1. Introduction
Tropical cyclones (TCs) over the North Atlantic (NA) have a major impact on the economy and environment in North America. Smith and Katz (2013) reported that the total estimated damages caused by TCs in the United States was about $418 billion in 1980–2011, about 47% of the total losses from natural disasters. Thus, a skillful seasonal prediction of TC activities is crucial for the preparation and reduction of TC-related losses in the North American coastal region.
A number of empirical models have been developed for the seasonal predictions of TC activities over the NA basin based on their time-lagged relationships with precursory environmental conditions (e.g., Gray et al. 1992; Elsner and Schmertmann 1993; Hess et al. 1995; Lehmiller et al. 1997; Blake and Gray 2004; Saunders and Lea 2005; Elsner and Jagger 2006; Klotzbach 2007; LaRow et al. 2010). Some of these models have shown useful skill and have been employed in operational TC predictions. However, the statistical relationships are at times difficult to interpret physically because of their time-lagged properties, which may eventually result in poor forecasting skill in operational TC predictions. In addition to statistical forecasts, high-resolution dynamic models have also been used for seasonal TC prediction by incorporating the TC detection algorithm based on key characteristics of TCs in model simulations (e.g., Knutson et al. 2007; Camargo and Barnston 2009; Zhao et al. 2010; Chen and Lin 2011). Despite demanding substantial computational resources, operational dynamic seasonal TC forecasts are experimentally used in various modeling centers (LaRow et al. 2010; Vecchi et al. 2014; Camp et al. 2015) with a significant improvement from just a few years ago (Camargo et al. 2007a).
As an alternative to these statistical or dynamical TC forecasts, several recent studies have introduced hybrid statistical–dynamical approaches for TC prediction over key ocean basins (e.g., Wang et al. 2009; Kim and Webster 2010; Vecchi et al. 2011; H.-S. Kim et al. 2012; Li et al. 2013). These hybrid methods utilize the simultaneous relationship between TC activities (i.e., predictand) and large-scale environmental conditions (i.e., predictors) forecasted by dynamical models to improve upon the traditional pure statistical forecasts. These hybrid forecasts overcome the weaknesses inherent in the statistical methods and also preserve the physical and direct connections between summertime TCs and the environmental conditions (H.-S. Kim et al. 2012). Furthermore, these approaches have operational advantages, because TC predictions can be updated according to forecasted atmospheric and oceanic conditions from coupled atmosphere–ocean climate models.
Most seasonal TC prediction approaches target the total number of TCs over the entire basin. However, the impact of TCs on human society are mainly connected to their landfalls rather than the total basinwide TC counts (Pielke and Landsea 1998; Pielke et al. 2008; Weinkle et al. 2012). Recognizing the practical importance of TC pathways, several recent studies attempted to predict the TC track density by a track-oriented-pattern categorization approach. Chu et al. (2010) and Chu and Zhao (2011) introduced this concept. They argued that this approach can eventually, in principle, lead to better performance via more detailed physical links with individual TC track patterns. For the western North Pacific (WNP) basin, a cost-effective probabilistic TC track density prediction model was developed on the basis of seven different TC track patterns (e.g., H.-S. Kim et al. 2012; J.-H. Kim et al. 2012; Ho et al. 2013). The National Typhoon Center of the Korea Meteorological Administration currently employs this model for the seasonal outlook of TC activities over the WNP basin. Similarly, classifications and predictions of the TC track patterns over the NA as well as their frequencies have been reported in several studies (Hall and Jewson 2007, 2008; Kossin et al. 2010; Kozar et al. 2012). However, comprehensive seasonal predictions of the spatial TC track pattern over the entire NA have not been reported thus far.
The objectives of this study are to develop a seasonal TC prediction model for the NA basin on the basis of a hybrid statistical–dynamical approach and to examine its skill. A number of previous studies have applied clustering techniques to TC tracks; such studies have shown that clustering analysis can be used to divide the overlapped effects of climate systems into individual categories of TC track patterns (Elsner 2003; Kossin et al. 2010). On the basis of these properties, we have developed statistical prediction models for each track pattern, and we merged the predictions for individual track patterns into a seasonal TC track density forecast for the entire basin. By comparing the model results to observations, we can evaluate the model’s forecast skill for the entire NA basin and subbasins. Specifically, this study targets the prediction of TCs that affect the U.S. East Coast, the Gulf of Mexico, and the Caribbean Sea, all of which are severely affected by landfalling TCs.
This paper is organized as follows. The datasets used in this study are described in section 2. Section 3 presents detailed procedures for developing the seasonal TC prediction model; verification of model predictability is presented in section 4. Section 5 shows the prediction of regional TC activities in NA subbasins. Conclusions from this study are in section 6.
2. Data
The data for TC activities over the NA basin during 1965–2012 are obtained from the hurricane database (HURDAT) at the National Oceanic and Atmospheric Administration (NOAA) National Hurricane Center, which posts the location and intensity of all NA TCs at 6-h intervals (McAdie et al. 2009). The HURDAT best-track data are available from 1851; however, this study has analyzed only the TCs from 1965, when satellite observations became available, to 2012 (Chu 2002). We focus on the TC activities from August through October (ASO), because the number of NA TCs in this 3-month period is relevant to about 80% of climatological mean annual NA TC counts. Only the cyclones with maximum sustained wind speeds greater than 17 m s−1 are defined as TCs.
To investigate the effects of the large-scale environment on TC activities, atmospheric circulation data are obtained from the National Centers for Environmental Prediction (NCEP) Reanalysis-2 dataset (R-2; Kanamitsu et al. 2002), which has a horizontal resolution of 2.5° × 2.5° in latitude and longitude. We analyzed the vertical wind shear (VWS) defined as the zonal wind difference between 200 and 850 hPa in addition to the zonal wind and relative vorticity at the 850-hPa level (U850 and VOR850, respectively). The monthly sea surface temperature (SST) data are obtained from the NOAA Extended Reconstructed SST, version 3 (ERSST.v3; Smith et al. 2008). ERSST.v3 data have a 2° × 2° resolution in latitude and longitude. The reanalysis data and SST during 1982–2012 are used for consistency with the dynamic seasonal forecast described below.
