
Previous studies have shown that the TC activity depends mainly on the conditions of the largescale atmosphere–ocean interaction associated with regional changes in SST anomalies, vertical wind shear (WS), lowlevel relative vorticity, and extensions to the monsoon trough and western Pacific subtropical high (Chia and Ropelewski, 2002; Emanuel, 2003; Hendricks et al., 2013; Huang and Chan, 2014; Lok and Chan, 2018). In the present study, seven factors related to LTCs in China are identified, which can be divided into two groups: three steering factors [sea level pressure (SLP), 500hPa geopotential height (H500), and zonal wind at 850 hPa (U850)] and four genesis factors [850hPa relative vorticity (Vor), 200hPa divergence (Div), 200–850hPa vertical zonal WS, and SST].
To select a potential predictor in the forecast model, the prediction skill for summer environmental factors related to the number of LTCs using CFSv2 for the four initial months from February to May are evaluated in this section. Figure 2 shows the TCCs between the DY of the JJA number of LTCs and the observed and CFSv2predicted largescale summer environmental circulations in the four initial months from February to May during 1983–2017. The boxes in Fig. 2 represent the definition areas of seven potential prediction factors. The key areas defined by the seven potential predictors are: P1–SLP (10°–25°N, 110°–150°E), P2–H500 (10°–25°N, 100°–140°E), P3–U850 (0°–15°N, 100°–140°E), P4–Vor (10°–25°N, 100°–140°E), P5–Div (0°–20°N, 120°–140°E), P6–WS (10°S–10°N, 100°–130°E), and P7–SST (0°–20°N, 80°–100°E). The lowpressure anomaly in China’s coastal area (Figs. 2a1, a2) and the easterly anomaly of the Maritime Continent (Fig. 2a3) are favorable for LTCs in China. During the years with more LTCs in China, they are accompanied by the positive anomalous lowlevel relative vorticity in the South China Sea (Fig. 2a4), more anticyclonic highlevel divergence, relatively strong WS in the Maritime Continent (Figs. 2a5, a6), and a cold SST anomaly (Fig. 2a7). For the 3month lead (February), CFSv2 has low predictive skill in terms of the relationship between the DY of the JJA number of LTCs and the CFSv2predicted predictors, except the divergence at 200 hPa (Figs. 2b1–b7). The TCC between the DY of the number of LTCs and P5–Div is 0.31, exceeding the 90% confidence level (Table 1). CFSv2 is able to make some skillful predictions with respect to the relationships between the number of LTCs and the predictors of P3–U850, P4–Vor, P5–Div, P6–WS, and P7–SST from March. For the 2 and 1month leads (April and May), CFSv2 shows high skill in the relationships between the number of LTCs and the CFSv2predicted potential predictors.
Figure 2. TCCs between the DY of the JJA number of LTCs and the observed (a1–a7) and CFSv2predicted JJA largescale environmental conditions in the four initial months from February to May (b1–e7) during 1983–2017: sea level pressure (SLP; a1, b1, c1, d1, and e1), geopotential height at 500 hPa (H500; a2, b2, c2, d2, and e2), zonal wind at 850 hPa (U850; a3, b3, c3, d3, and e3), relative vorticity at 850 hPa (Vor; a4, b4, c4, d4, and e4), divergence at 200 hPa (Div; a5, b5, c5, d5, and e5), zonal wind shear at 200–850 hPa (WS; a6, b6, c6, d6, and e6), and sea surface temperature (SST; a7, b7, c7, d7, and e7). Dotted areas indicate the statistical significance at the 95% confidence level, as estimated by a local Student’s ttest.
