-
As shown in Table 1, the seasonal mean TC frequency and ACE are systematically underestimated in all three basins. As seen in many other studies (e.g., Camargo et al., 2005; Manganello et al., 2012), the intensity of TCs in the low-resolution model is weaker than in observations. The coarse resolution of NUIST-CFS1.0 is one of the main reasons for the lower ACE. The TC frequency and ACE over the WNP predicted by NUIST-CFS1.0 are the largest among all the three basins, which is consistent with observations. The predicted TC frequency during the MJJASON season is similar between the ENP and the NA, whereas five more TCs are observed over the ENP than the NA.
WNP ENP NA TC frequency OBS 22.39 17.66 12.61 CFS 11.33 5.07 5.37 ACE (104 kt2) OBS 187.36 166.15 124.81 CFS 36.52 13.79 16.31 Table 1. Seasonal means of tropical cyclone (TC) frequency and accumulated cyclone energy (ACE) for the MJJASON season of 1982–2019 based on IBTrACS (OBS) and NUIST-CFS1.0 hindcasts.
The densities of cyclogenesis in IBTrACS and NUIST-CFS1.0 hindcasts are shown in Fig. 1. Cyclogenesis location in the model is defined as the point where the vortex with a
$ {V}_{\mathrm{m}\mathrm{a}\mathrm{x}} $ reaching 12.8 m s−1 or above is first identified, while the observed genesis is defined as the first record in IBTrACS. As shown in Fig. 1, the most active and concentrated area of cyclogenesis is located in the ENP in both observations and the hindcasts. The area of intense cyclogenesis in the WNP in the hindcasts extends more eastward. In addition, the model can reproduce the maximum cyclogenesis in the South China Sea well but with a more southward location. The largest differences appear in the NA, where the cyclogenesis in the hindcasts is found to be far away from the west coast of Africa. Moreover, the predicted TC activity decreases over the Gulf of Mexico and the east coast of the United States compared with IBTrACS.Figure 1. Genesis densities as number density per season per unit area equivalent to a 4.5° × 4.5° box based on (a) IBTrACS (OBS) and (b) NUIST-CFS1.0 hindcasts during the MJJASON season of 1982–2019. The red box shows the main development regions (MDRs) in the WNP, ENP, and NA, which are defined as 5°–22.5°N, 110°–180°E in the WNP, 7.5°–15°N, 160°–80°W in the ENP, and 7.5°–22.5°N, 80°–20°W in the NA, respectively. The WNP is further divided into eastern and western parts by the longitude of 140°E.
As shown in Fig. 2, the track densities in the WNP, ENP, and NA in hindcasts are underestimated. The track density areas of the three basins in the model are smaller than those in the observations. In the WNP, there is a local maximum to the east of the Philippines in both the observations and the hindcasts. However, the model fails to reproduce the secondary maximum over the South China Sea. The main area of track density along the east coast of the United States in the NA is also absent in the model. This may be attributed to the underestimated TC genesis frequency over the western part of the WNP (WNP-W) and the northern part of the NA in the model (Fig. 1).
It is well recognized that a reliable prediction of TC activity depends on a realistic representation of tropical SST and the associated atmospheric conditions (Gray, 1979; Knaff, 1997; DeMaria et al., 2001; Emanuel, 2007). The spatial distributions of climatological mean environmental factors (SST, VWS, low-level VOR, and mid-level RH) by model hindcasts are shown in Fig. 3. The highest mean SST and the lowest mean VWS (Figs. 3a, c) provide favorable conditions for TC genesis in the WNP. Figure 3b displays that the SST biases are noticeable in the tropical Pacific and Atlantic, which is common in the current generation of general circulation models (Wang et al., 2014; Xu et al., 2014a, b; Hsu et al., 2019). Large positive biases are found over the eastern tropical Atlantic and Pacific with magnitudes of up to 2°C. Cold SST biases appear in the northwestern Atlantic and Pacific, with weaker magnitudes. Hsu et al. (2019) proposed that the combination of the Pacific SST cold bias and Pacific SST warm bias would induce an eastward shift of WNP TC genesis. The strong VWS related to the east side of the tropical upper-tropospheric trough in the central North Pacific suppresses TC genesis (Figs. 3c, d; Kelley and Mock 1982; Fitzpatrick et al., 1995; Wu et al., 2015). The negative VWS biases in the eastern part of the WNP (WNP-E) may also contribute to the eastward shift of the mean location of TC genesis (Fig. 1). In the NA, the cold SST biases and positive VWS biases may lead to less TC genesis in the Gulf of Mexico and along the east coast of the United States (Fig. 1; Figs. 3b, d). It is noteworthy that the criterion of cyclogenesis occurring within 0°–30°N over oceans excludes TCs generated north of 30°N near the east coast of the United States. The westward shift of TC genesis in the NA MDR may be related to the warmer climatological mean SST, weaker VWS, and larger low-level VOR around 40°W longitude in the model (Figs. 3b, d, f). A large portion of ENP TCs form from easterly waves coming across the Central American mountainous region (Avila et al., 2003; Franklin et al., 2003), which are always poorly represented in low-resolution models (Martin and Thorncroft, 2015). It is likely that the model’s low resolution contributes to the deficient prediction of TC frequency over the ENP by the NUIST-CFS1.0 (Table 1).