For the dynamical component of this hybrid statistical–dynamical model, we adopt the NCEP Climate Forecast System, version 2 (CFSv2), a fully coupled global atmosphere–ocean–land modeling system. NCEP CFSv2 was updated in March 2011 from the earlier CFS version 1 model and has since been used for operational seasonal climate forecasting (Saha et al. 2014). In this study, the monthly NCEP CFSv2 retrospective forecasts at a 1° × 1° resolution in latitude and longitude for the period 1982–2012 is used. The retrospective data are reforecasted with the CFS Reanalysis as an initial condition, which is utilized to construct a simultaneous statistical relationship between TC activities and environmental fields for each TC cluster. The CFSv2 generates 9-month forecasts consisting of four ensembles per day with different initial conditions at 0000, 0600, 1200, and 1800 UTC. The retrospective data include 9-month forecasts issued every five days, beginning from 1 January of every year; thus, 24 ensemble members are generally included for each month except November, which has 28 members. To forecast the TC activity during the ASO, we use 12 ensemble members of NCEP CFSv2 forecasts representing the number of ensembles for half a month. These ensemble members are issued on three consecutive days, including two 5-day periods 10 and 5 days prior to the forecast in addition to the original forecast day for our TC prediction model (e.g., 25 and 30 June and 5 July for the case of a 5 July TC forecast). Additionally, we verified the model performance against observations every month by changing the forecast day in early February (5 February) to that in early July (5 July).
3. Development of seasonal TC forecast model
a. Pattern classification
The clustering results must be carefully examined, as they vary according to the number of clusters (Camargo et al. 2007b; Kim et al. 2011) and will eventually affect the entire model development. To objectively determine the optimal number of clusters in FCM, we examined the sensitivity of the four scalar indices, the partition coefficient (Bezdek 1981), the partition index (Bensaid et al. 1996), the separation index (Xie and Beni 1991), and the alternative Dunn index (Dunn 1973) to the number of clusters. The definitions and properties of these indices are described in Dunn (1973), Bezdek (1981), Xie and Beni (1991), Bensaid et al. (1996), and Kim et al. (2011). Larger (smaller) values of partition coefficient (partition index, separation index, and alternative Dunn index) indicate that corresponding cluster number is more optimal in FCM (Kim et al. 2011). The comprehensive optimum cluster number detection process using these four indices showed that four TC track patterns can adequately represent the NA TC track properties during the TC season (not shown).
Figure 1 displays the four TC track patterns and the entire 432 TC tracks over the NA during ASO from 1965 to 2012. The TCs of cluster 1 (C1 pattern) generally originate off the North American east coast (subtropical western NA) and propagate northeastward along the U.S. East Coast (Fig. 1a). These TCs often affect and make a landfall in the eastern U.S. region. The C2 pattern is characterized by tracks entering the Gulf of Mexico and the Caribbean Sea (Fig. 1b). These TCs generally form near Cuba and Haiti and move northwestward to the Gulf of Mexico. The devastating Hurricane Katrina in 2005 belongs to this cluster. The TCs in the C3 pattern originate over the vast open ocean of the NA and move to mid- and high-latitude regions with recurving tracks (Fig. 1c). Although this type of TC has the longest life-span and the strongest intensity, its damages are usually small because its track remains mostly over the ocean. The C4-pattern TCs are generated near the tropical region of the NA and move westward, affecting maritime islands (Fig. 1d). All together, the total TC track density of the NA clearly depicts two major paths of TCs. The first is a straight northwestward ridge to the Gulf of Mexico in the tropical regions (C2 and C4), and the second is a northeastward recurving ridge at midlatitude (Fig. 1e). The numbers (percentages) of C1, C2, C3, and C4 TCs are 112 (25.9%), 123 (28.5%), 98 (22.7%), and 99 (22.9%) of the total 432 (100%) TCs, respectively.
(a)–(d) Four track patterns of NA TCs during the period 1965–2012 TC season and (e) total tracks. Contours represent climatological track densities; the interval is 10, except in (e) it is 5. Black dots indicate the genesis position of each TC, and gray lines show individual TC tracks. The number of TCs for each pattern is shown in parenthesis.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
(a)–(d) Four track patterns of NA TCs during the period 1965–2012 TC season and (e) total tracks. Contours represent climatological track densities; the interval is 10, except in (e) it is 5. Black dots indicate the genesis position of each TC, and gray lines show individual TC tracks. The number of TCs for each pattern is shown in parenthesis.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
(a)–(d) Four track patterns of NA TCs during the period 1965–2012 TC season and (e) total tracks. Contours represent climatological track densities; the interval is 10, except in (e) it is 5. Black dots indicate the genesis position of each TC, and gray lines show individual TC tracks. The number of TCs for each pattern is shown in parenthesis.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
Our classification of the NA TC track patterns is highly consistent with previous studies. Kossin et al. (2010) and Kozar et al. (2012) suggested that the climatological TC tracks in the NA basin can be classified into four groups. Although the method and the period in their study are different from those in this study, their results are similar to those in the present study. Nakamura et al. (2009) clustered the TC tracks into six groups by using the k-means method with mass moments. In that study, TCs recurving to midlatitude were classified into four clusters. In the present study, however, such TCs were grouped in two clusters (i.e., C1 and C3 patterns). Classification of the TC tracks into more groups may be helpful for identifying the separated linkage with diverse climate variability. Nevertheless, more clusters lead to smaller numbers of TCs assigned to each cluster, which in turn makes it difficult to construct reliable statistical forecast models. Given the objective of this study, seasonal prediction of NA TC activities, we sorted the NA TC tracks into four patterns, which are the same as the aforementioned result of optimum cluster number detection process.
b. Prediction model for each cluster
Construction of the hybrid statistical–dynamical TC prediction model for each track pattern is a core step in this development. Prior to developing the model, we first examined the empirical relationships between TC activity and large-scale environmental fields for each track pattern. Candidate variables in constructing the model are decided as ASO-averaged SST, VWS, VOR850, and U850, all of which are well-known large-scale factors that influence the NA TC activities (Gray et al. 1992; Blake and Gray 2004; Saunders and Lea 2005; Klotzbach 2007). Other factors, such as moist stability and midlevel relative humidity, also affect TC activity (Gray 1998); however, these are not regarded as suitable for predictors because of the lack of reliability in today’s climate model simulations (John and Soden 2007; Saha et al. 2014).