TCC–NLTC ACC–P TCC–P Feb Mar Apr May Feb Mar Apr May Feb Mar Apr May P1–SLP −0.05 −0.23 −0.36^{**} −0.41^{**} 0.25^{***} 0.22^{**} 0.36^{**} 0.43^{***} 0.34^{**} 0.38^{**} 0.56^{***} 0.72^{***} P2–H500 −0.07 −0.19 −0.28^{*} −0.38^{**} 0.12 0.22^{**} 0.19^{*} 0.29^{*} 0.58^{***} 0.66^{***} 0.70^{***} 0.81^{***} P3–U850 0.23 0.35^{**} 0.35^{**} 0.38^{**} 0.24^{***} 0.23^{***} 0.31^{***} 0.37^{***} 0.70^{***} 0.69^{***} 0.73^{***} 0.70^{***} P4–Vor 0.25 0.38^{**} 0.39^{**} 0.41^{**} 0.13 0.11 0.16^{*} 0.23^{**} 0.62^{***} 0.60^{***} 0.66^{***} 0.66^{***} P5–Div 0.31^{*} 0.47^{***} 0.39^{**} 0.47^{***} 0.06 0.00 0.09 0.09 0.62^{***} 0.55^{***} 0.55^{***} 0.60^{***} P6–WS 0.23 0.32^{*} 0.36^{**} 0.47^{***} 0.30^{***} 0.33^{***} 0.36^{***} 0.40^{***} 0.54^{***} 0.52^{***} 0.62^{***} 0.65^{***} P7–SST −0.27 −0.37^{**} −0.36^{**} −0.46^{***} 0.07 0.03 0.03 0.17^{*} 0.68^{***} 0.66^{***} 0.65^{***} 0.74^{***} Note: TCC–NLTC refers to the TCC between the DY of the number of LTCs and the potential single predictor. ACC–P represents the spatial ACC between the observed and predicted JJA largescale atmospheric circulations in key areas. TCC–P means the TCC between the observed and predicted predictor. Triple, double, and single asterisks indicate the statistical significance at the 99%, 95%, and 90% confidence levels, respectively. Table 1. The prediction skill for environmental factors of the JJA number of LTCs during 1982–2017 using CFSv2predicted JJA predictors for the four initial months from February to May
The forecasting skill of CFSv2 in the four initial months is also assessed for the largescale environmental circulations associated with the JJA number of LTCs. Figure 3 shows TCCs between the observed and CFSv2predicted JJA environmental factors at the 0 to 4month leads associated with the number of LTCs during 1983–2017. Compared to the relationships between the DY of the number of LTCs and the predictors derived from CFSv2 in different initial months, CFSv2 shows high forecasting skill for the interannual variation of the environmental factor itself. The range of TCCs between the observed and predicted seven predictors is 0.34–0.70 at the 4month lead, 0.38–0.69 at the 3month lead, 0.56–0.73 at the 2month lead, and 0.60–0.81 at the 1month lead (Table 1). Based on the spatial anomaly correlation coefficients (ACCs) in key areas defined by the potential predictors, CFSv2 shows high predictive ability for P1–SLP, P3–U850, and P6–WS in February, with ACCs of 0.25, 0.24, and 0.30, all exceeding the 99% confidence level. For the 3 and 2month leads, ACCs are significant for P1–SLP, P2–H500, P3–U850, and P6–WS. For all potential predictors apart from P5–Div, CFSv2 shows high skill in the spatial pattern at the 1month lead.
Figure 3. TCCs between the observed and CFSv2predicted JJA largescale environmental conditions in the four initial months from February to May during 1983–2017: SLP (a1, b1, c1, and d1), H500 (a2, b2, c2, and d2), U850 (a3, b3, c3, and d3), Vor (a4, b4, c4, and d4), Div (a5, b5, c5, and d5), WS (a6, b6, c6, and d6), and SST (a7, b7, c7, and d7). Dotted areas indicate the statistical significance at the 95% confidence level, as estimated by a local Student’s ttest.

Although CFSv2 possesses the forecasting skill for the atmospheric circulation accompanying LTCs to some degree, the observed climate systems in the preceding seasons also play a vital role in predicting the number of LTCs. Previous studies have indicated that the influence of Southern Hemisphere circulation variability on largescale environments and tropical convection in the subtropical Northern Hemisphere show a possible use of Antarctic Oscillation variation for the longrange forecasting of WNP TC activity (Ho et al., 2005).
