Retrospective seasonal predictions of tropical cyclones (TCs) in the three major ocean basins of the Northern Hemisphere are performed from 1990 to 2010 using the Geophysical Fluid Dynamics Laboratory High-Resolution Atmospheric Model (HiRAM) at 25-km resolution. Atmospheric states are initialized for each forecast, with the sea surface temperature anomaly (SSTA) “persisted” from that at the starting time during the 5-month forecast period (July–November). Using a five-member ensemble, it is shown that the storm counts of both tropical storm (TS) and hurricane categories are highly predictable in the North Atlantic basin during the 21-yr period. The correlations between the 21-yr observed and model predicted storm counts are 0.88 and 0.89 for hurricanes and TSs, respectively. The prediction in the eastern North Pacific is skillful, but it is not as outstanding as that in the North Atlantic. The persistent SSTA assumption appears to be less robust for the western North Pacific, contributing to less skillful predictions in that region. The relative skill in the prediction of storm counts is shown to be consistent with the quality of the predicted large-scale environment in the three major basins. It is shown that intensity distribution of TCs can be captured well by the model if the central sea level pressure is used as the threshold variable instead of the commonly used 10-m wind speed. This demonstrates the feasibility of using the 25-km-resolution HiRAM, a general circulation model designed initially for long-term climate simulations, to study the impacts of climate change on the intensity distribution of TCs.
Using a global weather or climate model for seasonal predictions of tropical cyclone (TC) activity has been shown to be quite successful in recent years (Vitart 2006; Vitart et al. 2007; LaRow et al. 2010; Zhao et al. 2010, hereafter Z10; Chen and Lin 2011, hereafter C11). Unlike the more established statistical methods (Gray 1984; Elsner and Jagger 2006; Klotzbach and Gray 2009), with a global dynamical model, TCs of various intensities and sizes can naturally occur in regions with favorable storm-permitting large-scale environments. The seasonal evolution of large-scale atmospheric circulations associated with the TC activity can also be simultaneously available from the model prediction.
For the North Atlantic basin, it has been recently demonstrated that dynamic-based models are generally more skillful than pure statistical models in the retrospective TC’s seasonal prediction for the past recent decades. Vitart et al. (2007) achieved a high correlation of 0.81 between the observed and predicted Atlantic hurricane counts during 1993–2006 using the European Seasonal to Interannual Prediction multimodel ensemble of coupled ocean–atmosphere models. LaRow et al. (2010) obtained a correlation of 0.74 during 1986–2005 using the Florida State University–Center for Ocean Atmospheric Prediction Studies atmospheric global spectral model with sea surface temperature (SST) from the National Oceanic and Atmospheric Administration (NOAA) Climate Forecast System Model. With recent increase in computing capability, global models with 50 km or higher resolution have been increasingly applied to seasonal predictions of TCs. Z10, using the Geophysical Fluid Dynamics Laboratory (GFDL) High-Resolution Atmospheric Model (HiRAM) with the 50-km resolution, obtained correlations (between the observed and the model ensemble means hurricane counts) of 0.69 in the North Atlantic and 0.58 in the eastern Pacific during the 1982–2008 period by using the SST anomalies (SSTAs) averaged over June. It should be noted that Z10 did not initialize their forecasts. Instead, initial conditions are from a corresponding Atmospheric Model Intercomparison Project (AMIP)-type climate simulation (Gates 1992), and the ensembles were generated by extremely small random perturbations in the SST fields. Recently, CL11, using the 25-km version of the GFDL HiRAM, achieved a remarkable correlation of 0.96 between the observed and model predicted hurricane counts during the past decade (2000–10) in the North Atlantic. They adopted a similar strategy (as in Z10) of using the persistent SSTA, but with the atmospheric states for each forecast initialized with the National Centers for Environmental Prediction (NCEP)–Global Forecast System (GFS) analysis data.
For a skillful seasonal prediction, the initial atmospheric state is generally regarded as less important, at least not being considered as a dominant factor. From this perspective, it may be argued that a global model with “climate model heritage” is more advantageous for seasonal predictions than a model originally developed for short-term weather forecasts, which is considered mostly an initial-value problem. However, TCs typically originated from organized convective systems with the underlying ocean state playing a vital role in determining their life cycle from genesis to dissipation. Those organized convective systems, in turn, have their origin in cloud-scale motions that are severely underresolved by today’s global climate and weather models. Hence, for TC seasonal predictions, improved model resolution and higher quality of the oceanic state should be key to further improvement of the skills as demonstrated by the aforementioned studies. It was therefore suggested by CL11 that a coupled atmosphere–ocean global model with cloud and eddy resolving capability would be the ultimate tool. However, it is still cost-prohibitive at present to use a fully coupled model with resolution in the kilometer range for the retrospective predictions of interannual TC activities spanning several decades. Nonetheless, the version of the HiRAM at 25-km resolution used by CL11 and for this study is our attempt in that direction, and, in fact, the 25-km HiRAM can be regarded as a prototype global cloud system–resolving model (see section 2), consistent with the concept of seamless weather–climate prediction proposed by Shukla et al. (2009). Besides doubling the horizontal resolution from 50 km (as in Z10) to 25 km, CL11’s work indicated that the forecast in the tropics could be improved with an initialized atmospheric state. To support their argument, CL11 presented a successful prediction of the Madden–Julian oscillation, which would be impossible to achieve without the proper initialization of the atmospheric state.