Figure 2 shows the temporal correlations between the TC frequency and the selected atmospheric and oceanic variables for each cluster based on the ensemble mean of the NCEP CFSv2 retrospective issued on 5 July. The critical domains (rectangular boxes in Fig. 2) are identified as those when both correlations of the NCEP R-2 (not shown) and CFSv2 retrospective data with TC activity equal or exceed the 90% confidence level. The NCEP CFSv2 closely reproduces the observed relationships between TC frequency and large-scale atmospheric and oceanic forcings with high statistical significance in the critical domains. It is also required that the selected domains with variables should be physically related to the corresponding TC track patterns. Predictors are computed by obtaining the area averages of gridpoint values in the critical domains that meet the threshold of statistical significance at the 90% confidence level. The relative importance of individual predictors and their combinations for each track pattern can then be determined in sensitivity tests. As long as the NCEP CFSv2 effectively simulates the TC–environmental fields’ relationships, we can anticipate that the hybrid-type model will, in principle, provide credible predictions.
Distribution of correlation coefficients between observed C1–C4 TC frequencies and the ensemble average of CFSv2 retrospectives initialized on 5 Jul for each predictor. The contour interval is 0.2; the zero contour line is omitted. Shading indicates areas statistically significant at the 90% confidence level. Critical regions are presented as a rectangular box in each panel.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
Distribution of correlation coefficients between observed C1–C4 TC frequencies and the ensemble average of CFSv2 retrospectives initialized on 5 Jul for each predictor. The contour interval is 0.2; the zero contour line is omitted. Shading indicates areas statistically significant at the 90% confidence level. Critical regions are presented as a rectangular box in each panel.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
Distribution of correlation coefficients between observed C1–C4 TC frequencies and the ensemble average of CFSv2 retrospectives initialized on 5 Jul for each predictor. The contour interval is 0.2; the zero contour line is omitted. Shading indicates areas statistically significant at the 90% confidence level. Critical regions are presented as a rectangular box in each panel.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
The VWS is negatively correlated with the C1 TC frequency in the midlatitude region (Fig. 2a). It is noted that small VWS (i.e., weak baroclinity) is a key for TC formation as well as for maintaining TC activities. Based on this, VWS is included as a predictor since it is expected to affect TCs especially in the midlatitudes where strong westerlies in the upper troposphere can negatively affect TCs. Some regions, such as the equatorial eastern Pacific, show positive correlations between TCs and VWS, which can be recognized as a well-known anticorrelation of the North Atlantic TC activity and eastern Pacific VWS (Frank and Young 2007). However, this significant anticorrelation pattern is not well shown in the correlation map of the CFSv2 retrospectives because the number of grid points showing significant correlation with the C1-pattern TCs is too small (Fig. 2a). Thus, we exclude such regions. The C1-pattern TCs are also positively correlated with local VOR850 over the eastern U.S. coastal region, where this type of TC is dominant (Fig. 2b).
If a TC season is in a La Niña phase in the equatorial Pacific or if positive SST anomalies appear over low-latitude regions of the NA, C2-pattern TCs (i.e., TCs in the Gulf of Mexico) are found to be more active (Goldenberg et al. 2001; Holland and Webster 2007; Kossin and Vimont 2007; Vecchi and Soden 2007). Both the remote influence of the negative anomalous SST values in the eastern Pacific and the local positive SST effect in the NA induce a rising motion over the Gulf of Mexico. Because negatively correlated SST regions statistically significant at the 90% level are not reached in the CFSv2 correlation map, only positively correlated SST patterns in the low-latitude regions of the NA are used as a predictor (Fig. 2c). This physical mechanism also results in triggering positive vorticity anomalies and creates favorable conditions for TC genesis in the off-equatorial region (Fig. 2d). The U850 shows a significant positive relationship with C2-pattern TCs over the tropics; thus, U850 is selected as a predictor (Fig. 2e). Equatorial low-level zonal wind is related to the Madden–Julian oscillation (MJO), which has been shown to modulate TC activity in the Gulf of Mexico and the northwestern Caribbean Sea (Klotzbach 2010). In fact, the MJO itself may not be suitable for seasonal prediction because its time scale is subseasonal (i.e., about 30–60 days). Nevertheless, the dynamical mechanism of U850 for affecting C2 TC activity analogous to the MJO effect is still valid even in a seasonal time scale.
The prediction model of C3-pattern TCs uses SST, VOR850, and U850 as predictors. C3-pattern TCs activities show positive correlations with basinwide SST in the NA (Fig. 2f), which is a pattern similar to the Atlantic multidecadal oscillation. Along with that, C3-pattern TC frequency is positively correlated with VOR850 (Fig. 2g). Both create well-known favorable conditions for TC development and were thus selected as predictors. The U850 over the tropical NA also positively correlates to C3-pattern TCs (Fig. 2h). When the strong eastward zonal wind is dominant over the tropical NA, TCs (here, C3-pattern TCs) may be steered to recurve without making a landfall on the North American coast. For C4, we invite the positive relationships of midlatitude VWS as a predictor (Fig. 2i). Because the strengthening of midlatitude VWS can weaken TC activity, TCs usually have tracks that are active only in low-latitude regions (i.e., C4-pattern TCs). Moreover, the anomalous positive vorticity in the tropics apparently can contribute to C4 TC genesis (Fig. 2j), similar to that in the other cases. These results imply that, if favorable conditions of low-level vorticity are presented during the ASO season, the NA TCs are more likely to develop.
c. Construction of final forecasting map
4. Validation
Examination of the model skill employs a leave-one-out cross-validation method, which is widely used to assess the performance of statistical prediction (Gray et al. 1992; Elsner and Schmertmann 1993, 1994; Chu et al. 2007; H.-S. Kim et al. 2012; Ho et al. 2013). Specifically, when predictions are performed for the training period, the model is iteratively adjusted for all available retrospective forecasts data except the target year, assuming the predictors are independent on different years. This method allows us to examine whether the model fitting process is optimized to make prediction results as far as realistically possible.