In this study, a negative relationship is found between the number of LTCs and the preceding summer in the anomaly and DYform of SST in Southwest Indonesia (Fig. 4). The predictor x_{1} is defined as the regionally averaged DY of precedingsummer SST in Southwest Indonesia (14°S–0°, 90°–120°E). The TCC between the DY of the number of LTCs and the first predictor of the statistical prediction model is −0.34, exceeding the 95% confidence level estimated by a local Student’s ttest. TCCs between the predictor x_{1} and JJA SST are shown in Fig. 5a. It is found that the interannual JJA SST variability in Southwest Indonesia coincides with the interannual SST variability in the East Maritime Continent (EMC) and East Australia (EAU). The TCCs between the predictor x_{1} and the areaaveraged SST in EMC (10°S–20°N, 130°–160°E) and EAU (40°–10°S, 150°E–150°W) are 0.64 and 0.68 respectively, exceeding the 99% confidence level. There is also a close relationship between the SST in EMC and EAU, with a TCC of 0.82. The SST signal in Southwest Indonesia can persist from summer to the following winter (Figs. 5a–c). The SST signal in EMC and EAU can persist for one year, subsequently affecting the number of LTCs in the following summer. But how does the JJA SST affect the number of LTCs in the following JJA? To answer this question, an EMC–EAU SST index (SSTI) is defined as the regionally averaged SST in EMC and EAU. The JJA EMC–EAU SST is not only closely related to the predictor x_{1}, but also can last for one year. The JJA SST signal can be stored in EMC–EAU for one year, subsequently affecting the interannual variability of the number of LTCs in the following summer (Fig. 5). The teleconnection pattern from the mid–high latitudes of the Southern Hemisphere to WNP plays a key role in the linkage between the Antarctic Oscillation and TCs activity in WNP (Fan, 2007; Wang and Fan, 2007). Figure 6 shows the regression coefficients of the JJA SLP, relative vorticity at 850 hPa, geopotential height at 500 hPa, and horizontal wind at 850 hPa on the EMC–EAU SSTI in the preceding JJA. The SST in EMC–EAU increases in summer, which may lead to a negative SLP anomaly, lowlevel convergence over Indonesia, and divergence over the Philippines (Figs. 6a, b). The anomalous convection over the Maritime Continent may result in a strengthened WNP subtropical high in summer (Fig. 6c), contemporaneous with anomalous lowlevel southwesterly winds prevailing over eastern China (Fig. 6d). These conditions are unfavorable for LTCs over China.
Figure 4. TCCs between the JJA number of LTCs and preceding summer SST during 1983–2017 in (a) anomaly and (b) DYform for the observation. The red box is the definition region (14°S–0°, 90°–120°E) of predictor x_{1}. Dotted areas indicate the statistical significance at the 95% confidence level, as estimated by a local Student’s ttest.
Figure 5. TCCs between the predictor x_{1} and SST in the following seasons during 1983–2017: (a) JJA, (b) SON, (c) DJF, (d) MAM, and (e) the following JJA. Blue boxes are the definition regions of the predictor x_{1} (14°S–0°, 90°–120°E) and EMC–EAU (10°S–20°N, 130°–160°E; 40°–10°S, 150°E–150°W). Dotted areas indicate the statistical significance at the 95% confidence level, as estimated by a local Student’s ttest.
Figure 6. Regression coefficients of the (a) SLP, (b) Vor, (c) H500, and (d) horizontal wind at 850 hPa (UV850) on the EMC–EAU SSTI in JJA during 1983–2017. Dotted areas indicate the statistical significance at the 95% confidence level, as estimated by a local Student’s ttest.