In this follow-up study, we extend the retrospective seasonal predictions from 1990 to 2010 with the same model configurations and methodology as in CL11. We also extend the prediction one more month to the end of the official hurricane season (1 December). One more ensemble member is added to ensure the internal variability of the model will not be a significant factor affecting the skills of the ensemble. Detailed analyses are provided for the three major basins for the evaluation of the suitability of the persistent SSTA assumption on the TC seasonal prediction. The descriptions of the HiRAM, the data used for the initialization, and the methodology of storm tracking are given in section 2. Section 3 shows the results of storm counts in the three basins and the quality of the predicted large-scale environment. The examination of the persistent SSTA assumption will be presented in section 4. Section 5 provides the analyses of the storm intensity and a discussion of its relevance to potential application of the HiRAM for climate change. A summary and conclusions are presented in section 6.
2. Model, data, and methodology
The GFDL High-Resolution Atmospheric Model is a global model with flexibility in horizontal resolution designed for applications ranging from 5-day forecasts to century-long climate-change assessments. It is based on the GFDL Atmospheric Model, version 2 (AM2; Anderson et al. 2004), with key modifications described in Zhao et al. (2009). The version at 25-km resolution used in CL11 and this study is called C360-HiRAM; it uses the “vertically Lagrangian” finite-volume dynamical core (Lin 2004) on the cubed-sphere grid (Putman and Lin 2007) with 360 × 360 finite-volume cells on each face of the cube and 32 layers with the model top located at 1 hPa. A nonintrusive shallow convective scheme is adopted, replacing the traditional deep convective parameterization [see appendix A in Zhao et al. (2009)]. Different from the C180-HiRAM (Zhao et al. 2009; Z10), and designed for eventual global cloud-resolving simulations, a six-category bulk cloud microphysics scheme [similar to that of Lin et al. (1983)] is adopted for the resolved component of the moist processes. With the bulk cloud microphysics scheme, the rapid fall of condensates (e.g., rainwater) usually imposes a severe restriction on the size of the model time step. To improve the computational efficiency and to reduce numerically induced vertical mixing, terminal fall of all condensates (rain, snow, graupel, and cloud ice) are treated in a Lagrangian manner using the same high-order conservative remapping scheme in the dynamical core. The impact of high-wind speed on the parameterized drag coefficient Cd for the momentum exchange over ocean is included following Moon et al. (2007). The efficiency of the orographic gravity wave drag parameterization is, unfortunately, highly dependent on horizontal resolution. For the C360-HiRAM, a multiplying factor of 2.5 is used (vs. 1.0 in the standard GFDL AM2). The full model integration uses a time-split strategy with different time step for different modules. The time steps for the slow physics (radiation and ocean coupling), fast physics (e.g., shallow convection, boundary layer, gravity wave drag, and land surface model coupling), and the cloud microphysics are 3600, 600, and 300 s, respectively. The time step for the dynamics is adaptive, starting with the vertical remapping time step the same as the fast physics (600 s). There are 20 time splits for the fast horizontally processes (Lin 2004), resulting in a small time step of 30 s. The tracer time step is usually the same as the fast physics; further time splits are automatically activated whenever needed to maintain stability [the Courant–Friedrichs–Lewy (CFL) condition must be less than 1 for each horizontal direction].
The NCEP–GFS analysis data (http://www.emc.ncep.noaa.gov/GFS/doc.php) are used for the initialization of the model atmospheric states. Because of the fundamental differences in the physical parameterizations in general, and convective parameterizations and cloud microphysics in particular, the three-dimensional wind and temperature fields in HiRAM are nudged toward the NCEP analysis for one month, starting from 1 June for each season. HiRAM’s cloud microphysics fields and the land surface model are considered spun up this way. To constitute an ensemble, one-day-lagged initial conditions are used to provide five forecast members. The first member begins at 0000 UTC 1 July, while the second begins at 0000 UTC 30 June, and so on. For each member, the SST used for the following 5-month prediction to 1 December is based on the monthly-varying SST climatology plus the observed SSTA, which is defined as the departure of the NCEP–GFS ocean skin temperature from the SST climatology. The observed SSTA is “frozen” (from initial forecast time) during the 5-month forecast period with the intraseasonal variation of SST entirely determined by the background SST climatology. Climatology values for ozone and greenhouse gases are used. With this approach, true forecasts are performed by using only information available up to forecast time.
The GFDL vortex tracker (Marchok 2002), which has been operationally used at NCEP since 1998 with the updated warm-core criterion (Gall et al. 2011), is adopted for tracking TCs in the forecasts. The tracker was originally designed for the GFDL regional hurricane model for short-range TC forecasts with the capability to detect and track vortices even if their intensities only meet the tropical depression (TD) threshold (with the maximum wind speed at 17.5 m s−1 or less). Landsea et al. (2010) and Villarini et al. (2011) indicated that a false upward trend in observations on decadal time scales may be due to the better identification of the short-lived storms than that in the pre-satellite era. To avoid this issue, short-lived storms are filtered out from our model results. Furthermore, based on the wind speed and warm-core information provided by the tracker, we set up the following criteria to better define tropical storms (TSs) in our model.
The total lifetime of a TS (including the TD stage) must be longer than 72 h.
The TS must have a warm core lasting longer than 48 h cumulatively during its lifetime.
The maximum 10-m wind speed of the TS must be greater than 17.5 m s−1 together with a warm core for longer than 36 h, consecutively.
The genesis location of the storm must be lower than 40°N.