Figure 3 shows the ensemble-mean hindcasts for the TC frequency for individual clusters (Figs. 3a–d) as well as the total TC numbers (Fig. 3e) initialized in early February to early July on a monthly interval. The 12 members of the NCEP CFSv2 retrospective forecasts are used to calculate the ensemble means. Although individual forecasts oscillate with relatively large variances (not shown), the ensemble mean for constructing the final forecast agrees well with observations (black lines in Figs. 3a–e) regardless of the forecast lead time. These are consistent with those reported by Kwon et al. (2007), who showed that multimember ensemble means generally outperform individual members in statistical predictions of TC activity. In addition, to objectively evaluate the model performance compared to the reference forecast, reforecasts from the model based on NCEP R-2 instead of CFSv2 retrospectives are overlapped (black dashed lines in Figs. 3a–e). In this comparison, the reforecast using the reanalysis data (i.e., NCEP R-2 in this study) can be regarded as the reference forecast. The reforecasts using the NCEP R-2 compare well with observations with no forecast biases. As shown in the hindcasts from the CFSv2 retrospectives and reforecasts using NCEP R-2 data, the overall climatology and variability of TC activities are well represented in the model despite some forecast uncertainties due to the errors in CFSv2 seasonal forecasts. Thus, we conclude that the ensemble-mean hindcasts from CFSv2 are reliable.
Time series of TC frequency from observations (black solid line), from reforecasts using the NCEP R-2 data (black dashed line), and from the ensemble mean of the model hindcast driven by the CFSv2 retrospective run (gray solid line) for the period 1982–2012.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
Time series of TC frequency from observations (black solid line), from reforecasts using the NCEP R-2 data (black dashed line), and from the ensemble mean of the model hindcast driven by the CFSv2 retrospective run (gray solid line) for the period 1982–2012.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
Time series of TC frequency from observations (black solid line), from reforecasts using the NCEP R-2 data (black dashed line), and from the ensemble mean of the model hindcast driven by the CFSv2 retrospective run (gray solid line) for the period 1982–2012.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
It is noteworthy that the hindcast of CFSv2 effectively resolves the interannual variability and most of the observed extreme TC activity seasons (e.g., 2000 in C1, 2005 in C2, and 1995 in C3), suggesting that our model is skillful also in predicting abnormal TC activities. Note that the predicted frequency does not reach the zero TC count because the Poisson regression model cannot generate zero (e.g., 1994 and 1995 in C1, 2008 in C3, and 1985 in C4). One notable forecast error is that the hindcast substantially overestimates the observed TC frequency in 2010, particularly for clusters C2, C3, and C4 (Figs. 3b–d). This discrepancy is mainly caused by large-scale positive VOR850 anomalies in the CFSv2 retrospective forecast in 2010. The CFSv2 retrospective shows broad positive VOR850 anomalies over the low-latitude NA region concentrated over the Gulf of Mexico (not shown), which were subsequently adopted as predictors for C2–C4 patterns (Figs. 2d,g,j).
Correlation coefficients, root-mean-square errors, mean square skill scores, and Gerrity skill scores of ensemble-averaged hindcasts from CFSv2 retrospectives in six issue days and the reforecast using the NCEP R-2 with the observation for the period 1982–2012.
As well as the TC number prediction, it is also worth examining the forecast results in terms of categories compared to its climatology: below normal (BN), normal (N), and above normal (AN). When the number of TCs is 0.5 standard deviations above (below) the average TC frequency, that year is assigned to the AN (BN) category. Understandably, the other years are defined as category N. The categorized results of observation and ensemble-averaged hindcasts are shown in contingency tables for each cluster and total TC number (Table 2). When the diagonal components of each contingency matrix have nonzero values, both observations and predictions are in the same category, this indicates there was a successful prediction. The other components can be interpretable as forecast failures. Table 2 shows that the C1 model correctly predicts 21 of the 31 years (1982–2012), which is about a 68% success rate. Similarly, the success rates of the C2, C3, C4, and total TC models are 65%, 74%, 71%, and 74%, respectively.
Contingency tables between the observed TC activity and hindcast results issued on 5 Jul for the period 1982–2012.
In summary, our model is skillful with various forecast lead times. Some of the long-lead predictions (e.g., 5–6-month lead forecast) show slightly better skill than the short-lead-time predictions (e.g., 1–2-month lead), although the predictability of CFSv2 normally decreases as the lead time increases. This is expected, as all prediction models of multiple leads are optimized for their own training period. As mentioned earlier, the hybrid statistical–dynamical model picks statistically significant grids for obtaining predictors. Although the number of significant grids for longer-lead-time prediction decreases, the skillful predictors from selected variable sets and grids maintain the high prediction skill of the model. Thus, stable skill with varying forecast lead times is a unique virtue of the hybrid statistical–dynamical model for predicting ASO TC activities, which differs from that in previous traditional statistical models (e.g., Gray et al. 1992; Elsner and Schmertmann 1993; Hess et al. 1995; Lehmiller et al. 1997; Blake and Gray 2004; Saunders and Lea 2005; Elsner and Jagger 2006; Klotzbach 2007; LaRow et al. 2010).
5. Regional distribution of forecasted TCs
Our model is used to predict seasonal TC track patterns for the entire NA basin. To assess the forecast skill in generating spatial TC distributions during the training period, we analyzed the rank correlation for each 5° × 5° grid point in the NA basin for different forecast lead months (Fig. 4). Because the observed and forecasted TC frequencies do not follow a Gaussian distribution at each grid point, we use rank correlation as a statistical measure for the model skill in order to consider prediction model errors (Vecchi et al. 2014). By using rank correlation, which represents the degree of similarity between two rankings, we can evaluate the significance of the relationship between them. Strong rank correlations (>0.5) for all forecast lead months appeared in regions with high TC frequency, suggesting the model is more skillful in regions with high TC activity. Most of these high predictability grids are located in regions in which the four TC track patterns overlap. Although these results imply that the seasonal prediction results represented as a TC track density distribution may vary with time, the rank-correlation patterns are largely invariant of the forecast lead time.