Based on a statistical forecast model, a previous study proved that the preceding October SLP in southeastern Australia is an important factor for the number of LTCs (Fan, 2009). In this study, the preceding October SLP in South Australia is also found to be an important predictor in the prediction model for the number of LTCs. TCCs between the number of LTCs and the preceding October SLP during 1982–2016 in anomaly and DYform for the observation are shown in Fig. 7. Compared to the variables in anomalyform, the negative correlation between the number of LTCs and South Australia SLP increases significantly in DYform. The predictor x_{2} is defined as the areaaveraged DY of SLP in the preceding October in South Australia (65°–45°S, 90°–160°E). The TCC between the number of LTCs and the second predictor is −0.48, exceeding the 99% confidence level. But how can the preceding October SLP in South Australia affect the number of LTCs in China? Previous studies have pointed out that the atmospheric circulations in the Southern Hemisphere are closely related to the TC activity over WNP (Ho et al., 2005; Sun et al., 2007; Wang and Fan, 2007). The boreal spring SST anomaly east of Australia may lead to simultaneous change in the tropical atmospheric circulation via the teleconnection wave train, and then affect the WNP TC activity (Zhou and Cui, 2011). Figure 8a shows the regressions of JJA vertical velocity in DYform along 90°–150°E upon the JJA number of LTCs. When the summer number of LTCs is frequent, strong ascending motion appears over approximately 10°–25°N, which is conducive to TC genesis and landfall in China. As shown in Fig. 8b, the preceding October negative SLP anomaly in South Australia may weaken the ascending motion in WNP, which is detrimental to LTCs. The influence of Southern Hemisphere circulation variability on the largescale environments and tropical convection in the subtropical Northern Hemisphere may be carried out by the teleconnection wave train.
Figure 7. TCCs between the JJA number of LTCs and preceding October SLP during 1982–2016 in (a) anomaly and (b) DYform for the observation. The red box is the definition region (65°–45°S, 90°–160°E) of predictor x_{2}. Dotted areas indicate the statistical significance at the 95% confidence level, as estimated by a local Student’s ttest.
Figure 8. Regressions of the JJA vertical velocity (10^{−3} Pa s^{−1}) in DYform along 90°–150°E upon the JJA number of (a) LTCs and (b) predictor x_{2} multiplied by −1.0. Dotted areas indicate the statistical significance at the 95% confidence level, as estimated by a local Student’s ttest.
The preceding DJF SST in the Sea of Japan is also found to be a potential predictor for the number of LTCs in summer. The predictor x_{3} is defined as the areaaveraged DY of DJF SST in the Sea of Japan (34°–44°N, 130°–150°E). Figure 9 shows TCCs between the JJA number of LTCs and the preceding winter SST during 1983–2017 in anomaly and DYform for the observation. The significant negative correlation between the number of JJA LTCs and the preceding winter SST exists only in DYform. Previous studies have examined the impact of ENSO on LTCs along the coast of China (Wang and Chan, 2002; Liu and Chan, 2003; Wu et al., 2004; Fudeyasu et al., 2006). Seasonal variations of the LTC activity along the South China coast are closely related to the largescale circulation accompanied with ENSO (Liu and Chan, 2003). The largescale circulation anomalies associated with El Niño events result in the eastward shift in the mean TC genesis position, which will reduce the number of LTCs (Wu and Lau, 1992; Wang and Chan, 2002; Wu et al., 2004; Chen et al., 2006). The TC genesis location shifts eastward, which is not conducive to LTCs in China. There is a significant negative correlation between the predictor x_{3} and number of LTCs, with a TCC of −0.52. The TCC between the DY of the number of LTCs and DJF Niño4 index is −0.33, exceeding the 95% confidence level estimated by a local Student’s ttest. Because the explained variance of the winter SST in the Sea of Japan is larger than that of the DJF Niño4 index, the winter SST in the Sea of Japan is selected as a potential predictor in the prediction model. It is found that the significant relationship between the predictor x_{3} and Niño4 index in DYform continues from December to the following April (Fig. 10). The winter SST in the Sea of Japan may affect the number of LTCs via influencing the winter and spring ENSO, which in turn influences the interannual variation of the number of LTCs. As shown in Fig. 11a, the anomalous lowlevel divergence occurs over the Philippine Sea in summer, with anomalous cold DJF SST over the Niño4 region. The suppressed ascending motion in the Maritime Continent may result in an enhancement and westward shift of the WNP subtropical high (Fig. 11b). As shown in Fig. 11c, during La Niña phases, an anomalous anticyclonic cell is situated over the South China Sea and westerlies tend to prevail between 20° and 30°N in WNP. In addition, there is a weak vertical WS in the Maritime Continent, accompanied by the anomalous cold DJF SST over the Niño4 region (Fig. 11d). These conditions are unfavorable for LTCs over China.
Figure 9. TCCs between the JJA number of LTCs and preceding winter SST during 1983–2017 in (a) anomaly and (b) DYform for the observation. The red box is the definition region (34°–44°N, 130°–150°E) of predictor x_{3}. Dotted areas indicate the statistical significance at the 95% confidence level, as estimated by a local Student’s ttest.