For an identified long-lived TS to qualify as a hurricane, the maximum 10-m wind must be greater than 32.5 m s−1 at some point during its lifetime. Walsh et al. (2007) discussed the relationship between model resolution and criterion for storm detection. They suggested that a detection scheme should use wind speeds on the 10-m height corrected by the model resolution. Their analyses also showed that the variation of the TS threshold does not change much when the model resolution is higher than 25 km (Fig. 2 in Walsh et al. 2007). Therefore, it is reasonable and less ambiguous to use the official wind definition of the TS without executing resolution correction for storms generated by the HiRAM at 25-km resolution.
The observed hurricane numbers in the North Atlantic and the eastern North Pacific basins (typhoon numbers in the western North Pacific basin) are based on the tropical cyclone reports from the National Hurricane Center (Joint Typhoon Warning Center) (Jarvinen et al. 1984; Davis et al. 1984; Chu et al. 2002). To be consistent with the selection of the model TS, the observed TSs are further filtered from the International Best Track Archive for Climate Stewardship (IBTrACS; Kruk et al. 2010) data based on the same criteria but without the warm-core criterion since this information is not available. Therefore, the observed numbers of TS used in this study can and will differ from those of the “named storms” in the official TC reports. The information of wind speed, sea level pressure, and storm locations provided by the IBTrACS data is used as the observations for the analyses and comparisons in the following sections. The monthly-mean NCEP optimum interpolation SST (OI SST) and the interim European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-Interim) monthly-mean analysis data (Simmons et al. 2006) are used to examine the persistent SSTA assumption and the model predicted climatology.
3. Main results
The model-predicted climatology is first examined to evaluate the quality of the model-predicted large-scale environments that may influence the genesis and geographical distribution of the TCs. Relevant to TC activities in the Atlantic basin, we compare the African easterly jet (AEJ) predicted by the model to that in the ERA data (Fig. 1). It has been suggested that the North Atlantic (NA) TC genesis can be related to the “African easterly waves,” which are associated with the instability of the AEJ (Carlson 1969; Shen et al. 2010). From the vertical structure of the mean zonal wind during August–October (ASO; the peak period of the hurricane season) from the 21-yr ERA data, the AEJ between 30°W and 10°E (Fig. 1a) shows a distinct maximum 8 m s−1 at 14°N around 600 hPa. A very similar AEJ structure can also be identified from the ensemble mean of the HiRAM forecast (Fig. 1b). Besides the AEJ, both the ERA data and the model forecast show an upper-level tropical easterly jet south of the AEJ as well as an strong upper-level westerly jet at around 30°N. Another low-level westerly flow around 5°–10°N is predicted in good agreement with the ERA data. The Hovmöller diagrams of the zonal wind at 600 hPa illustrate the intraseasonal variation of the AEJ (Figs. 1c,d). Both the model and ERA data show that the maximum wind speed (10 m s−1) lasted to September and then gradually become weaker with a southward shift. Climatologically speaking, these large-scale AEJ characteristics that may be important to TC genesis are simulated very well by the C360-HiRAM in the seasonal prediction mode.
Relevant to the TC activities in the western North Pacific basin, the diagnostic fields associated with the subtropical high and the Asian monsoon are investigated. The expansion or contraction of subtropical high can affect TC tracks in the western Pacific (Chan et al. 2001; Wu et al. 2007; Chen et al. 2009), while the summer monsoon circulation in East and Southeast Asia is very important to the TC movement and genesis (Davidson 1995; Lander 1996; Chen et al. 2004). The model-predicted 21-yr (1990–2010) average sea level pressure (SLP) and 850-hPa wind fields during ASO are compared with those in the ERA-Interim data (cf. Figs. 2a,b). It shows that the HiRAM predicted subtropical high is relatively weaker than that in the ERA-Interim analysis. Meanwhile, the range of the monsoon trough is more extended to the northeast in the model than in the ERA-Interim analysis.
The daily precipitation of the HiRAM forecast during ASO is also compared with the Tropical Rainfall Measuring Mission (TRMM) 3B43 surface precipitation data (Huffman et al. 1997). To be consistent with the available period of the TRMM data, the average period is during 1998–2010 in Fig. 3. From the comparison, it can be found that the global pattern of the model-predicted precipitation is quite similar to that from the TRMM data, except for the overprediction shown in the Indian Ocean around the equator and the underprediction in the intertropical convergence zone (ITCZ) between 150° and 165°E. Our model shows overall stronger precipitation rate than that in the observation, especially in the ITCZ over the central to eastern Pacific and the Atlantic. Note that some previous studies indicated that the TRMM’s precipitation rate is often underestimated in the heavy rainfall region (Bowman et al. 2003; Berg et al. 2006; Liao and Meneghini 2009). The pattern of the model predicted precipitation extended from southwest to northeast around 10°–20°N, 130°–150°E is consistent with the low pressure area associated with the monsoon trough shown in Fig. 2b. The northeastward extension of the trough circulation may contribute to the overpredicted storm number in this basin, which will be illustrated later in this section. We therefore proceed to the presentation of the retrospective predictions of TC frequency (storm counts).