Spatial distribution of retrospective rank correlation between the observed TC passages and ensemble average of hindcasts results in a 5° × 5° grid box. The contour interval is 0.25; shading indicates areas with rank correlation greater than 0.5. Multiple forecasts initialized in (a) 5 Feb, (b) 2 Mar, (c) 1 Apr, (d) 1 May, (e) 5 Jun, and (f) 5 Jul. Three vulnerable TC-influenced domains, defined as R1, R2, and R3 regions, are also shown.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
Spatial distribution of retrospective rank correlation between the observed TC passages and ensemble average of hindcasts results in a 5° × 5° grid box. The contour interval is 0.25; shading indicates areas with rank correlation greater than 0.5. Multiple forecasts initialized in (a) 5 Feb, (b) 2 Mar, (c) 1 Apr, (d) 1 May, (e) 5 Jun, and (f) 5 Jul. Three vulnerable TC-influenced domains, defined as R1, R2, and R3 regions, are also shown.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
Spatial distribution of retrospective rank correlation between the observed TC passages and ensemble average of hindcasts results in a 5° × 5° grid box. The contour interval is 0.25; shading indicates areas with rank correlation greater than 0.5. Multiple forecasts initialized in (a) 5 Feb, (b) 2 Mar, (c) 1 Apr, (d) 1 May, (e) 5 Jun, and (f) 5 Jul. Three vulnerable TC-influenced domains, defined as R1, R2, and R3 regions, are also shown.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
The direct impact of TCs to human society is closely linked to TC landfall accompanied by strong wind gusts and heavy rainfall; therefore, coastal regions are more vulnerable to TC effects. To examine the model skill for these vulnerable coastal regions, we focused on three major regions particularly vulnerable to TCs (Fig. 4a). The first region (R1) includes the U.S. East Coast area, where population and economic activities are heavily concentrated. The second region (R2) covers the Gulf of Mexico, the western Caribbean Sea, and their neighboring countries. The third region (R3) includes the eastern Caribbean Sea, a main region of TC genesis and their pathways. Spatially, these regions show substantial rank correlation, indicating high forecast skill (Fig. 4). By assessing TC distribution over these three vulnerable regions, we can interpret the results of TC activity prediction at mesoscales as well as a larger-scale view of the entire basinwide map.
Figure 5 depicts the temporal variations of the observed TC frequency and the ensemble-mean hindcast for six different lead times in the three regions. To compare observations and hindcasts consistently, both of parameters are converted into track densities in 5° × 5° grid boxes covering the entire NA basin. The area averages for the three regions are then evaluated. The three time series show no climatological forecast biases for all regions, which is consistent with the aforementioned time series of each cluster. Although there are some forecast failures, such as overestimation for R2 in 2010, the correlations of different lead times are all above 0.70 for the three vulnerable regions, which is statistically significant at the 99% confidence level. In addition to the interannual variability of regional TC activities, the model also effectively captured the extreme TC activity years (e.g., 2005 in R2 and 1995 in R3; Fig. 4). Therefore, our model is also reliable and skillful for predicting subbasin coastal TC activity.
Time series of regional averaged TC passages in (a) R1, (b) R2, and (c) R3 regions. The black line indicates observation, and gray lines show ensemble-averaged values of CFSv2 retrospectives for six forecast days, including 5 Feb, 2 Mar, 1 Apr, 1 May, 5 Jun, and 5 Jul for the period 1982–2012.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
Time series of regional averaged TC passages in (a) R1, (b) R2, and (c) R3 regions. The black line indicates observation, and gray lines show ensemble-averaged values of CFSv2 retrospectives for six forecast days, including 5 Feb, 2 Mar, 1 Apr, 1 May, 5 Jun, and 5 Jul for the period 1982–2012.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
Time series of regional averaged TC passages in (a) R1, (b) R2, and (c) R3 regions. The black line indicates observation, and gray lines show ensemble-averaged values of CFSv2 retrospectives for six forecast days, including 5 Feb, 2 Mar, 1 Apr, 1 May, 5 Jun, and 5 Jul for the period 1982–2012.
Citation: Journal of Climate 29, 2; 10.1175/JCLI-D-15-0407.1
6. Conclusions
We have developed a seasonal prediction model based on four TC track patterns of ASO TC activities over the NA basin and evaluated its forecasting skill for various lead months from February to July. Unlike previous studies focused on the seasonal forecast of the total number of TCs in the NA basin, our model can also predict the basinwide spatial distributions of TC activities in addition to the total number of TCs. Because TC-related damages are more closely related to the TCs’ proximity to coastal areas or landfall rather than the total TC genesis number in the entire NA basin, this study can contribute directly to TC-impact preparedness, which can significantly reduce damages to life and property in TC-prone coastal areas.