Figure 3. Climatological mean (a) SST (units: °C), (c) VWS (units: m s−1), (e) 850-hPa VOR (units: 10−5 s−1), and (g) 600-hPa RH (units: %) based on NUIST-CFS1.0 hindcasts for the MJJASON season of 1982–2019. The panels in the right column display the differences in the (b) SST, (d) VWS, (f) 850-hPa VOR, and (h) 600-hPa RH between model hindcasts and observations. Only significant differences are presented.
-
NUIST-CFS1.0 performs well in forecasting ACE in the WNP. The correlation coefficient of the seasonal ACE between the observations and the calibrated ensemble means during 1982–2019 is 0.65 (Table 2 and Fig. 4b). However, the performance of NUIST-CFS1.0 in predicting TC frequency over the entire WNP is relatively poor, with the correlation coefficient being only 0.24 (Table 2 and Fig. 4a). In 1992, 2013, and 2014, the observed TC frequency is outside of the 10th–90th percentile range of ensemble hindcasts. However, the hindcasts can still successfully capture some extreme cases, such as those in 1998 and 2010 (Fig. 4a). The RMSE for the calibrated TC frequency is larger than one standard deviation of the observed TC counts (Table 2), indicating that the prediction error for TC frequency is quite large in the WNP. The correlation coefficients of seasonal TC activity (including TC frequency and ACE) between the observations and individual ensemble members of NUIST-CFS1.0 (Table 2) are also calculated. It is found that the correlation coefficients for individual members are generally lower than those of the ensemble means, suggesting that the ensemble forecast approach is helpful in improving the prediction of seasonal TC activity over the WNP.
Figure 4. Hindcasts of MJJASON (left column) TC frequency and (right column) ACE (units: 104 kt2) in the (a, b) WNP, (c, d) WNP-E, and (e, f) WNP-W, respectively. Red lines indicate the observed time series, and black lines display the calibrated ensemble mean. Blue dots denote the calibration from individual ensemble members. Box-and-whisker plots display the 25th–75th and 10th–90th percentile ranges, respectively. Correlation coefficients between the observed time series and ensemble means are shown in each panel. Gray lines indicate the observed climatological means for each basin.
CC_M CC_I RMSE STDV_O SPRvERR TC frequency WNP 0.24 0.13 (±0.11) 4.54 3.71 1.32 WNP-E 0.61 0.45 (±0.06) 3.79 4.34 1.11 WNP-W 0.41 0.22 (±0.19) 3.49 3.47 1.35 ACE (104 kt2) WNP 0.65 0.41 (±0.08) 42.77 55.82 1.33 WNP-E 0.72 0.57 (±0.05) 48.42 59.51 1.20 WNP-W 0.23 0.11 (±0.13) 19.06 19.46 0.90 Table 2. Correlation coefficients (CC_Ms) and root-mean-square errors (RMSEs) of seasonal TC frequency (top row) and ACE (bottom row) between the observations and the calibrated ensemble means during 1982–2019. CC_I denotes the simple average of the correlation coefficients of seasonal TC frequency and ACE between observations and individual members. The standard deviations of the CC_Is are given in parentheses. STDV_O is one standard deviation of the observations. The SPRvERR indicates the ratio of the ensemble spread (averaged over all forecast years) to the RMSE. The values in bold indicate that the correlation coefficients are statistically significant at the 95% confidence level as examined by the one-tailed Student’s t-test.