The 5-month (July–November) observed and model predicted hurricane–typhoon and TS counts in the three basins during 1990–2010 are shown in Fig. 4. The correlations, mean root-mean-square errors (RMSEs), and mean biases for the IBTrACs observation for the five forecast members and the ensemble mean for three different time periods are listed in Table 1. The correlation between the observed and model-predicted hurricane counts is 0.88 for the entire 21-yr period, while the correlations are 0.78 and 0.94 for the first and second decades, respectively (Table 1). For the same 2000–10 period, the slight drop from 0.96 (as obtained by CL11) to 0.94 is due to the one-month extension to 1 December. Negative mean biases (~1) and small mean RMSEs (lower than 2) are obtained for all three periods (Table 1). The forecast skill of the TS category is similarly good in the NA (Fig. 4b), with correlations of 0.89, 0.84, and 0.92 for the 21-yr period and the first and the second decades, respectively (Table 1). Positive mean biases are shown in the TS category, which is different from those in the hurricane category and the result in CL11. Based on the analyses of intraseasonal variation (Fig. 5b), these biases mostly come from the overprediction in October and November. Consequently, the bias changes from small negative (2000–10; CL11) to positive when the month of November is included in this study. In contrast, the negative mean bias for hurricanes is mainly from the underpredictions in September, which is the peak month of the NA hurricane season (Fig. 5a). One of the plausible reasons is that the signal of the warmest SST peak during individual season is excluded from the model since the background SST is from the climatology, which is time-smoothed.
The prediction in the eastern North Pacific (ENP) is not as outstanding as that in the NA, but still can be considered skillful (Figs. 4c,d). The correlations between the observed and model-predicted storm counts are 0.6 for both the hurricane and the TS categories with negative mean biases during the past 21 years (Table 1). Note that the correlation in the first decade is superior to that in the second. The inconsistency between the two decades is also shown in the NA but the higher skill is in the second period and the difference is much smaller. The intraseasonal variation of the storm numbers in this basin is different from that in the NA (Figs. 5c,d). Instead of showing a peak in September, most of the observed storms happened in the first three months with rather even distribution. The shortage of the model-predicted storms in the three months results in the large negative biases (Table 1). Interestingly, the bias in TS numbers is less than that of hurricanes, which indicates that it is more difficult for an ENP TS to intensify into a hurricane in our model than in nature. This may be explained by the fact that hurricanes over ENP are generally found to be smaller in size than those over other basins (Chavas and Emanuel 2010), and it is more difficult for the 25-km-resolution HiRAM to support the development of ENP compact hurricanes.
Relative to the results in the NA and the ENP basins, the prediction in the western North Pacific (WNP) is not skillful (Figs. 4e,f). The correlations between observed and model-predicted storm counts are around 0.35 for both categories for all different time periods (Table 1). Our model generated too many storms in this basin, showing large positive bias in the TS category. The overprediction can also be found in typhoon category, especially in November (Fig. 5e). In CL11, the correlations were 0.77 (0.6) for the four-month seasonal TS (typhoon) predictions during 2000–10, which should be considered very skillful predictions. The huge degradation in skill partly came from the large deviation in the last month (November) included only in the present study (Figs. 5e,f). It is speculated here that over the WNP, oceanic feedback is relatively more important (as compared with NA), and the prescribed “ocean” could not response to any external forcing. For example, the accumulated cooling effects of storms during the season are not included, which would tend to overproduce the storm counts. Despite the year-to-year correlation of the seasonal predictions being not good in the WNP, the decreasing trend of the storm count during the past 21 years is still well captured by the model (Figs. 4e,f).
Figure 6 shows the comparison of the 21-yr hurricane/typhoon occurrence locations between the observation and the model predictions. It is interesting that the geographical distribution patterns are quite different among the three basins. The pattern is rather scattering in the NA but much more concentrated in the other two basins, which can be found in both observations and model forecasts. In the WNP, the typhoon occurrence in the model forecast is mostly located within the same area as that in the observations, but without showing distinct peaks. The maximum frequency in the ENP is located between 100° and 115°W from the observation (Fig. 6a), while the occurrence locations in the model are similar to the observed but more concentrated, showing a smaller range of the peak (Fig. 6b). In the NA, this scattering characteristic may related to the relationship between the TC genesis and the African easterly waves (Chen et al. 2008). The pattern of the model forecast is generally in agreement with that of the observation, which could be attributed to the good reproducibility of the AEJ in the HiRAM shown in Fig. 1.
Some relationships between biases in the storm counts of the three basins and the large-scale atmospheric patterns can be exhibited by the biases (relative to the ERA-Interim analysis) in model-predicted mean SLP and vertical wind shear during ASO period (Fig. 7). A broad negative SLP bias is located within 20°–30°N, 130°E–180°, which is consistent with the positive bias of the storm number in the WNP. In contrast, the positive bias of the SLP shown in the main TC genesis region of the ENP is in agreement with the negative model storm counts in that basin. The positive wind shear difference is also found in the same area of the ENP, implying that the TC development in the model could be inhibited by the stronger wind shear than that in the observation. As compared to the Pacific Ocean, over the NA main development region [MDR; 10°–20°N, 80°–20°W, as defined by Vecchi et al. (2011)], both the SLP and the wind shear biases are much smaller, which is consistent with the more skillful storm-count prediction for the NA basin. Note that the storm-inhibiting positive wind shear is found over the southwestern area of the Gulf of Mexico, which may provide an explanation for the model’s larger negative biases in this region (as shown in Fig. 6).