TCs in the NA basin are objectively classified into four TC track patterns by using the FCM, in which each pattern has its own unique track characteristics. The C1-pattern TCs pass along the U.S. East Coast, and C2-pattern TCs develop over the Gulf of Mexico and the Caribbean Sea and remain in these areas for their entire lifetimes. C3-pattern TCs are generated in the subtropical NA, and C4-pattern TCs develop in the open ocean of the equatorial-central NA. C3 TCs move to the midlatitude NA with recurving pathways, whereas C4 TCs are confined to low-latitude regions and mainly move northwestward toward the islands in the Caribbean. After classifying the four track patterns, the prediction model for an individual pattern is constructed. Identifying the simultaneous relationships between each TC pattern and climate variables enables us to select appropriate predictors of model for each pattern. Combinations of candidate predictors (e.g., SST, VWS, VOR850, and U850) well known for affecting TC activities are determined by conducting several cross-validation tests with various sets of candidate predictors to yield the best predictability for each pattern. A hybrid dynamical–statistical model is developed by using CFSv2 retrospective forecasts and TC frequencies in each cluster based on the Poisson regression. To verify the performances of the forecasts initialized in six consecutive months from early February to early July, we conducted leave-one-out cross validation for all forecasts. Validation of the TC frequency of individual clusters and the total counts in the NA basin suggests high prediction skill at the 99% confidence level regardless of forecast lead time. Our model also shows better predictability than the reforecast based on the NCEP R-2 predictors. In addition, the spatial distributions of rank correlation for different forecast lead times are calculated to investigate predictability by regional groups. Because substantially high values of rank correlation occur in a wide region, we can anticipate reliable regional TC activity prediction. To investigate regional TC activity more concretely, the model performance is evaluated for the three TC-vulnerable regions (i.e., R1, R2, and R3 in Fig. 4). Temporal variations of observed TC activities and the ensemble-mean hindcasts for the three vulnerable regions confirm that the model is skillful in predicting regional TC activities.
The advantage of our model is that it can predict spatial patterns of the TC activity in the entire NA basin with minimal computational costs compared to using fine-resolution dynamical models, either regional or global. Recently, in light of this advantage, a track pattern–based model for the WNP basin has been employed for quasi–real-time operational forecasting by the National Typhoon Center of the Korea Meteorological Administration (J.-H. Kim et al. 2012; Ho et al. 2013). This implementation of a WNP TC prediction model by a meteorological agency and its satisfactory operation until now could partly dispel doubts about artificial skill, as argued in DelSole and Shukla (2009), of our methodology for application to real-time forecasts. Moreover, the high forecast skill of our model shown in this study also supports the applicability of the hybrid model to operational seasonal TC prediction for the NA basin in conjunction with the seasonal forecast dataset from the CFSv2 as predictors.
Two major challenges have been identified in our model. First, unlike dynamical models, it does not forecast individual TCs, because our model is based on statistical regression. Although there is physical consistency between TC genesis and favorable environmental conditions, such relationships are not always linked directly to TC formation. Second, TCs with irregular tracks showing a significant amount of fuzziness cannot be easily assigned to one of the four defined track patterns, which may lead to larger forecasting error. Improvements to resolve the limitations in our model will be performed in future research.
Acknowledgments
This study was funded by the Korea Ministry of Environment as the “Climate Change Correspondence Program.” Mr. W. Choi was supported by the BK21 project of the Korean government. We acknowledge the critical comments from two anonymous reviewers.
REFERENCES
Bensaid, A. M., L. O. Hall, J. C. Bezdek, L. P. Clarke, M. L. Silbiger, J. A. Arrington, and R. F. Murtagh, 1996: Validity-guided (re)clustering with applications to image segmentation. IEEE Trans. Fuzzy Syst., 4, 112–123, doi:10.1109/91.493905.
Bezdek, J. C., 1981: Pattern Recognition with Fuzzy Objective Function Algorithms.Kluwer Academic, 256 pp.
Blake, E. S., and W. M. Gray, 2004: Prediction of August Atlantic basin hurricane activity. Wea. Forecasting, 19, 1044–1060, doi:10.1175/814.1.
Camargo, S. J., and A. G. Barnston, 2009: Experimental dynamical seasonal forecasts of tropical cyclone activity at IRI. Wea. Forecasting, 24, 472–491, doi:10.1175/2008WAF2007099.1.
Camargo, S. J., A. G. Barnston, P. Klotzbach, and C. W. Landsea, 2007a: Seasonal tropical cyclone forecasts. WMO Bull., 56, 297–309.
Camargo, S. J., A. W. Robertson, S. J. Gaffney, P. Smyth, and M. Ghil, 2007b: Cluster analysis of typhoon tracks. Part I: General properties. J. Climate, 20, 3635–3653, doi:10.1175/JCLI4188.1.
Camp, J., M. Roberts, C. MacLachlan, E. Wallace, L. Hermanson, A. Brookshaw, A. Arribas, and A. A. Scaife, 2015: Seasonal forecasting of tropical storms using the Met Office GloSea5 seasonal forecast system. Quart. J. Roy. Meteor. Soc., 141, 2206–2219, doi:10.1002/qj.2516.
Chen, J.-H., and S.-J. Lin, 2011: The remarkable predictability of inter-annual variability of Atlantic hurricanes during the past decade. Geophys. Res. Lett., 38, L11804, doi:10.1029/2011GL047629.
Chu, P.-S., 2002: Large-scale circulation features associated with decadal variations of tropical cyclone activity over the central North Pacific. J. Climate, 15, 2678–2689, doi:10.1175/1520-0442(2002)015<2678:LSCFAW>2.0.CO;2.
Chu, P.-S., and X. Zhao, 2007: A Bayesian regression approach for predicting seasonal tropical cyclone activity over the central North Pacific. J. Climate, 20, 4002–4013, doi:10.1175/JCLI4214.1.
Chu, P.-S., and X. Zhao, 2011: Bayesian analysis for extreme climatic events: A review. Atmos. Res., 102, 243–262, doi:10.1016/j.atmosres.2011.07.001.
Chu, P.-S., X. Zhao, C.-T. Lee, and M.-M. Lu, 2007: Climate prediction of tropical cyclone activity in the vicinity of Taiwan using the multivariate least absolute deviation regression method. Terr. Atmos. Oceanic Sci., 18, 805–825, doi:10.3319/TAO.2007.18.4.805(A).
Chu, P.-S., X. Zhao, C. Ho, H.-S. Kim, M.-M. Lu, and J.-H. Kim, 2010: Bayesian forecasting of seasonal typhoon activity: A track-pattern-oriented categorization approach. J. Climate, 23, 6654–6668, doi:10.1175/2010JCLI3710.1.
DelSole, T., and J. Shukla, 2009: Artificial skill due to predictor screening. J. Climate, 22, 331–345, doi:10.1175/2008JCLI2414.1.