Based on the above research, a statistical model using three predictors is built by using the DY method. Crossvalidation testing is used to evaluate the predictive skill of the forecast model (Michaelsen, 1987). The accuracy of the prediction model is quantitatively measured by using the TCC, mean absolute error (MAE), and meansquare skill score (MSSS). MSSS is defined by the following equation:
where MSE refers to the meansquared error, P and O are the prediction and observation data respectively (Murphy, 1988). The MSSS is the most stringent of the three indicators.
The statistical prediction model is represented by Eqs. (2) and (3):
where Δy_{i} represents the DY of LTCs for one current year, y_{i1} refers to the observed LTCs of the previous year, and y_{i} refers to the predicted LTCs for one current year in the statistical prediction model. As shown in Fig. 12a, the predicted DY of the number of LTCs in the crossvalidation tests agrees well with the observation during 1983–2017, with a TCC of 0.73. The predicted number of LTCs can be obtained by adding the predicted DY of the number of LTCs to the observed number of LTCs in the previous year (Fig. 12b). The TCC between them is 0.63, with an MAE of 1.18 and MSSS of −0.38. The years 1984, 1985, and 2014 are the three years during which the observed JJA number of LTCs falls outside the 95% predicted range.
Figure 12. (a) DY of the number of LTCs for the observation (red line) and statistical model (turquoise line) in the crossvalidation during 1983–2017. (b) Predicted (turquoise line) and observed (red line) number of LTCs in the crossvalidation during 1983–2017. The grey shading indicates the 95% prediction interval.

Several studies have pointed out that the prediction ability of the hybrid statistical–dynamical model for TCs activity is not only higher than that of the statistical model, but also higher than that of the dynamical model. (Wang et al., 2009; Kim and Webster, 2010; Vecchi et al., 2011). Based on the relationship between the number of LTCs and the physical significance of predictors, a hybrid statistical–dynamical model is constructed to predict the number of LTCs during the summer at 3 to 0month leads from February to May. The model outputs used in the hybrid dynamical–statistical model are obtained from CFSv2. Predictors in the prediction models are selected based on correlations between the observed number of LTCs and the largescale environmental circulation. Because the seven predictors derived from CFSv2 for the four initial months are not independent of each other, only a single predictor is selected in the hybrid prediction model.
Table 2 shows the forecasting skill using the three statistical predictors and one CFSv2predicted predictor for the four initial months from February to May. For the 4month lead, TCCs for the seven predictors are 0.64 for P1–SLP, 0.61 for P2–H500, 0.71 for P3–U850 and P4–Vor, 0.70 for P5–Div, 0.69 for P6–WS, and 0.66 for P7–SST. Compared to the statistical prediction model, MAE is reduced by 0.4%–14.4%. For February, P4–Vor and P5–Div show the best skill, with an MSSS of 0.13 and 0.12 respectively. For the 3month lead, the range of TCCs for the seven predictors is 0.62–0.73. MAE decreases by 4.2%–25.2%. The predictors P4–Vor and P5–Div also show the best forecasting skill among these predictors. For the 2month lead, the predictors P1–SLP and P4–Vor show higher predictive ability than the other predictors, with an MSSS of 0.12 and 0.09 respectively. In May, both P1–SLP and P5–Div demonstrate considerable predictive skill for the JJA number of LTCs. For the 4, 3, and 2month leads, the predictor P2–H500 shows the worst forecasting skill. In general, P4–Vor is a modest predictor, with the stable and good forecasting skill at all lead times.
