1.Met Office Hadley Centre, Met Office, FitzRoy Road, Exeter, EX 1 3PB, UK 2.NCAS-Climate and Department of Meteorology, University of Reading, Reading, RG 6 6BB, UK 3.College of Engineering, Mathematics and Physical Sciences, Exeter University, Exeter, EX 4 4QJ, UK 4.Laboratory for Climate Studies, National Climate Center, China Meteorological Administration, Beijing 100081, China Manuscript received: 2018-05-01 Manuscript revised: 2018-09-04 Manuscript accepted: 2018-10-11 Abstract:Predicting monsoon onset is crucial for agriculture and socioeconomic planning in countries where millions rely on the timely arrival of monsoon rains for their livelihoods. In this study we demonstrate useful skill in predicting year-to-year variations in South China Sea summer monsoon onset at up to a three-month lead time using the GloSea5 seasonal forecasting system. The main source of predictability comes from skillful prediction of Pacific sea surface temperatures associated with El NiÑo and La NiÑa. The South China Sea summer monsoon onset is a known indicator of the broadscale seasonal transition that represents the first stage of the onset of the Asian summer monsoon as a whole. Subsequent development of rainfall across East Asia is influenced by subseasonal variability and synoptic events that reduce predictability, but interannual variability in the broadscale monsoon onset for East Asian summer monsoon still provides potentially useful information for users about possible delays or early occurrence of the onset of rainfall over East Asia. Keywords: SCSSM, South China Sea summer monsoon, EASM, East Asian summer monsoon 摘要:季风爆发的预测对农业和社会经济的规划有着重要作用,数以百万计人们的生活都依赖于季风雨季的到来。本文利用GloSea5季节预测模型研究了南海季风爆发年际变化提前三个月的可预测性。这种可预测性的主要来源是与ENSO有关的太平洋海温的预测技巧。南海夏季风爆发是大范围季节转型的标志,代表了整个亚洲夏季风爆发的第一阶段。尽管随后东亚地区降水由于受到次季节尺度和天气尺度事件的影响而造成可预测性降低,但对东亚夏季风建立年际变化的预测仍然能够为用户提供季风雨季到来或早或晚这种有用的信息。 关键词:SCSSM, 南海夏季风, EASM, 东亚夏季风
HTML
--> --> --> -->
3.1. Prediction skill of SCSSM onset using the (Wang et al., 2004) criterion
Figure 1 shows the SCSSM onset pentads identified using the (Wang et al., 2004) criterion for each forecast member with start dates of 17 and 25 March, and 1 and 9 April, in each year, with the ensemble-mean pentad and that identified in the reanalyses. The average interannual standard deviation of onset dates from individual ensemble members is 2.2 pentads, which compares reasonably well with that of the reanalyses (2.6 pentads), and there is a statistically significant (at the 0.75% level, for a one-tailed t-test) correlation of 0.5 between the interannual variations of the ensemble-mean dates and those from the reanalyses, indicating significant predictability. The hindcasts also predict the mean onset pentad to match that of the reanalyses, i.e., pentad 28 (16-20 May). Figure1. Predictability of the SCSSM wind onset: onset pentads derived using the method proposed by (Wang et al., 2004) from the GloSea5 ensemble predictions initialized on 17 and 25 March, and 1 and 9 April (green dots represent individual members of the 52-member ensemble, with the size of the dot scaled by the number of members predicting the same onset pentad), and their ensemble mean (green line), compared with the equivalent onset pentads derived from ERA-Interim (black line). The yellow line shows the Ni?o3.4 SST anomaly in March for each year taken from the HadISST1.1 dataset. Pearson correlation coefficients are given in the legend: r (ens, obs) represents the correlation between the GloSea5 ensemble mean and ERA-Interim; r (ens, sst) represents the correlation between the GloSea5 ensemble mean SCS onset pentads and the observed March Ni?o3.4 SST anomaly; r (obs, sst) represents the correlation between the ERA-Interim SCS onset pentads and the observed March Ni?o3.4 SST anomaly.
(Luo and Lin, 2017) suggested that a more objective measure of the SCSSM onset can be determined using a daily cumulative U SCS and specifying the onset as where this time series changes from decreasing to increasing (indicating that the flow is becoming predominantly westerly). (Wang et al., 2004) also checked their SCSSM onset dates against a cumulative U SCS criterion, DU, which compares the accumulated U SCS in the three days prior to and after the onset. They showed that, although their onset criteria do not explicitly require an abrupt change in westerly speed across the onset pentad, the resultant onset pentads were coincident with such a change. We found that including the additional criterion of DU >7 m s-1 makes very little difference to our results (not shown). We carried out the same analysis for four start dates (1, 9, 17 and 25) in January, February and March taken from the standard operational hindcast ensemble of seven members per start date, and also for a 56-member combined ensemble using start dates of 25 March, and 1, 9 and 17 April (see Table 1). The correlation coefficient increases with decreasing lead time, becoming statistically significant at the 1.5% level (for a one-tailed t-test) from February start dates onwards. Thus, there is significant skill in the SCSSM onset prediction using the (Wang et al., 2004) index at a near three-month lead time over this hindcast period.