Dunn, J. C., 1973: A fuzzy relative of the ISODATA process and its use in detecting compact well-separated clusters. J. Cybern., 3, 32–57, doi:10.1080/01969727308546046.
Elsner, J. B., 2003: Tracking hurricanes. Bull. Amer. Meteor. Soc., 84, 353–356, doi:10.1175/BAMS-84-3-353.
Elsner, J. B., and C. P. Schmertmann, 1993: Improving extended-range seasonal predictions of intense Atlantic hurricane activity. Wea. Forecasting, 8, 345–351, doi:10.1175/1520-0434(1993)008<0345:IERSPO>2.0.CO;2.
Elsner, J. B., and C. P. Schmertmann, 1994: Assessing forecast skill through cross validation. Wea. Forecasting, 9, 619–624, doi:10.1175/1520-0434(1994)009<0619:AFSTCV>2.0.CO;2.
Elsner, J. B., and T. H. Jagger, 2006: Prediction models for annual U.S. hurricane counts. J. Climate, 19, 2935–2952, doi:10.1175/JCLI3729.1.
Frank, W. M., and G. S. Young, 2007: The interannual variability of tropical cyclones. Mon. Wea. Rev., 135, 3587–3598, doi:10.1175/MWR3435.1.
Gerrity, J. P., Jr., 1992: A note on Gandin and Murphy’s equitable skill score. Mon. Wea. Rev., 120, 2709–2712, doi:10.1175/1520-0493(1992)120<2709:ANOGAM>2.0.CO;2.
Goldenberg, S. B., C. W. Landsea, A. M. Mestas-Nuñez, and W. M. Gray, 2001: The recent increase in Atlantic hurricane activity: Causes and implications. Science, 293, 474–479, doi:10.1126/science.1060040.
Gray, W. M., 1998: The formation of tropical cyclones. Meteor. Atmos. Phys., 67, 37–69, doi:10.1007/BF01277501.
Gray, W. M., C. W. Landsea, P. W. Mielke, and K. J. Berry, 1992: Predicting Atlantic seasonal hurricane activity 6–11 months in advance. Wea. Forecasting, 7, 440–455, doi:10.1175/1520-0434(1992)007<0440:PASHAM>2.0.CO;2.
Hall, T. M., and S. Jewson, 2007: Statistical modeling of North Atlantic tropical cyclone tracks. Tellus, 59A, 486–498, doi:10.1111/j.1600-0870.2007.00240.x.
Hall, T. M., and S. Jewson, 2008: Comparison of local and basinwide methods for risk assessment of cyclone landfall. J. Appl. Meteor. Climatol., 47, 361–367, doi:10.1175/2007JAMC1720.1.
Hess, J. C., J. B. Elsner, and N. E. LaSeur, 1995: Improving seasonal hurricane predictions for the Atlantic basin. Wea. Forecasting, 10, 425–432, doi:10.1175/1520-0434(1995)010<0425:ISHPFT>2.0.CO;2.
Ho, C.-H., J.-H. Kim, H.-S. Kim, W. Choi, M.-H. Lee, H.-D. Yoo, T.-R. Kim, and S. Park, 2013: Technical note on a track-pattern-based model for predicting seasonal tropical activity over the western North Pacific. Adv. Atmos. Sci., 30, 1260–1274, doi:10.1007/s00376-013-2237-6.
Holland, G. J., and P. J. Webster, 2007: Heightened tropical cyclone activity in the North Atlantic: Natural variability or climate trend? Philos. Trans. Roy. Soc. London, 365A, 2695–2716, doi:10.1098/rsta.2007.2083.
John, V. O., and B. J. Soden, 2007: Temperature and humidity biases in global climate models and their impact on climate feedbacks. Geophys. Res. Lett., 34, L18704, doi:10.1029/2007GL030429.
Kanamitsu, M., W. Ebisuzaki, J. Woollen, S.-K. Yang, J. J. Hnilo, M. Fiorino, and G. L. Potter, 2002: NCEP–DOE AMIP-II Reanalysis (R-2). Bull. Amer. Meteor. Soc., 83, 1631–1643, doi:10.1175/BAMS-83-11-1631.
Kim, H.-M., and P. J. Webster, 2010: Extended-range seasonal hurricane forecasts for the North Atlantic with a hybrid dynamical-statistical model. Geophys. Res. Lett., 37, L21705, doi:10.1029/2010GL044792.
Kim, H.-S., C.-H. Ho, P.-S. Chu, and J.-H. Kim, 2010: Seasonal prediction of summertime tropical cyclone activity over the East China Sea using the least absolute deviation regression and the Poisson regression. Int. J. Climatol., 30, 210–219, doi:10.1002/joc.1878.
Kim, H.-S., J.-H. Kim, C.-H. Ho, and P.-S. Chu, 2011: Pattern classification of typhoon tracks using the fuzzy c-means clustering method. J. Climate, 24, 488–508, doi:10.1175/2010JCLI3751.1.
Kim, H.-S., C.-H. Ho, J.-H. Kim, and P.-S. Chu, 2012: Track-pattern-based model for predicting seasonal tropical cyclone activity in the western North Pacific. J. Climate, 25, 4660–4678, doi:10.1175/JCLI-D-11-00236.1.
Kim, J.-H., C.-H. Ho, H.-S. Kim, and W. Choi, 2012: 2010 western North Pacific typhoon season: Seasonal overview and forecast using a track-pattern-based model. Wea. Forecasting, 27, 730–743, doi:10.1175/WAF-D-11-00109.1.
Klotzbach, P. J., 2007: Recent developments in statistical prediction of seasonal Atlantic basin tropical cyclone activity. Tellus, 59A, 511–518, doi:10.1111/j.1600-0870.2007.00239.x.
Klotzbach, P. J., 2010: On the Madden–Julian oscillation–Atlantic hurricane relationship. J. Climate, 23, 282–293, doi:10.1175/2009JCLI2978.1.
Knutson, T. R., J. J. Sirutis, S. T. Garner, I. M. Held, and R. E. Tuleya, 2007: Simulation of the recent multidecadal increase of Atlantic hurricane activity using an 18-km-grid regional model. Bull. Amer. Meteor. Soc., 88, 1549–1565, doi:10.1175/BAMS-88-10-1549.
