSC index can reasonably reflect the decadal variability of extreme precipitation over southern China. Moreover, the 3-year decadal increment of the EPSC (DI_EPSC) is then calculated, which will be treated as the predictand in the following section.

3.2. Predictors

This section mainly discusses the predictors of EPSC and the associated mechanisms. Based on previous studies, potential predictors are chosen mainly from external forcings, including the SST, the SIC, the soil moisture, and the snow depth. The stepwise regression with the F-test at the 95% confidence level is used to determine the main predictors of EPSC. The determined decadal predictors must be well-correlated with the predictand and independent of each other. Two final predictors, including the summer SST of the North Atlantic leading the DI_EPSC by 7 years and the spring SIC near the East Siberian Sea leading the DI_ EPSC by 5 years, which are well-correlated with the DI_EPSC and independent of each other, are therefore chosen to predict the decadal summer EPSC. The physical processes of these two predictors affecting the EPSC will be explored in detail.

Previous studies have shown that triple or dipolar SST anomalies in the North Atlantic can induce a wave train over the North Atlantic through Eurasia, which is one of the factors influencing precipitation and extreme precipitation in eastern China [37,38,39,40]. Additionally, anomalous warming in the North Atlantic can also induce cooling in the tropical eastern Pacific Ocean and warming in the Indo-Pacific Ocean through a series of air–sea interactions [22,40,41,42,43,44]. This is a significant factor in maintaining an anticyclone over the western Pacific and transporting water vapor to eastern China [40], thus affecting extreme precipitation over southern China.

Notably, the leading 7-year summer (June–August) SST over the North Atlantic is significantly related to the EPSC (Figure 2a). There significant positive and negative centers emerged in the mid- to high-latitude regions of the North Atlantic, indicating the influence of the SST dipole over the North Atlantic ahead 7 years on the EPSC. The SST index (DI_SST) is therefore defined as the regional averaged decadal increment in the SST over the positive region (26°–39° N, 43°–60° W) minus that of the negative region (49°–57° N, 14°–49° W).

How do the SST anomalies affect EPSC 7 years in advance? The wavelet analysis of the DI_SST during 1959–2011 is shown in Figure 3a, which indicates that the dipole anomalies of the North Atlantic SST have a quasi-period of 12–14 years (Figure 3a) with a significant negative autocorrelation of 5–7 years (Figure 3b). Both the period and significant auto-correlation imply that the dipole SST anomalies of the North Atlantic can provide useful predictive information for EPSC 7 years ahead.

What are the specific physical processes by which the SST dipole pattern over the North Atlantic affects EPSC?

Figure 4a shows the spatial pattern of the correlation coefficient between the DI_SST and surface winds. The SST dipole over the North Atlantic can lead to significant anomalous easterlies from the tropical North Atlantic to the tropical East Pacific (as shown in the rectangle region of Figure 4a). The northeasterly/easterly anomalies (as shown in the rectangle region of Figure 4b) favor cooling in the tropical eastern Pacific and warming in the western Pacific SST (Figure 4b). This SST pattern can enhance the Walker circulation, which is characterized by descending motion over the eastern Pacific and ascending motion over the western Pacific (Figure 4c). There are associated cyclonic anomalies over the northwestern Pacific in both low troposphere to mid troposphere (Figure 4d,e). In the west of the cyclone, significant northeasterly winds prevail over southern China. In the mid troposphere, there are significant easterly winds from southern China to Tibetan Plateau (Figure 4e), which may generate descent motion (Figure 4f) through the zonal temperature gradient between southern China and Tibetan Plateau [45,46,47]. Both the northeasterly winds and descending motion is unfavorable for the increase of EPSC [48,49]. Thus, the preceding 7-year SST dipole pattern over the North Atlantic is chosen to be one predictor for EPSC based on its significant influence and period of 12–14 years.

In addition, previous studies have also explored the connection between the decrease in spring Arctic SIC and the increase in summer precipitation even extreme precipitation in East and southern China [17,18,50]. The decrease in Arctic SIC can induce a Eurasia teleconnection, which is inductive to transport water vapor to southern China [17]. In this paper, the leading 5-year spring SIC near the East Siberian Sea also influences the EPSC. Figure 2b shows a significant negative center near the East Siberian Sea, which indicates the influence of the SIC over the East Siberian Sea ahead of 5 years on EPSC. Hence, the SIC index (DI_SIC) is defined as the regional averaged decadal increment in the SIC over the negative region (71°–77° N, 149°–157° E).