A frequently asked question is how many ensemble numbers are required for the seasonal prediction of TCs using the GFDL HiRAM. From Table 1, it is noted that the correlations from individual members are generally lower than that of the ensemble mean in both categories. This result is expected since the use of the ensemble mean is to filter out the internal variability of the particular model. It has been examined that if any four of five ensemble members are considered, the average correlations of all four-member combinations are 0.87, 0.6, and 0.34 for the NA, the ENP, and the WNP hurricanes (typhoons), respectively. Moreover, when the ensemble number is reduced to three, the average correlations of all three-member combinations for the three basins in the same order as above are 0.83, 0.59, and 0.33. It can be found that the differences in correlations between the uses of different member sizes are insignificant for the ENP and the WNP. However, the difference (degradation in skills) is much evident if the ensemble members are reduced to three in the NA. Accordingly, for the GFDL HiRAM, four to five ensemble members should be sufficient for seasonal predictions of TCs in the NA basin. Nevertheless, the model ensembles do not show much of the internal variability in the other two basins.
Wu et al. (2012) used regional climate model to demonstrate the importance of the internal variability to the TC activity simulations on seasonal time scale. In Table 2, the correlations of individual member to the mean of the remaining four members are listed to provide some measure of the internal variability of the HiRAM seasonal prediction. It is found that the correlations of model mean to the individual member in the NA are significantly lower than those in the ENP and WNP. It is an interesting finding that is consistent with the above results that the HiRAM has higher internal variability in the NA than in the other two basins. However, recall from Table 1 that the skill of seasonal prediction in the NA is the best among the three under the persistent SSTA assumption. It demonstrates that the initial SSTA is the dominant factor affecting the TC’s seasonal activity in the NA. Therefore, even with large ensemble spread, our model can still achieve a skillful storm number prediction in this basin.
4. Examination of persistent SST anomaly assumption
In this study, a true forecast mode is enforced by using the SSTA at the forecast initial time, which is “frozen” during the 5-month prediction. The validity of this critical assumption is examined by comparing the persisted SSTA to the observed monthly-mean SSTA, which is computed as the difference between the NCEP OI SST and the climatology SST used in the model. The 21-yr correlation and mean RMSE of the two SSTA fields are shown in Fig. 8. As expected, the SSTA used in the model (five-member mean) is highly correlated with the monthly-mean observed SSTA in July (Fig. 8a), with the correlation gradually decreasing to the end of a TC season (Figs. 8c,e). A similar transition is found in the 21-yr mean RMSE. In July, small errors are shown over most ocean regions (Fig. 8b) and the SSTA used by the model gradually loses its similarity to the observation in the late months (Figs. 8d,f). The good correlation can be maintained up to November in the Niño-3.4 region (5°S–5°N, 120–170°W), while the SSTA in the NA MDR is also well correlated to the observations during the entire season. The 21-yr mean correlations within the Niño-3.4 region (NA MDR) are 0.94 (0.94), 0.86 (0.78), and 0.85 (0.64) for July, ASO, and November, respectively. In contrast, the SSTA correlations are relatively low in the respective TC genesis regions of the ENP and the WNP. The 21-yr mean correlations are 0.75 (0.71), 0.75 (0.37), and 0.58 (0.34) within the ENP (WNP) MDR for July, ASO, and November, respectively. The definitions of the ENP MDR (10°–20°N, 90°–140°W) and the WNP MDR (5°–15°N, 130°E–180°) are followed Collins and Roache (2011) and Emanuel (2005).
To discuss the relationship between the local SSTA and the model bias of the predicted storm numbers, the difference between the model used and observed SSTAs within the basinwide region of different bias groups are shown in Fig. 9. The ratio of the bias (the model-predicted count minus the observed count) to the observed hurricane/typhoon count in each year is first computed for each basin (Table 3). For the NA and the WNP, the years with positive (negative) bias ratio over 20% are grouped into the “positive-year” (“negative-year”) group. The threshold is 30% for the ENP since ratios in this basin are generally large. As listed in Table 3, there are four positive years and six negative years in the NA, while the numbers of positive and negative years are 2 (6) and 11 (4) in the ENP (the WNP).
Different from the RMSE shown in Fig. 8, the SSTA differences during ASO period are presented in Fig. 9 to identify the warmer or cooler SSTA within the basinwide areas in the model (as compared to the observed). The 21-yr, positive-year, and negative-year average SSTA differences for the three basins are shown. In Fig. 9a, a cooler pattern is within the belt region (10°–20 °N) between western Africa and the Gulf of Mexico. It is consistent with the negative bias of the hurricane counts in this basin (see NA HU in Table 1). This cooler feature can be even more significant in the six largest negative years (Fig. 9c). In contrast, in the four positive years, no similar cooler pattern is found in the MDR, but there are some warmer areas located at 20°N, 60°–30°W and within the Caribbean Sea (Fig. 9b). These characteristics can be linked to the positive biases of model predicted hurricane counts in these years.
The negative biases of the ENP hurricane counts are much larger than those in the NA (see ENP HU in Table 1). The patterns of the average SSTA differences are similar for the 21 years and the negative group (Figs. 9d, f), which can be expected since more than half of the years are grouped into the negative years even with a higher threshold applied in this basin. The cooler area is located around 14°–18°N, 130°–110°W, which is northwest to the peak hurricane occurrence region shown in Fig. 6. It is worth noting that the average SSTA differences of the two positive years shows a totally different pattern in Fig. 9e. The regions with warmer SSTA spread over the entire tropical ENP, indicating the influence of the SST on the hurricane genesis in this basin.
In contrast to the above two basins, the model-predicted typhoon counts show positive biases in the WNP (see WNP TY in Table 1). From the 21-yr average SSTA differences (Fig. 9g), there is not much signal in the tropical WNP, but a warmer SSTA belt (between 135°E and 180°) is located at higher latitude (17°–30°N) (Fig. 9g). The warmer SSTA areas between 160°E and 180° expand to lower latitude in the average of the six largest positive bias years (Fig. 9h), while some cooler SSTA signals are shown within 0°–15°N and 135°E–180° and when negative bias years are considered (Fig. 9i).