ACE is controlled by not only TC frequency but also TC intensity and lifetime. TCs in El Niño years tend to be more intense and long-lived than those in La Niña years (Chia and Ropelewski, 2002; Camargo and Sobel, 2005; Chen et al., 2006), as indicated by the large correlation coefficient between ACE and the Niño-3.4 index (Table 3; the correlation coefficient is 0.73 in observations). The hindcasts can reproduce the positive relationship between ACE and the Niño-3.4 index but with the correlation coefficient (0.45) being much lower than that seen in the observations. The NUIST-CFS1.0 seasonal hindcast skill for the Niño-3.4 index is very high with the correlation coefficient being 0.83 (Fig. 5a). Wang and Chan (2002) proposed that the equatorial central and eastern Pacific warming during El Niño years tends to induce pronounced equatorial westerly anomalies in the western Pacific. Large meridional shears associated with the equatorial westerly anomalies increase low-level relative VOR, which would promote TC genesis, especially in the southeastern part of the WNP. The hindcasts capture this relationship well, i.e., the interannual variations of the seasonal ACE in model predictions are positively related to the average low-level relative VOR (Table 3). However, the negative relationship between MDR VWS and ACE is not captured by NUIST-CFS1.0.
WNP WNP-E WNP-W OBS CFS OBS CFS OBS CFS Niño3.4 0.01 0.25 0.45 0.54 −0.55 −0.41 0.73 0.45* 0.74 0.64 −0.17 0.00 EIO SSTA −0.57 −0.25 −0.43 −0.32 −0.08 0.09 −0.17 −0.27 −0.12 −0.28 −0.14 −0.12 MDR SST −0.20 −0.21 −0.40 0.02* 0.37 0.12 −0.36 −0.24 −0.24 0.06 −0.06 −0.21 MDR VWS −0.26 −0.02 −0.50 −0.32 0.12 −0.27* −0.34 0.08* −0.46 −0.26 0.18 0.05 MDR VOR 0.26 0.45 0.56 0.69 −0.05 −0.08 0.78 0.61 0.79 0.76 0.26 0.29 MDR RH 0.00 0.26 −0.05 0.34* 0.63 0.34 −0.31 0.19* 0.05 0.28 0.33 0.20 Table 3. Correlation coefficients between large-scale environmental factors and TC frequency (top row) and ACE (bottom row) in the WNP, WNP-E, and WNP-W during the MJJASON season of 1982–2019. The correlation in the model is the mean value of the correlation in each member. The large-scale environmental indexes include 1) the Niño-3.4 index (5°S−5°N, 120°−170°W), 2) EIO SSTA averaged over 10°S−22.5°N, 75°−100°E (Zhan et al., 2011), 3) MDR SST, 4) MDR VWS, 5) MDR VOR at 850-hPa, and 6) MDR RH at 600-hPa. The values in bold indicate that the correlation coefficients are statistically significant at the 95% confidence level as examined by the one-tailed Student’s t-test. Model-produced correlations that are obviously different from the observations in terms of Fisher’s Z statistic are marked with an asterisk.
The interannual variations of WNP TC frequency are hard to capture in many prediction systems, especially in those low-resolution ones (e.g., Camargo et al., 2005; Camp et al., 2015). Previous studies (Ramage and Hori, 1981; Lander, 1994; Wang and Chan, 2002; Chen et al., 2006) have shown that annual TC frequency over the WNP has no close relationship with the Niño-3.4 index, with the correlation coefficient reaching only 0.01 in the observations (Table 3). In NUIST-CFS1.0 hindcasts, the average correlation coefficient between TC frequency and the Niño-3.4 index over all nine ensemble members is 0.25, which is slightly larger than that in the observations. In the entire WNP, the interannual variations of seasonal TC frequency are not closely correlated to the changes of any environment variable, as indicated by the low correlation coefficients of the observed TC frequency with the dynamic and thermodynamic factors averaged over the WNP MDR. The correlation of the predicted variables is not significantly different from that of the observed variables (Table 3). The interannual variations of seasonal mean SST and low-level VOR averaged over the respective MDR of each basin are reproduced well in the model, and the correlations between the ensemble means and the observations are around 0.80 (Table 4). As one of the dominant factors modulating the variations of WNP TC activity associated with ENSO (Camargo et al., 2007), the prediction skill for locally averaged VWS over the entire WNP is relatively low (with the correlation coefficient being 0.33).