How do the SIC anomalies influence EPSC 5 years in advance? As shown in Figure 5, the DI_SIC has a quasi-period of 10 years with a significant negative autocorrelation of 5–6 years (Figure 5b). Both the period and the significant autocorrelation indicate that the SIC anomalies of the East Siberian Sea can also provide useful predictive information for EPSC 5 years in advance.

How can the SIC over the East Siberian Sea physically influence EPSC?

Figure 6a shows the spatial pattern of the correlation coefficient between the DI_SIC and the spring SLP, which is similar to the Arctic Oscillation (AO) pattern. When the SIC in the East Siberian decreases, the local surface temperature increases, which makes the temperature differences between the pole and mid latitudes’ decrease, and AO accordingly shows a negative phase. This is consistent with previous studies [51,52]. When SIC in the East Siberian decreases, there appears to be a significant warming of the anomalous SST near the North Pacific under the influence of AO (Figure 6b). The anomalous increased SST can store useful information from spring to summer, which, in turn, induces cyclonic anomalies over the northwest Pacific (Figure 6c). The northerly in the west of the cyclone can weaken the water vapor transport to southern China, where the water vapor flux appears to be significantly divergent (Figure 6e). Figure 6d shows the negative correlation of DI_SIC with the vertical motion of southern China. When the sea ice decreases, there is descending motion in southern China, which may be generated from the divergent motion. Both the descending motion and divergent water vapor flux are unfavorable for the increase of EPSC. These findings are consistent with previous studies, which found that an anomalous AO can enhance the connection between the warm SST in the North Pacific and the cyclone in the northwest Pacific, thus affecting the summer climate in East Asia [53,54]. Thus, the preceding 5-year SIC in the East Siberian Sea is chosen to be the other predictor for EPSC based on its significant effects on EPSC and its period of 10 years.

3.3. Decadal Prediction Model

The two predictors mentioned above, including the DI_SST and DI_SIC, have significant influences on the predictand (DI_EPSC) with correlation coefficients of 0.79 and −0.51, respectively (Figure 7). The two predictors are generally independent of each other with the correlation coefficient of −0.2, which is not statistically significant. The multiple linear regression method is therefore applied to build the statistical model for predicting the DI_EPSC based on the DI_SST and DI_SIC.

$$DI\text{\_}EPSC=0.71\times DI\text{\_}SST-0.36\times DI\text{\_}SIC$$

(7)

The main factor affecting summer EPSC is the summer SST in the North Atlantic, which explains 50% of the variance of the DI_EPSC. The second factor is the spring SIC in the East Siberian Sea, which explains 13% of the variance of the DI_EPSC.

Cross-validation and independent hindcast are applied to measure the prediction skill of the statistical model. The results of cross-validation (Figure 8a) show that the predicted DI_EPSC is significantly related to the observed DI_EPSC, with a correlation coefficient of 0.83 and an MSSS of 0.69. The phase of the predicted DI_EPSC is consistent with that of the observed DI_EPSC in most years, apart from the time period 1976–1982. The predicted DI_EPSC effectively captures the variability of the observed DI_EPSC, especially in recent years. The final predicted EPSC is obtained by adding the predicted DI_EPSC to the observed EPSC 3 years ago (Figure 8c). The final predicted EPSC also shows a high consistency with the observed EPSC, with a correlation coefficient of 0.86 and an MSSS of 0.72. The phase of the predicted EPSC is consistent with that of the observed EPSC in most years, although the extreme values in the 1990s are not captured well by this model. Noteworthily, the final predicted EPSC captures the shift in the early 1990s. The other decadal shifts with small fluctuations are also well-captured by this statistical model.