The results in Fig. 9 clearly indicate that the SSTA can affect the hurricane/typhoon genesis in all three basins. Nevertheless, the skills of the TC’s seasonal prediction show large variations among the three basins. In an effort to evaluate if these variations are due to the inadequacy of the persistent SSTA assumption within the ENP and WNP MDRs (see Fig. 8), we carried out 6-yr simulations (2002–07) using the Hadley Centre monthly SST data (Rayner et al. 2003). The mean RMSEs of the storm counts are slightly reduced in the simulations with the Hadley Centre SST (not shown). These limited results suggest that the skill of the TC’s seasonal prediction using the persistent SSTA may be comparable to the simulation with observed SST data. However, longer-period experiments are required to fully address the impact of this assumption, while the use of the SST data with finer temporal resolution (e.g., weekly SST dataset) should be necessary. These experiments cannot be completed in this study since we are restricted by the available computing resources.
Another speculation for the high predictability in the NA is that the impact from the SSTs on the TC’s seasonal genesis is more direct than that in the other basins where more indirect impacts of the SSTs through other atmospheric circulations could also affect the TC predictability. For example, the Asian monsoon is strongly affected by the SSTs and also affects the TC genesis in the WNP. The air–sea interaction between the Asian monsoon and ocean also needs to be considered since a coupled model is not used in this study. In contrast, there may be less indirect impacts on the NA TCs associated with the local SSTs. For example, the AEJ, which is an important factors for TC genesis in the NA, is sensitive to the temperature discontinuities due to the Sahara, but less so to the SSTs. Therefore, applying the same persistent SSTA assumption, a model could have a better chance to achieve a skillful prediction in the NA than in the other basins, particularly in the WNP.
The applicability of using the persistent SSTA assumption on the seasonal prediction of TC in the NA can also be verified by the statistic model. It has been established in Zhao et al. (2009) that the changes in hurricane frequency in the NA are strongly correlated with the relative SSTs over the NA MDR versus the average SST of the entire tropical oceans. Vecchi and Knutson (2011) and Villarini et al. (2010) have further demonstrated this connection between NA MDR relative SSTs and NA hurricane counts from the century-long observations. In Vecchi et al. (2011), a statistical hurricane prediction method has been proposed by using the Poisson regression model described as a logarithmic link function of SSTAs in the NA MDR and in the global tropical belt region (30°S–30°N). The statistical method was built from 212-yr model integration from the C180-HiRAM, which included climate-change experiments and historical ensembles with the observed SSTs over the period 1981–2008.
In this study, we applied this statistical method to the observed and model used SSTAs to obtain hurricane counts in the NA for individual years. To distinguish from the C360-HiRAM predicted hurricanes, the numbers predicted by the statistical method are called “SSTA converted hurricane counts” hereafter. The 21-yr SSTA converted hurricane counts are shown in Fig. 10, comparing directly to the observed and model predicted NA hurricane counts. It is found that the correlation between the observed and “observed SSTA converted” hurricane counts is 0.78. The correlation is higher between the model-predicted and “model-used SSTA converted” hurricane counts (0.82). The high correlations in both situations demonstrate the skill of the statistical method in Vecchi et al. (2011). The reason for the higher correlation shown in the model group can be attributed to the fact that the training of the Poisson regression was based on the HiRAM experiments. Furthermore, it also benefited from the ensemble averaging in the model, which reduces the internal variability.
The correlation between the observed and model-used SSTA converted hurricane counts is 0.76, which is a skillful statistic model prediction in the NA. The result also confirms that the hurricane number prediction in the NA is highly related to the interannual variation of the SSTA at the early hurricane season. In other words, the assumption of the persistent SSTA is reasonable and applicable for the seasonal hurricane prediction in this basin. Furthermore, note that the correlation of 0.76 from the statistic model prediction is lower than the one from the C360-HiRAM prediction (0.88). It illustrated that the computationally inexpensive statistical method is highly skillful but a high-resolution dynamical model can perform even better for the seasonal hurricane prediction.
It should be noted that the statistical method established by Vecchi et al. (2011) is only applicable for the seasonal hurricane prediction in the North Atlantic basin. For the WNP basin, Zhan et al. (2011a) proposed that the East Indian Ocean (EIO) SSTA affects the TC genesis in the entire genesis region over the WNP and largely determines the numbers of both the total and weak TCs. Zhan et al. (2011b) further examined the involved physical mechanisms by using the International Pacific Research Center regional atmospheric model (Wang et al. 2003; Wu et al. 2012) driven by the reanalysis and the observed SSTs. In our study, the relationships between the 21-yr EIO (10°S–22.5°N, 75°–100°E) SSTA indices and the WNP storm counts are also investigated. The correlations between the observed 5-month mean SSTA indices and observed storm counts are −0.57 and −0.74 for the typhoon and TS categories, respectively. The results confirm that the EIO SSTA does show some impacts on the WNP storm genesis frequency. Moreover, the index has larger control on weaker storms (TSs) rather than on the stronger storms (typhoons). It is interesting to note that there is no significant statistical relation between the EIO indices based on the model-used SSTA and model-predicted storms in this basin. One possible reason is that the air–sea interaction is relatively more important in this region so that the persistent SSTA would be a less robust assumption (as compared with the NA basin). In other words, the atmospheric processes in the model might also have some bias in the late months of the season resulting partly from the SST bias, which could affect TC activities.