2 3.2. Predictability of SCSSM onset using the (Gao et al., 2001) criterion -->
3.2. Predictability of SCSSM onset using the (Gao et al., 2001) criterion
Figure 2 shows the SCSSM dates identified using the (Gao et al., 2001) criterion in each year by each of the 52 ensemble members with start dates of 17 and 25 March, and 1 and 9 April, with the ensemble-mean pentad and that identified in the reanalyses. In contrast with the findings using the (Wang et al., 2004)U SCS index, we find low skill in onset prediction using the (Gao et al., 2001) index at a greater than one-month lead time. Table 1 shows that the correlation increases slightly if the lead time is reduced to around one month, but remains barely statistically significant at the 6% level (using a one-tailed t-test). Figure2. As in Fig. 1 but for SCSSM thermodynamic onset as determined by a sustained increased of θ e above 340 K accompanied by the establishment of westerly winds over the region (10°-20°N, 110°-120°E), as proposed by (Gao et al., 2001) [with the threshold modified by (Ding and He, 2006)].
The difference in prediction skill between the two methods of determining SCSSM onset may be in part related to the region used for the (Gao et al., 2001) index; (Wang et al., 2004) commented that: "… the northern SCS is open to the invasion of a cold front from the north. The westerly flow occurring before the onset is located north of the subtropical ridge and is not of tropical origin". They stated, therefore, that the northern part of the SCS should be excluded when defining the tropical monsoon burst over the SCS. (He et al., 2017) also commented on the influence of northern cold air entering this region of the SCS contributing to ambiguous or intermittent onset. They highlighted the case of 2009, when the strong westerly flow established in mid April was interrupted by easterlies propagating from the northern SCS for several days in early May. Other examples of years where this occurred were given in He et al. (2017, Figs. 1 and 2), and included 2007, 2009 and 2011. Additionally, although (He et al., 2017) did not identify 2004 as an intermittent onset year, the U850 averaged over the (Gao et al., 2001) SCS box fluctuates between easterly and westerly during May, making the onset ambiguous when the (Gao et al., 2001) index is used. He et al. (2017, Fig. 1) showed that this is related to the variability of the winds in the northern part of the SCS. In contrast, the U850 winds over the southern part of the SCS [as covered by the (Wang et al., 2004) box] do not fluctuate to the same extent. (Chan et al., 2000) showed that, in 1998, incursion of cold air into the northern SCS promoted the release of convective available potential energy, which helped to trigger the onset earlier than may have been expected given the ENSO conditions. (Liu et al., 2002) further linked the cold-air incursion to a Rossby wave train triggered over the Bay of Bengal. The additional influence of variability from the subtropics in the northern SCS, which, like ISO, is unpredictable on seasonal time scales, is likely to be a contributing factor in the reduced seasonal prediction skill for SCSSM onset using the (Gao et al., 2001) criteria. In recognition of this, forecasters at CMA release their SCSSM onset forecasts using the (Gao et al., 2001) criteria only on the extended range (11-30 days) timescale (D. Zhang, personal communication, 30 March 2018), on which models have been shown in previous work to possess skill in predicting intraseasonal variability (e.g., Lee et al., 2015; Lim et al., 2018).
2 3.3. Drivers of SCSSM onset predictability using the (Wang et al., 2004) index -->
3.3. Drivers of SCSSM onset predictability using the (Wang et al., 2004) index
Several studies have shown that ENSO is one of the main drivers of large-scale interannual variability in the Asian monsoon region (e.g., Zhou and Chan, 2007; Luo et al., 2016). Westerly (easterly) equatorial wind anomalies associated with El Ni?o (La Ni?a) and a weaker (stronger) Walker circulation are typically associated with negative (positive) SST anomalies over the SCS and a delayed (advanced) seasonal transition (He et al., 2017). This relationship is not symmetrical, however: (He et al., 2017) suggested that both ISOs and changes in west-east thermal contrasts across the Indian Ocean and western Pacific can influence the timing of onset in La Ni?a years. (Hardiman et al., 2018) found a similar asymmetry in the relationship between seasonal mean Yangtze River rainfall and ENSO in observations and hindcasts. We also show on Fig. 1 the observed March Ni?o3.4 SST anomaly time series from HadISST1.1 (yellow line). The correlation coefficient between the ensemble-mean SCSSM onset pentad time series derived using the (Wang et al., 2004) index and the Ni?o3.4 SST time series is 0.9, indicating that the predictable component of the hindcast SCSSM onset is driven mainly by ENSO, which itself is highly predictable on this time scale in GloSea5 (Scaife et al., 2014; MacLachlan et al., 2015). The correlation between observed estimates of SCSSM onset and the observed March Ni?o3.4 SST is rather lower (0.41), indicating the influence of other drivers of SCSSM onset variability that may not be predictable, particularly the ISO (e.g. Shao et al., 2015; Wang et al., 2018), which is itself subject to interannual variations relating to large-scale modes, such as the Pacific-Japan teleconnection (Li et al., 2014). The skill of the ensemble (0.5) is therefore marginally