WNP WNP-E WNP-W ENP NA MDR SST 0.79 0.74 0.77 0.88 0.77 MDR VWS 0.33 0.67 0.47 0.78 0.62 MDR VOR 0.81 0.84 0.68 0.08 0.45 MDR RH 0.64 0.61 0.63 0.18 0.23 Table 4. Correlation coefficients between predicted and observed large-scale environmental factors during the MJJASON season of 1982–2019. The correlation skills are calculated based on ensemble-mean forecasts. The values in bold indicate that the correlation coefficients are statistically significant at the 95% confidence level as examined the one-tailed Student’s t test.
Zhan et al. (2011) demonstrated that EIO SSTA is one of the factors that modulate seasonal WNP TC frequency. In the observations, the correlation coefficient between interannual variations of the seasonal mean EIO SSTAs and TC genesis frequency during 1982–2019 is −0.57 (Table 3). The prediction skill for the EIO SSTA index is rather good, with the correlation coefficient reaching 0.61 (Fig. 5b). However, the relationship between EIO SSTA and seasonal TC frequency in NUIST-CFS1.0 is not significant (the average correlation coefficient of individual members is −0.25).
The years with EIO SSTA greater (less) than one standard deviation in both the observations and each member of NUIST-CFS1.0 are selected as strong warm (cold) EIO years. Figure 6 shows the differences of the observed and predicted 850-hPa wind and sea level pressure between the warm and cold EIO years, respectively. It is found that the lower troposphere is dominated by an anomalous anticyclonic circulation over the WNP. In the model, the anomalous equatorial easterlies over the South China Sea and WNP driven by the anomalous convective heating over the EIO are more obvious. Compared with observations, the anomalous anticyclonic circulation in the model does not cover the main WNP TC genesis region and it is located further northward and eastward (Fig. 1b). Therefore, the relationship between EIO SSTA and seasonal TC frequency in model hindcasts is weak. This may be one of the reasons for the poor performance of NUIST-CFS1.0 in forecasting annual variations of TC frequency.
Figure 6. Composite differences of the (a) observed and (b) NUIST-CFS1.0-predicted 850-hPa wind (vector, units: m s−1,) and sea level pressure (shading, units: hPa) in warm and cold EIO years. Only significant sea level pressure differences are presented. Green vectors denote significant wind differences.
-
Since TC frequency over the WNP shows distinct regional features (Wang and Chan, 2002; Kim et al., 2010b) and the predictability sources of TC genesis in individual WNP subregions may differ from each other (Kim et al., 2010a; Lu et al., 2010), the performance in predicting TC activity in subregions over the WNP is assessed. Here, the longitude of 140°E is used to divide the WNP into the WNP-E and WNP-W, following Wang and Chan (2002). The ensemble mean of the hindcasts skillfully predicts the interannual variations of the seasonal TC frequency over the WNP subregions (Figs. 4c, e; Table 2). The correlation coefficients are 0.61 and 0.41 for the WNP-E and WNP-W, respectively. The model performs better for WNP-E ACE, with the correlation coefficient reaching 0.72, and relatively worse (0.23) for WNP-W ACE. NUIST-CFS1.0 can realistically reproduce most of the observed environment–TC correlations (Table 3) and capture the interannual variation characteristics of large-scale environmental indexes well (Table 4).
As shown in Table 3, observed TC frequency over the WNP-E (WNP-W) is positively (negatively) correlated with the Niño-3.4 index, indicating great impacts from ENSO. During El Niño (La Niña) years, TC formation is enhanced (weakened) in the WNP-E (WNP-W) (Wang and Chan, 2002; Wang et al., 2019). Over the open ocean, TCs that form over the WNP-E tend to be longer and stronger. The seasonal variation of ACE in the WNP-E is closely related to the Niño-3.4 index, with the correlation coefficient being 0.74 in the observations and 0.64 in the model. The Niño-3.4 index could explain ~55% of observed WNP-E ACE variations. As suggested by Wang and Chan (2002), the enhanced TC formation in the WNP-E is attributed to the increase in low-level VOR generated by El Niño-induced equatorial westerlies. Observed WNP-E ACE also has a close relationship with 850-hPa VOR, as shown in Table 3. The good performance in predicting interannual variations of TC activity over the WNP-E by NUIST-CFS1.0 is attributed to the high skill for forecasting the Niño-3.4 index (Fig. 5a). However, in the WNP-W, the relationship between ACE and the Niño-3.4 index is not significant in either the observations or the model hindcasts, and thereby, interannual variations of ACE in this region are hard to capture in NUIST-CFS1.0.
