1. Introduction
Tropical cyclones (TCs) that form in the Coral Sea region of the southwest Pacific (0°–30°S, 142.5°–170°E; Fig. 1) pose an annual risk to human life and property, affecting the east coast of Australia, New Zealand, and the islands of Melanesia (including Vanuatu, New Caledonia, and Fiji). The Great Barrier Reef, located in the Coral Sea, is susceptible to both ecological and structural damage from TCs (Done 1992). The Coral Sea region is situated between two different large-scale steering flow regimes (Ramsay et al. 2012); consequently, many TCs tend to meander and may also exhibit looping behavior.
(a) Correlation map showing the relationship between SSTA in JJA and TCC for the following season. The color bar gives the signs and magnitudes of the correlations. The three domains used to compute the EMI are denoted as box A, box B, and box C. The Coral Sea region and the three Niño regions also are identified. (b) Locations of TC genesis in the Coral Sea, delimited by vertical lines at 142.5° and 170°E. The SST variance (°C) for JFM over the period 1977–2012 is shaded (see color bar). The red dashed lines show the mean SST (°C) for JFM over the same period.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
(a) Correlation map showing the relationship between SSTA in JJA and TCC for the following season. The color bar gives the signs and magnitudes of the correlations. The three domains used to compute the EMI are denoted as box A, box B, and box C. The Coral Sea region and the three Niño regions also are identified. (b) Locations of TC genesis in the Coral Sea, delimited by vertical lines at 142.5° and 170°E. The SST variance (°C) for JFM over the period 1977–2012 is shaded (see color bar). The red dashed lines show the mean SST (°C) for JFM over the same period.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
(a) Correlation map showing the relationship between SSTA in JJA and TCC for the following season. The color bar gives the signs and magnitudes of the correlations. The three domains used to compute the EMI are denoted as box A, box B, and box C. The Coral Sea region and the three Niño regions also are identified. (b) Locations of TC genesis in the Coral Sea, delimited by vertical lines at 142.5° and 170°E. The SST variance (°C) for JFM over the period 1977–2012 is shaded (see color bar). The red dashed lines show the mean SST (°C) for JFM over the same period.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
Interannual variability of TC frequency over the southwest Pacific/Coral Sea region and the broader Australian TC region (90°–160°E) has been linked to the El Niño–Southern Oscillation (ENSO) (e.g., Nicholls 1979, 1984, 1985; Solow and Nicholls 1990; Evans and Allan 1992; Basher and Zheng 1995; Ramsay et al. 2008; Chand and Walsh 2009; Vincent et al. 2011). The strength and sign of the ENSO relationship varies across subregions of the tropical Pacific Ocean. The substantial association between TC activity and ENSO has led to the development of a number of seasonal prediction schemes of varying complexity (e.g., Nicholls 1979, 1985; Chand et al. 2010; Werner and Holbrook 2011; Liu and Chan 2012). Some of these schemes also have utilized a combination of atmospheric predictors to optimize performance. For instance, Werner and Holbrook (2011) found that a three-predictor model, comprising convective available potential energy (CAPE), 850-hPa zonal winds, and 500-hPa geopotential height, explained 53% (r = 0.73) of the hindcast variance of annual Coral Sea TC frequency from 1968 to 2008.
The genesis locations of southwest Pacific TCs differ between positive and negative phases of ENSO, with a tendency for TCs to migrate northeastward (southwestward) during an El Niño (La Niña). TC activity is enhanced westward of 170°E during a La Niña phase, whereas during an El Niño more TCs form east of 170°E (Revell and Goulter 1986a,b; Hastings 1990; Basher and Zheng 1995; Kuleshov et al. 2008; Chand and Walsh 2009; Dowdy et al. 2012). The north–south shift of TC activity has been attributed to concomitant changes in the location of the South Pacific convergence zone (SPCZ), which in turn is modulated by ENSO (e.g., Trenberth 1976; Folland et al. 2002; Vincent et al. 2011). Chand et al. (2013a,b) extended the relationship between canonical ENSO and TCs by partitioning the ENSO neutral phase into “positive-neutral” and “negative-neutral” components. For the Coral Sea region, they found large and significant differences in annual TC counts (TCC) between positive-neutral (fewer TCs) and negative-neutral (more TCs) phases, similar to the TC count magnitude differences between El Niño and La Niña.