5. Analyses of storm intensity distributions
Because of the limitation of model resolution, it is a challenging task for global models to represent the full spectrum of the TC’s intensity distribution during long-term simulations or predictions. In the 50-km HiRAM used by Zhao et al. (2009), very few storms could exceed a maximum wind speed of 50 m s−1, which is the threshold of category 3 hurricane on the Saffir–Simpson scale. For the C360-HiRAM with 25-km resolution, it has been shown in CL11 that the wind–SLP relationship was well captured for the TS and hurricane categories 1 and 2, but the 10-m wind speed could only reach up to category 3 (50–57 m s−1) in the NA basin. In this study, the probability density functions of the storm intensity from the 21-yr model-predicted and observed storms in all three basins characterized by the maximum 10-m wind speed are shown in Fig. 11a. For all basins, it is not surprising that the model underpredicted major hurricanes (category 3 and above; shaded area), while the numbers of TSs and category 1 and 2 hurricanes are overpredicted. Comparing the distributions among the three basins (Figs. 11c,e,g), it is worth noting that more intense typhoons can be predicted over the WNP where the maximum wind speed can exceed 60 m s−1 (category 4). However, the model-predicted hurricanes in the NA are rather weak (few reached 55–60 m s−1). It is even weaker over ENP where there are no model-predicted hurricanes with maximum wind speed beyond 55 m s−1. The differences among the three basins could be partly explained by the fact that more intense and larger in size TCs tend to form over the WNP than over the NA or the ENP (Chavas and Emanuel 2010). The general size of the observed hurricanes over ENP is the smallest, which may provide an explanation as to why the model has the most difficulty in simulating the storm intensity there.
Because of the insufficient model resolution in most global models, a model’s ability in simulating the storm intensity distribution may be related to the observed storm size distribution. Clearly, it is more difficult for a 25-km-resolution model to resolve the eyewall structure of a compact hurricane over the ENP than that of a large (in size) typhoon over the WNP. Furthermore, it should be noted that the structure of the maximum 10-m wind (wind speed versus radius from the center) would clearly exhibit sharper variation than that of the SLP distribution. Therefore, it should be easier for a model with insufficient horizontal resolution to better represent the horizontal structure of the SLP, which tends to be a Gaussian-like smooth distribution. With this reasoning, we investigated using the minimum SLP as an indicator of storm intensity to reexamine the performance of storm intensity. The probability density function of the minimum SLP for storms in all three basins during the past 21 years is shown in Fig. 11b. As expected, the model-predicted and observed frequency distributions are well matched. Among the three basins, the NA has the best match (Fig. 11d). Over the ENP and the WNP, the model still significantly underpredicted major hurricanes (using 960 hPa as the threshold; shaded area) but the differences are dramatically reduced (cf. Figs. 11e–h). It should be noted that the number of the observations characterized by the minimum SLP is fewer than that by the maximum 10-m wind over the WNP, which is due to the missing SLP information from some storms in the IBTrACs dataset.
It is encouraging that, with the SLP as the storm intensity indicator, the HiRAM at 25-km resolution can match well the observed intensity distribution, particularly in the NA basin where the storm count prediction is also the best. The mean ratios of the major hurricane number to the TS (using 1000 hPa as the threshold) number over the NA are 0.36 and 0.32 for the observations and the model predictions, respectively. The result suggests that the C360-HiRAM can be a useful tool for studying the impacts of global warming on the intensity changes of Atlantic hurricanes, which was impossible previously for most global models.
6. Summary and conclusions
The GFDL HiRAM at 25-km resolution that is designed for both weather forecast and climate-change studies is used for retrospective seasonal predictions of TCs in the three ocean basins of the Northern Hemisphere (the NA, the ENP, and the WNP) during 1990–2010. The model configurations and forecast methodology followed CL11, which achieved a remarkable correlation of 0.96 between the observed and model-predicted hurricane counts over the NA during the past decade (2000–10). In this study, five ensemble members are initialized by nudging the C360-HiRAM from 1 June to 1 July for each season with the NCEP–GFS analysis data. As in CL11, the SST used for the subsequent 5-month forecast (July–November) is computed as the SST climatology plus the observed SSTA at each forecast’s initial time. The observed SSTA is “frozen” during the entire forecast period while the intraseasonal variation of SST is completely determined by the background month-varying SST climatology. With this approach, our forecasts are true and independent forecasts, using only information available up to forecast time and without relying on another forecast model (or data source) to provide boundary conditions during forecast period.
The GFDL vortex tracker with the warm-core criterion is adopted for tracking TCs in the model forecasts. Several criteria are applied to define a TS or a hurricane/typhoon in the model based on the wind speed and warm-core information provided by the tracker. The observed TSs are identified from the IBTrACS based on the same criteria (except the warm-core condition, which is not available). Therefore, short-lived tropical storms are consistently filtered out (as in the storm tracker used in the model predictions). The observed hurricane (typhoon) numbers are based on the tropical cyclone reports from the National Hurricane Center (Joint Typhoon Warning Center) without additional filtering.