Apart from ENSO, the SSTA in the EIO also exerts statistically significant impacts on TC frequency over the WNP-E (Table 3). The correlation coefficient between WNP-E TC frequency and EIO SSTA is −0.43 in observations, while it is not significant in model hindcasts. However, the difference in correlation coefficients between model hindcasts and observations is not obvious. In addition, MDR SST, VWS, and low-level VOR are all significantly related to the interannual variation of seasonal TC frequency over the WNP-E. However, in the WNP-W, the local environmental factors are not significantly related to TC frequency, except for mid-level RH. Therefore, the interannual variations of TC activity over the WNP-W is harder to capture compared with those over the WNP-E. Moreover, it is found that the standard deviations of correlation coefficients between individual ensemble members and the SPRvERR are quite large over the WNP-W compared to those over the WNP-E (Table 2). In the WNP-E, the SPRvERR of predicted TC frequency is closer to one (after calibration), indicating the model uncertainty is well accounted for and the model predictions of TC frequency are more reliable. In the WNP-W, the ensemble spread is much larger than the RMSE (the SPRvERR equals 1.3 for predicted TC frequency) and the standard deviations of correlation coefficients are comparable to the average correlation coefficients. This indicates that there is an over-dispersion of TC frequency over the WNP-W in the model.
-
NUIST-CFS1.0 shows good performance in predicting the interannual variations of seasonal mean TC frequency over the ENP and NA (Table 5 and Figs. 7a, c). The correlation coefficients between observed and ensemble mean TC frequency are 0.58 and 0.54 in the ENP and NA, respectively. NUIST-CFS1.0 predictions of TC frequency over the NA are relatively more reliable, with the SPRvERR being one. Additionally, the multidecadal variations are captured well, especially in the ENP. Before 1994 and after 2012, TC genesis frequency over the ENP is above the climatological mean, while TCs occur less frequently during 1995–2011. The correlation coefficient of ACE in the ENP reaches 0.70 (Fig. 7b), indicating good performance by the model. The model also performs well in predicting NA ACE (Fig. 7d). The RMSEs for calibrated TC frequency in the ENP and NA are comparable to one standard deviation of the observed TC counts (Table 5). For ACE, the RMSE is less than one standard deviation of the observations, which means the prediction errors for ENP and NA ACE are acceptable. The correlation coefficients of individual members are also smaller than those of the ensemble means (Table 5).
CC_M CC_I RMSE STDV_O SPRvERR TC frequency ENP 0.58 0.39 ($ \pm $0.09) 4.80 4.79 1.38 NA 0.54 0.41 ($ \pm $0.09) 4.76 4.81 1.00 ACE ($ {10}^{4}{\mathrm{k}\mathrm{t}}^{2} $) ENP 0.70 0.48 ($ \pm $0.13) 58.87 82.51 1.28 NA 0.47 0.36 ($ \pm $0.09) 65.73 70.34 0.78 Table 5. Similar to Table 2, but for TC frequency over the ENP and NA.
Seasonal TC activity over the ENP and NA is modulated by climate variabilities like ENSO (Camargo et al., 2010; Zhao et al., 2010). In addition, SST_REL is also found to be useful in predicting seasonal TC activity (see also Zhao et al., 2010; Vecchi et al., 2011). As shown in Table 6, the interannual variations of TC frequency and ACE in the ENP are significantly related to SST-based indexes (Niño-3.4 index, ENP SST_REL, and MDR SST) and MDR VWS. It is found that the observed correlations of TC frequency and ACE in the ENP with the corresponding climatic indexes are reproduced well by NUIST-CFS1.0. In particular, the hindcast skill for ENP MDR SST and VWS is fairly high (Table 4). As proposed by Zhao et al. (2010), the quality of the prediction of seasonal TC activity in a coupled atmosphere–ocean model depends largely on the model’s ability to predict the evolution of SST_REL. In NUIST-CFS1.0, SST_REL in the ENP is well forecasted, with the correlation coefficient between the predicted and observed values reaching 0.79 (Fig. 8a). However, the relationship between TC activity and ENP SST_REL in the model is significantly underestimated compared to observations. Figure 9 shows the correlation between the relative SST anomaly and ENP TC frequency in observations and the hindcasts. It is found that in observations, the strongest correlations are located near the western coast of Mexico and the ENP MDR region (Fig. 9a). In the model, the maximum correlations appear in the region of 10°S–30°N, 180°–100°W (Fig. 9b). Therefore, ENP SST_REL is not a major predictor for annual TC counts in the model.