higher than using predicted ENSO conditions alone to predict monsoon onset, though both are skillful. Figure 3a provides additional insight by showing the correlation between the ensemble-mean SCSSM onset dates for the 23 years from the hindcast and observed global monthly mean SSTs in March over the same period. This illustrates that the predictable part of the SCSSM onset from the hindcast is strongly correlated with an ENSO-like pattern of Pacific SSTs, consistent with the findings of (Zhu and Li, 2017). There is also a strong positive correlation with SSTs in the equatorial Indian Ocean, again indicating that warmer SSTs are associated with later SCSSM onset dates. For the observed onset dates derived from ERA-Interim (Fig. 3b), the correlations with SST are far smaller, due to the presence of additional factors in the observations that are not predicted by the ensemble mean. The average correlations between the SSTs and 1000 pseudo-time-series of SCSSM onset created by randomly choosing an individual ensemble member hindcast for each year (Fig. 3c) are naturally smaller than with the ensemble-mean time series, but not as low as those in observations (Fig. 3b). This suggests that some of the subseasonal variations (e.g., ISOs) that affect SCSSM onset in reality may not be sufficiently well represented by the model to capture such influences, even at the relatively high horizontal resolution used by GloSea5 (N216; about 60 km at 50°N). This is consistent with the findings of (Fang et al., 2016), who showed that while several aspects of the boreal summer ISO were improved in the Met Office Unified Model at this resolution, difficulty remained in realistic representation of the variance and propagation characteristics. Figure3. Correlation coefficients between the SCSSM onset pentad derived using the (Wang et al., 2004) index and observed March average SSTs from HadISST1.1 for the period 1993-2015, using: (a) ensemble-mean onset dates from the hindcast; (b) onset dates from ERA-Interim; (c) 10 000 pseudo-time-series of onset dates created by randomly selecting an individual ensemble member from each year (panel shows average over all correlations). Contours and darker shades indicate correlations significant at the 1% (r=0.48) and 3% (r=0.40) levels respectively, for a one-tailed t-test.
2 3.4. Robustness of SCSSM wind onset predictability to ensemble size -->
3.4. Robustness of SCSSM wind onset predictability to ensemble size
To assess the influence of ensemble size on the prediction skill using the (Wang et al., 2004) index, we randomly sample small ensembles of between 1 and 51 members from the 52 members in our combined ensembles with start dates between 17 March and 9 April, and recalculate the correlation between the ensemble-mean time series and that from the observations for different numbers of ensemble members. Figure 4 indicates that, for this measure of monsoon onset, the prediction skill (black line) rises quickly with ensemble size, reaching a mean value of 0.5 for a 28-member ensemble (which is the size of the standard operational hindcast set), and is robust (correlation coefficients averaged over all ensemble-mean time series are statistically significant at the 1% level for a one-tailed test) for around 10 ensemble members or more. This is a reflection of the strong and predictable influence of ENSO on wider tropical rainfall (Kumar et al., 2013; Scaife et al., 2017) and, here, on the SCSSM onset dates in the hindcast: in most of the summers following strong El Ni?o/La Ni?a years (e.g., 1998, 1999, 2000, 2001, 2005, 2008, 2010), the spread among ensemble members is small and several members identify the same onset pentad (see Fig. 1), thereby constraining the values selected by random sampling of the ensemble for those years. Figure4. Effect of ensemble size on the skill of SCSSM onset predictions using the (Wang et al., 2004) index (solid line), denoted r (ens, obs), and the signal-to-noise ratio (correlation of ensemble-mean time series with a pseudo-time-series created by randomly selecting a single model ensemble member for each year; dashed line), denoted r (ens, mod). In both cases, for each choice of ensemble size, up to 10 000 ensemble-mean time series are generated by randomly selecting the chosen number of ensemble member onset dates (independently and without replacement) from the 52 onset dates diagnosed in each year in the combined ensemble and averaging over the chosen number of ensemble members. Dot-dashed lines indicate the values of r that are significant at the 1% and 0.1% levels for a one-tailed t-test.
Several authors (e.g., Eade et al., 2014; Scaife et al., 2014; Dunstone et al., 2016) have demonstrated that the model's North Atlantic Oscillation is less predictable than that observed, so that a large number of ensemble members is required for good prediction skill. This was confirmed by repeatedly randomly selecting a single member to be the truth and using the ensemble mean of the remaining members to predict that member. In contrast, the dashed line in Fig. 4 indicates that the model's SCSSM onset dates are more predictable than those from reanalyses, i.e., that the model is over-confident in its predictions, as is often found for tropical rainfall (Weisheimer and Palmer, 2014). This again illustrates the dominant role of ENSO in providing the predictability in the model, while the observed onset dates are also influenced by intraseasonal variations that are unpredictable on the seasonal time scale.