Kossin, J. P., and D. J. Vimont, 2007: A more general framework for understanding Atlantic hurricane variability and trends. Bull. Amer. Meteor. Soc., 88, 1767–1781, doi:10.1175/BAMS-88-11-1767.
Kossin, J. P., S. J. Camargo, and M. Sitkowski, 2010: Climate modulation of North Atlantic hurricane tracks. J. Climate, 23, 3057–3076, doi:10.1175/2010JCLI3497.1.
Kozar, M. E., M. E. Mann, S. J. Camargo, J. P. Kossin, and J. L. Evans, 2012: Stratified statistical models of North Atlantic basin-wide and regional tropical cyclone counts. J. Geophys. Res., 117, D18103, doi:10.7916/D80G3J9N.
Kwon, H. J., W.-J. Lee, S.-H. Won, and E.-J. Cha, 2007: Statistical ensemble prediction of the tropical cyclone activity over the western North Pacific. Geophys. Res. Lett., 34, L24805, doi:10.1029/2007GL032308.
LaRow, T. E., L. Stefanova, D.-W. Shin, and S. Cocke, 2010: Seasonal Atlantic tropical cyclone hindcasting/forecasting using two sea surface temperature datasets. Geophys. Res. Lett., 37, L02804, doi:10.1029/2009GL041459.
Lehmiller, G. S., T. B. Kimberlain, and J. B. Elsner, 1997: Seasonal prediction models for North Atlantic basin hurricane location. Mon. Wea. Rev., 125, 1780–1791, doi:10.1175/1520-0493(1997)125<1780:SPMFNA>2.0.CO;2.
Li, X., S. Yang, H. Wang, X. Jia, and A. Kumar, 2013: A dynamical-statistical forecast model for the annual frequency of western Pacific tropical cyclones based on the NCEP Climate Forecast System version 2. J. Geophys. Res., 118, 12 061–12 074, doi:10.1002/2013JD020708.
McAdie, C. J., C. W. Landsea, C. J. Neumann, J. E. David, E. Blake, and G. R. Hammer, 2009: Tropical cyclones of the North Atlantic Ocean, 1851–2006. NCDC/TCP/NHC Historical Climatology Series 6-2, 238 pp. [Available online at http://www.nhc.noaa.gov/pdf/TC_Book_Atl_1851-2006_lowres.pdf.]
Nakamura, J., U. Lall, Y. Kushnir, and S. J. Camargo, 2009: Classifying North Atlantic tropical cyclone tracks by mass moments. J. Climate, 22, 5481–5494, doi:10.1175/2009JCLI2828.1.
Pielke, R. A., Jr., and C. W. Landsea, 1998: Normalized hurricane damages in the United States: 1925–95. Wea. Forecasting, 13, 621–631, doi:10.1175/1520-0434(1998)013<0621:NHDITU>2.0.CO;2.
Pielke, R. A., Jr., J. Gratz, C. W. Landsea, D. Collins, M. A. Saunders, and R. Musulin, 2008: Normalized hurricane damage in the United States: 1900–2005. Nat. Hazards Rev., 9, 29–42, doi:10.1061/(ASCE)1527-6988(2008)9:1(29).
Saha, S., and Coauthors, 2014: The NCEP Climate Forecast System version 2. J. Climate, 27, 2185–2208, doi:10.1175/JCLI-D-12-00823.1.
Saunders, M. A., and A. S. Lea, 2005: Seasonal prediction of hurricane activity reaching the coast of the United States. Nature, 434, 1005–1008, doi:10.1038/nature03454.
Smith, A. B., and R. W. Katz, 2013: US billion-dollar weather and climate disasters: Data sources, trends, accuracy and biases. Nat. Hazards, 67, 387–410, doi:10.1007/s11069-013-0566-5.
Smith, T. M., R. W. Reynolds, T. C. Peterson, and J. Lawrimore, 2008: Improvements to NOAA’s historical merged land–ocean surface temperature analysis (1880–2006). J. Climate, 21, 2283–2296, doi:10.1175/2007JCLI2100.1.
Vecchi, G. A., and B. J. Soden, 2007: Global warming and the weakening of the tropical circulation. J. Climate, 20, 4316–4340, doi:10.1175/JCLI4258.1.
Vecchi, G. A., M. Zhao, H. Wang, G. Villarini, A. Rosati, A. Kumar, I. M. Held, and R. Gudgel, 2011: Statistical–dynamical predictions of seasonal North Atlantic hurricane activity. Mon. Wea. Rev., 139, 1070–1082, doi:10.1175/2010MWR3499.1.
Vecchi, G. A., and Coauthors, 2014: On the seasonal forecasting of regional tropical cyclone activity. J. Climate, 27, 7994–8016, doi:10.1175/JCLI-D-14-00158.1.
Wang, H., J. K. E. Schemm, A. Kumar, W. Wang, L. Long, M. Chelliah, G. D. Bell, and P. Peng, 2009: A statistical forecast model for Atlantic seasonal hurricane activity based on the NCEP dynamical seasonal forecast. J. Climate, 22, 4481–4500, doi:10.1175/2009JCLI2753.1.
Weinkle, J., R. Maue, and R. A. Pielke Jr., 2012: Historical global tropical cyclone landfalls. J. Climate, 25, 4729–4735, doi:10.1175/JCLI-D-11-00719.1.
Wilks, D. S., 2006: Statistical Methods in the Atmospheric Sciences.2nd ed. Academic Press, 627 pp.
Xie, X. L., and G. A. Beni, 1991: A validity measure for fuzzy clustering. IEEE Trans. Pattern Anal. Mach. Intell., 13, 841–846, doi:10.1109/34.85677.
Zhao, M., I. M. Held, and G. A. Vecchi, 2010: Retrospective forecasts of the hurricane season using a global atmospheric model assuming persistence of SST anomalies. Mon. Wea. Rev., 138, 3858–3868, doi:10.1175/2010MWR3366.1.