ENP NA OBS CFS1.0 OBS CFS1.0 Niño-3.4 0.46 0.51 −0.32 −0.56 0.43 0.53 −0.40 −0.57 SST_REL 0.71 0.31* 0.67 0.61 0.68 0.33* 0.64 0.61 MDR SST 0.52 0.41 0.65 0.46 0.45 0.44 0.58 0.48 MDR VWS −0.54 −0.35 −0.67 −0.54 −0.55 −0.35 −0.65 −0.56 MDR VOR 0.10 0.12 0.43 0.56 0.06 0.12 0.47 0.57 MDR RH −0.05 0.16 0.33 0.56 −0.10 0.16 0.37 0.52 Table 6. Similar to Table 3, but for the correlations in the ENP and NA. MDR SST, VWS, VOR, and RH are averaged over 7.5°–15°N, 160°–80°W in the ENP, and over 7.5°–22.5°N, 80°–20°W in the NA, respectively. SST_REL is defined as the difference in SST between the MDR of the ENP and NA and that in the global tropics (30°S–30°N).
Figure 8. Same as Fig. 4, but for the predictions of (a) ENP SST_REL anomaly and (b) NA SST_REL anomaly.
Figure 9. Spatial distributions of the correlation coefficients between SST_REL and ENP TC frequency in (a) observations and (b) NUIST-CFS1.0. The dotted areas denote correlation coefficients are statistically significant at the 95% confidence level.
In the NA, the observed correlations between TC activity and corresponding climatic indexes are reproduced well in model hindcasts (Tables 4, 6, and Fig. 8b). In the model, TC frequency over the NA is significantly related to the Niño-3.4 index (the correlation coefficient is −0.56), while the correlation is not significant in observations (Table 6). As discussed in Wang and Lee (2009), TC activity over the NA is out of phase with that over the ENP, which is closely related to the factors of VWS and convective instability. In observations, the correlation coefficient of TC frequency between these two basins is −0.47. In the model, this out-of-phase relationship is well represented. The mean correlation coefficient of predicted TC frequency between the NA and ENP is −0.75, larger than that in observations. Figure 10 shows the regression of VWS onto the Niño-3.4 index and NA genesis density in observations and the model hindcasts. The spatial patterns of regressed VWS are quite similar between observations and the hindcasts. However, there is an absence of cyclogenesis over the Gulf of Mexico and the east coast of the United States in the hindcasts where VWS is negatively correlated with the Niño-3.4 index. TC genesis appears in the region where VWS is positively correlated with the Niño-3.4 index in the hindcasts. This could explain why the correlation between TC frequency and the ENSO index is much larger in NUIST-CFS1.0 than in observations.
Figure 10. Regression coefficients of the MJJASON VWS (shading) onto the time series of Niño-3.4 index during 1982–2019 in (a) observations and (b) the NUIST-CFS1.0 hindcasts. The regression coefficients in (b) are the mean values of the coefficients over all nine members. The purple contours denote genesis densities in the NA.
-
The good performance in predicting tropical circulation and tropical SST-based variations such as ENSO, EIO SSTA, and SST_REL in NUIST CFS1.0 lays a foundation for the skillful prediction of TC activity variations. The model also captures the observed relationships between environmental factors and TC activity well. However, the model’s performance in predicting interannual variations of TC frequency and ACE in the NH can be further improved. As discussed in section 4.1, the predictability sources of TC activity vary between different subregions of the WNP. In the WNP-E, the SST-related indexes, such as the Niño-3.4 index and EIO SSTA, are identified as good predictors for TC frequency variations. However, in the WNP-W, the relationship between the Niño-3.4 index and TC frequency is not that significant. The eastward shift in the location of TC genesis in NUIST-CFS1.0 causes a higher correlation coefficient between the ENSO index and WNP TC frequency than that in observations. The results in section 4.2 also show that the decrease in TC activity over the Gulf of Mexico and the east coast of the United States in the model leads to a more significant relationship between NA TC frequency and the Niño-3.4 index than what is seen in observations. Therefore, the deviation of predicted TC genesis location may lead to the inaccurate TC–environment relationship, which is essential to achieving good prediction skill for interannual variations of TC activity.