P1–SLP P2–H500 P3–U850 P4–Vor P5–Div P6–WS P7–SST TCC Feb 0.64 0.61 0.71 0.71 0.70 0.69 0.66 Mar 0.65 0.62 0.73 0.72 0.70 0.69 0.68 Apr 0.70 0.64 0.69 0.70 0.69 0.70 0.66 May 0.70 0.66 0.66 0.68 0.70 0.69 0.66 MAE Feb 4.5% 0.4% 11.0% 14.4% 13.8% 14.1% 7.8% Mar 11.5% 4.2% 22.1% 25.2% 20.6% 19.1% 17.1% Apr 16.7% 4.8% 13.7% 17.2% 11.6% 14.3% 6.4% May 19.3% 12.2% 8.5% 11.8% 16.3% 13.7% 8.4% MSSS Feb −0.15 −0.31 0.06 0.13 0.12 0.06 −0.09 Mar −0.05 −0.19 0.20 0.24 0.27 0.09 0.09 Apr 0.12 −0.10 0.01 0.09 0.03 0.07 −0.07 May 0.14 0.05 −0.09 0.02 0.16 0.08 −0.05 Table 2. The prediction skill of the hybrid model for the JJA number of LTCs during 1982–2017 using three statistical predictors and CFSv2predicted single predictors for the four initial months from February to May
In addition, because LTCs are one of the most serious natural disasters, the need to well predict the number of LTCs over China is urgent. A summer climate consultation is held by the National Climate Center in March every year. Therefore, a realtime prediction model for the number of LTCs in summer is also constructed, based on the above research (Fig. 13). The realtime prediction model is established based on the three statistical predictors (x_{1}, x_{2}, and x_{3}) and the CFSv2predicted P4–Vor (x_{4}) in JJA from February. The predicted JJA DY of the number of LTCs in the hybrid prediction model is obtained by Eq. (4) below:
Figure 13. DY of the number LTCs for the observation (red line) and hybrid model using three statistical predictors and the CFSv2predicted relative vorticity at 850 hPa in February (turquoise line) in the crossvalidation during 1983–2017. (b) Predicted (turquoise line) and observed (red line) number of LTCs in the crossvalidation during 1983–2017. The grey shading is the 95% prediction interval.
The TCC between the predicted JJA number of LTCs (DY of the number of LTCs) is 0.71 (0.82), exceeding the 99% confidence level. MAE decreases by 14.4%, and MSSS is 0.13. The realtime prediction model shows better prediction skill with respect to the JJA number of LTCs for 2014–17.
TCC–NLTC  ACC–P  TCC–P  
Feb  Mar  Apr  May  Feb  Mar  Apr  May  Feb  Mar  Apr  May  
P1–SLP  −0.05  −0.23  −0.36^{**}  −0.41^{**}  0.25^{***}  0.22^{**}  0.36^{**}  0.43^{***}  0.34^{**}  0.38^{**}  0.56^{***}  0.72^{***}  
P2–H500  −0.07  −0.19  −0.28^{*}  −0.38^{**}  0.12  0.22^{**}  0.19^{*}  0.29^{*}  0.58^{***}  0.66^{***}  0.70^{***}  0.81^{***}  
P3–U850  0.23  0.35^{**}  0.35^{**}  0.38^{**}  0.24^{***}  0.23^{***}  0.31^{***}  0.37^{***}  0.70^{***}  0.69^{***}  0.73^{***}  0.70^{***}  
P4–Vor  0.25  0.38^{**}  0.39^{**}  0.41^{**}  0.13  0.11  0.16^{*}  0.23^{**}  0.62^{***}  0.60^{***}  0.66^{***}  0.66^{***}  
P5–Div  0.31^{*}  0.47^{***}  0.39^{**}  0.47^{***}  0.06  0.00  0.09  0.09  0.62^{***}  0.55^{***}  0.55^{***}  0.60^{***}  
P6–WS  0.23  0.32^{*}  0.36^{**}  0.47^{***}  0.30^{***}  0.33^{***}  0.36^{***}  0.40^{***}  0.54^{***}  0.52^{***}  0.62^{***}  0.65^{***}  
P7–SST  −0.27  −0.37^{**}  −0.36^{**}  −0.46^{***}  0.07  0.03  0.03  0.17^{*}  0.68^{***}  0.66^{***}  0.65^{***}  0.74^{***}  
Note: TCC–NLTC refers to the TCC between the DY of the number of LTCs and the potential single predictor. ACC–P represents the spatial ACC between the observed and predicted JJA largescale atmospheric circulations in key areas. TCC–P means the TCC between the observed and predicted predictor. Triple, double, and single asterisks indicate the statistical significance at the 99%, 95%, and 90% confidence levels, respectively. 