From the independent hindcast results (Figure 8b), the predicted DI_EPSC is also consistent with the observed DI_EPSC, with a significant correlation coefficient of 0.9 and an MSSS of 0.81. The phase between the predicted DI_EPSC and the observed DI_EPSC is generally consistent in most years except 2010. The predicted DI_EPSC can also effectively captures the variability and amplitude of the observed DI_EPSC. Additionally, the extreme value of the DI_EPSC in 2016 is captured by this model. The shift during 2014–2015 is also captured by this model. The final predicted EPSC also shows high consistency with the observed EPSC, with a correlation coefficient of 0.78 and an MSSS of 0.52 (Figure 8d). The phase of the predicted EPSC is generally consistent with the phase of the observed EPSC in most years apart from 2012 and 2015. The variability and amplitude between the predicted EPSC and the observed EPSC are also generally consistent in most years, although there are small deviations, which shows generally high prediction skills.

3.4. Real-Time Prediction

The results of the cross-validation and independent hindcast show that this statistical model has reasonable skills in predicting the EPSC, including the shift, phase, and amplitude. Thus, in this section, we further try to make a real-time prediction of EPSC.

Figure 7 indicates that there are remarkable positive anomalies in DI_SST and slight negative anomalies in DI_SIC during 2020–2022. Since the percentage of the explained variance from DI_SST (50%) is higher than that from the DI_SIC (13%) and the anomalous amplitude is also higher in DI_SST than that in DI_SIC, DI_EPSC would increase during 2020–2022 (Figure 8b). As shown in Figure 8d, the real-time predicted EPSC is in a positive phase during 2020–2022, indicating that EPSC would increase during 2020–2022, with a remarkable increase in 2022. Figure 9 shows the observed recent 5-years running mean of the summer extreme precipitation over Eastern China, there are remarkable positive anomalies over Southern China. Noteworthily, the historic Dragon-Boat extreme precipitation event occurred in southern China in May–June 2022 [55], which is consistent with the real-time prediction results. Thus, this statistical model can therefore predict the real-time results 3 years in advance and shows confidence in the future prediction of EPSC.

4. Discussion

Previous studies show that PDO and AMO have significant effects on the decadal summer rainfall over East China [56], which may provide useful information for the prediction of EPSC. We also attempted to use PDO and AMO as predictors. However, the correlations between the PDO/AMO with EPSC are not statistically significant at the 95% confidence level, especially in the decadal increment form, no matter the simultaneous or leading correlations. The results are also true in DI form (

Table 1. The correlation coefficients between the EPSC and the leading 3–7 years summer PDO/AMO. The value in parentheses indicates the correlation coefficient in decadal increment (DI) form.

	Leading 3 Years	Leading 4 Years	Leading 5 Years	Leading 6 Years	Leading 7 Years
PDO	0.23 (0.39)	0.21 (0.24)	0.16 (−0.06)	0.15 (−0.26)	0.19 (−0.39)
AMO	0.30 (−0.33)	0.30 (−0.29)	0.31 (−0.22)	0.33 (−0.09)	0.36 (−0.05)

). The above results will not change when using other seasonal mean PDO/AMO predictors. One possible reason is that the extreme precipitation is discussed in this study compared to the normal precipitation in previous research. The other reason is that the DI_SST/DI_SIC suggested in this study may play the role of bridge for the influence of PDO/AMO on EPSC. Therefore, these two predictors contribute more significantly to the EPSC and predict the EPSC better than PDO and AMO.

5. Conclusions

This paper analyzed the summer EPSC and explored the main influencing factors. The summer extreme precipitation over southern China has been increasing over recent decades, with a significant shift in the early 1990s. The summer SST of the North Atlantic and the spring SIC in the East Siberian Sea are indicated to affect the EPSC. Both predictors have a quasi-period of 10–14 years, which can provide useful prediction information for summer EPSC leading 5–7 years. Therefore, these two predictors were chosen to establish the prediction model based on the decadal increment method.

Both the cross-validation and independent hindcast results illustrate that the decadal increment method can predict the EPSC well, especially the shift in the early 1990s. Additionally, the real-time prediction results show that the EPSC would increase during 2020–2022, which is consistent with the actual situation of more EPSC. These prediction results will help with future local disaster prevention and mitigation.

The decadal increment method not only adds leading observational data to obtain useful decadal information, but also adds valid samples in the correlation analysis, which predicts the summer EPSC well. This statistical approach provides new ideas for the current decadal climate prediction, especially decadal changes. In the future, the statistical-dynamical model or machine learning combined with the increment method can also be attempted to improve decadal predictions.