-
The skill of NUIST-CFS1.0 in predicting seasonal TC activity is compared with the skill of other forecast systems built based on an ocean–atmosphere coupled model. These forecast systems are developed by the UK Met Office (Camp et al., 2015), Geophysical Fluid Dynamics Laboratory (GFDL) (Murakami et al., 2016), European Centre for Medium-Range Weather Forecasts (ECMWF) (Manganello et al., 2016), and Japan Meteorological Agency (JMA) (Takaya et al., 2010). The information for predicted seasonal TC activity in each forecast system is summarized in Table 7, and the correlation coefficients are presented in Fig. 11.
Forecast system Resolution Predicted years Target months References Met Office GloSea4 ~120 km 1996–2009 6–11 Camp et al., 2015 Met Office GloSea5 ~60 km 1996–2009 6–11 Camp et al., 2015 1992–2013 6–11 GFDL FLOR ~50 km 1980–2015 7–11 Murakami et al., 2016 GFDL HiFLOR ~25 km 1980–2015 7–11 Murakami et al., 2016 Project Minerva T319, ~60 km 1980–2011 5–11 Manganello et al., 2016 JMA/MRI-CGCM ~180 km 1979–2006 6–10 Takaya et al., 2010 Table 7. List of forecast systems that predict interannual variations of seasonal TC activity: approximate horizontal resolution of the atmospheric component, predicted years, target months, and references.
Figure 11. Correlation coefficients of seasonal TC frequency and ACE between observations and the model hindcasts. The models with low horizontal resolutions (> 100 km) are marked in red, and those with high horizontal resolutions are marked in blue (~50–60 km) or green (~25 km). Bold markers indicate the correlation coefficients are statistically significant.
Similar to NUIST-CFS1.0, the interannual variation of seasonal TC frequency in the WNP is difficult to capture in the models with low horizontal resolutions (>100 km), such as the UK Met Office Global Seasonal Forecast System version 4 (GloSea4) and the JMA/Meteorological Research Institute (MRI) CGCM (Fig. 11a). For high-resolution models, the UK Met Office Global Seasonal Forecast System version 5 (GloSea5) and the experimental coupled prediction system based on ECMWF system 4 (Project Minerva) can predict the interannual variations of WNP TC frequency well. However, the correlation coefficients in the GFDL 50-km-resolution Forecast-oriented Low Ocean Resolution model (FLOR) and its 25-km-resolution version (HiFLOR) are not significant (0.13 for FLOR and 0.28 for HiFLOR). The skill in predicting seasonal ACE in the WNP for NUIST-CFS1.0 is comparable to the skill for the forecast systems listed in Table 7 (Fig. 11b). All the models can capture interannual variations of seasonal WNP ACE well, with the correlation coefficient for GloSea5 being the highest (>0.8).
The correlation coefficients for ENP and NA TC frequency in NUIST-CFS1.0 are above 0.5, which are higher than those in GloSea4 (Fig. 11a). With a similar horizontal resolution, GloSea4 cannot capture the interannual variations of seasonal TC frequency over the ENP and NA. The skill in reproducing ENP and NA TC frequency for NUIST-CFS1.0 is comparable to the skill for the high-resolution models (GloSea5, FLOR, and HiFLOR). It is noteworthy that the correlation coefficients for ENP TC frequency during 1996–2009 from GloSea5 and the NA TC numbers in the T319 experiments of Minerva are not significant. NUIST-CFS1.0 also shows a relatively higher skill in predicting the seasonal ACE index over the ENP among the forecast systems listed here (Fig. 11b). The correlation coefficient for NA ACE in NUIST-CFS1.0 is higher than that in GloSea4 and is comparable to that in GloSea5 and Minerva. Additionally, FLOR and HiFLOR show better skill for NA ACE. In general, the prediction skill for TC frequency and ACE from NUIST-CFS1.0 is comparable to that in the other forecast systems.
WNP | ENP | NA | ||
TC frequency | ||||
OBS | 22.39 | 17.66 | 12.61 | |
CFS | 11.33 | 5.07 | 5.37 | |
ACE (104 kt2) | ||||
OBS | 187.36 | 166.15 | 124.81 | |
CFS | 36.52 | 13.79 | 16.31 |