A skillful prediction of storm counts without a corresponding skillful prediction of the large-scale environment may be a cause for concern—that right prediction may be obtained for the wrong reason. To demonstrate that is not the case, we compared the predicted 21-yr mean “seasonal climatology” with the corresponding ERA-Interim analysis. We showed that the temporal (intraseasonal) variations and horizontal structure of the African easterly jet are well captured by the HiRAM. However, relative to the ERA analysis data, the HiRAM-predicted subtropical high is weaker, while the range of the Asian summer monsoon trough is more extended to the northeast. The global pattern of the model-predicted precipitation is in fair agreement with that from the TRMM 3B43 data. The HiRAM prediction shows overall stronger precipitation than the TRMM data, especially in the ITCZ over the central to eastern Pacific and the Atlantic. It is consistent with previous studies that indicated that the TRMM’s precipitation rate is often underestimated in the heavy rainfall region.
The storm counts are shown to be most predictable over the NA, achieving a high correlation of 0.88 (0.89) between the observed and model predicted hurricane (TS) counts during the 21-yr period. Small mean RMSEs (lower than 2) are shown for both hurricane and TS categories in the NA basin. The negative mean bias for hurricanes is mainly from the underprediction in September which is also the peak month of the NA hurricane season. One possible reason is that the peak SST signal in September was not well represented in the model due to the use of the background climatology SST.
The prediction over the ENP is not as outstanding as that of the NA but can still be considered skillful. For the ENP basin, the correlations between the observed and model-predicted storm counts are 0.6 for both the hurricane and the TS categories, with negative biases from the underprediction during July to September. The TS counts are less underpredicted than for the hurricanes, indicating that it is more difficult for an ENP TS to intensify into a hurricane in our model than in nature. This viewpoint is supported by the storm intensity distribution described in section 5, which shows larger difference between the observed and model-predicted major hurricanes in the ENP than in the other two basins.
When compared with the NA and the ENP, the prediction in WNP is much less skillful. The correlations between observed and model-predicted storm counts are roughly 0.35 for both TS and typhoon categories. In CL11, for the four-month period (July–October) during 2000–10, the correlations were 0.77 and 0.6 for TSs and typhoons, respectively. The huge degradation shown in the correlation is partly due to the poor results in November (the last month) extended in this study. Furthermore, the addition of the fifth ensemble also causes some skill degradation.
It is interesting to note that the patterns of hurricane/typhoon occurrence locations are quite different among the three basins. The pattern can be described as scattering in the NA and concentrated in the other two basins, which can be seen in both observations and model forecasts. Some connections between the storm number biases of the three basins and the predicted large-scale atmospheric environments can be made by comparing the model-predicted atmospheric fields to the ERA-Interim analysis during ASO, the peak season. A broad negative (positive) SLP bias over the WNP (ENP) is consistent with the positive (negative) bias of the respective storm counts, while the differences (i.e., errors) between the model prediction and the ERA analysis are much smaller in both SLP and wind fields over the NA, which is consistent with the highly skillful TC prediction in this basin.
Close examinations of the SSTA used in the model reveal that the persistent assumption is adequate over the NA MDR and the Niño-3.4 region to the very end of the hurricane season, but the differences between the model and observed SSTAs are larger in the main TC genesis areas of the ENP and the WNP. The hurricane frequency prediction of the statistical model using the SSTA information proposed by Vecchi et al. (2011) indicates that the prediction skill of hurricanes over the NA is strongly correlated to the initial SSTA. Besides, the impacts of the local SSTA on the TC activities in the three basins are likely not the same. For example, it has been shown by Zhan et al. (2011a) that there is a strong nonlocal negative correlation between the East Indian Ocean SSTA and TCs activity in the WNP basin. However, no significant statistical relation of the EIO indices was found in the model results. One possible reason is that the air–sea interaction is relatively more important over the WNP as compared to the NA. In other words, the atmospheric processes in the model might also have some bias in the later season resulting partly from the larger SST bias, which could affect TC activities. Furthermore, the excluded oceanic feedback from the using of the prescribed “ocean” in the atmosphere-only model may affect the TC’s seasonal prediction in the WNP more than that in the NA.
Because of the limitation of model resolution, it is a challenging task for a global weather or climate model to represent the full spectrum of TC’s intensity distribution. From the intensity analyses based on the 10-m maximum wind speed, it is not surprising to see that our 25-km-resolution model underpredicted major hurricanes and overpredicted TSs and category 1 and 2 hurricanes. We also found basin-specific differences that more intense typhoons can be formed over the WNP, with weaker and the weakest hurricanes over the NA and the ENP, respectively. Given the insufficient model resolution, this behavior can be related to the different storm size distributions among the basins. For a rather compact hurricane typically found over the ENP, it would naturally be more difficult for the model to resolve the wind structure than to resolve that of a larger typhoon over the WNP. Based on this reasoning, we also used the minimum central SLP to reexamine the storm intensity distribution as predicted by the C360-HiRAM. A very encouraging finding is that major hurricanes/typhoons are well represented if the minimum SLP is used instead of the usual maximum 10-m wind speed. The representation of the SLP-based intensity distribution is better over the NA than the other two basins. This finding demonstrates the potential application of using the HiRAM at 25-km resolution to study the impacts of global warming on the changes of Atlantic hurricane frequency as well as intensity, without resorting to the usual downscaling strategy using an independent regional model (Bender et al. 2010).
We thank Tim Marchok for his assistance on the use of the tracker program he developed, and Zhitao Yu for his assistance to process the storm filter program. We also thank Yuqing Wang for the discussion about the seasonal TC activities over the western Pacific Ocean. Funding from NOAA’s Hurricane Forecast Improvement Project made this project possible.
Current affiliation: Naval Research Laboratory, Monterey, California.