Commonly, seasonal TC prediction schemes use a single Niño region as a predictor of TC activity—for example, Niño-3.4 is used by the Bureau of Meteorology (http://www.bom.gov.au/climate/ahead/tc.shtml#tabs=About-the-outlook) and Niño-4 is used by the Guy Carpenter Asia–Pacific Climate Impact Centre (Liu and Chan 2012) for the Australian region. An alternative is to amalgamate the individual regions into a set of weighted components. Niño regions were first combined by Trenberth and Stepaniak (2001), who constructed the Trans-Niño Index (TNI) to capture the sea surface temperature (SST) gradient between Niño-1 + 2 and Niño-4. Subsequently, Ashok et al. (2007) connected the east, central, and west tropical Pacific, forming the El Niño Modoki Index (EMI). El Niño Modoki is also known as date line El Niño (Larkin and Harrison 2005), warm pool El Niño (Kug et al. 2009), and central Pacific El Niño (Yu and Kao 2007). For reference, the locations of the Niño boxes are as follows: Niño-1 + 2 (10°S–0°, 80°–90°W); Niño-3 (5°S–5°N, 90°–150°W); Niño-3.4 (5°S–5°N, 120°–170°W); and Niño-4 (5°S–5°N, 150°W–160°E).
El Niño Modoki is known to influence TC frequencies elsewhere, such as the western North Pacific and North Atlantic (e.g., Kim et al. 2009; Chen and Tam 2010; Kim et al. 2011). The influence of the TNI on TC activity in the Atlantic basin was explored by Larson et al. (2012). The present study augments previous work by establishing statistical linkages between the seasonal number of TCs that form in the Coral Sea and 1) the ENSO regions that define the EMI and 2) the areas that form the TNI. To our knowledge, this study is the first to draw an explicit link between Coral Sea TCC and the regions defining EMI and TNI.
2. Data and methods
TC data were obtained from the International Best Track Archive for Climate Stewardship (IBTrACS; Knapp et al. 2010) for July 1977 to June 2010, and supplemented by the Southwest Pacific Enhanced Archive for Tropical Cyclones (SPEArTC; Diamond et al. 2012) for July 2010 to June 2012. The base period for the present study is therefore 1977–2012. The data for five storms from IBTrACS were corrected with data from SPEArTC: Stan 1978, Elinor 1982, Fritz 1983, Gavin 1984, and Nute 1997 [for details, see Table AII of Diamond et al. (2012)]. As the TC season spans two calendar years in the Southern Hemisphere, the first year will be used to refer to a particular season.
The rationale for this study beginning with the 1977 season is that it coincides with the introduction of routine 3-hourly geostationary satellite observations (e.g., Harper et al. 2008; Kossin et al. 2013). Previous work has suggested a positive bias exists in the TCC during the 1970s through the mid-1980s (e.g., Nicholls et al. 1998; Buckley et al. 2003; Ramsay et al. 2008), owing to the inclusion of weak or hybrid systems that would not qualify as a TC later in the observational record. Integrated TC intensity metrics, such as accumulated cyclone energy (ACE) and Power Dissipation Index (PDI), are not considered here, because of the large amount of missing wind data from IBTrACS–World Meteorological Organization (WMO), particularly in the 1980s.
To predict TCC, Modoki-based linear regression models were created, incorporating the SST in Modoki boxes as predictor variables, defined as [SST]A, [SST]B, and [SST]C. When employed as TCC model components, [SST]A, [SST]B, and [SST]C, are referred to simply as A, B, and C. All 35 additive and multiplicative models therefore are groupings of A, B, and C. Additionally, Niño-1 + 2, Niño-3.4, and Niño-4 boxes were used to create three additional models, including an optimized TNI model, for a total of 38 TCC prediction models (Table 1).
List of all models investigated; A, B, and C in models 1 to 35 refer to the Modoki boxes, as defined by Ashok et al. (2007) and as shown in Fig. 1.
Owing to the relatively limited sample size (35 yr), all models are tested using a standard procedure, the leave-one-out cross validation (LOOCV; Wilks 2011). The cross validation process defines a dataset to “test” the model to gain knowledge on how a model will generalize to an independent dataset. LOOCV uses a single observation from the original sample as the testing data, and the remaining observations as the training data to fit the model (e.g., Klotzbach 2011). The process is repeated as a jackknife process until each observation in the sample is used once as the testing data. In the present analysis, LOOCV uses 34 years of paired SST and TCC data (defined as “training data”) to build the regression model coefficients and applies those coefficients to predict the year withheld (defined as “testing data”). The LOOCV approach generates 35 sets of model coefficients and performance indices for evaluation. LOOCV assumes that there is no significant autocorrelation at lag 1 (Wilks 2011). This assumption was tested, and the lag-1 correlation (~0.17) was well within the white noise limit (0.33). Confidence intervals on the correlation (r) between the 35 pairs of predicted TCC and observed TCC are created by bootstrapping r, using 10 000 bias-corrected and accelerated (BCa) replications (Efron and Tibshirani 1993).
3. Results
a. TC climatology for the Coral Sea
During the 1977 to 2011 seasons, a total of 133 TCs formed in the Coral Sea basin. Most TCs develop during January, February, and March, and comprise 22%, 29%, and 20%, respectively, of the seasonal TCC, and approximately 90% of all TCs form between December and April. The geographical distribution of TC genesis, clustered between 5° and 20°S, on the southwestern edge of the Pacific warm pool, is located within a region of relatively low SST variance in January–March (JFM) of ~0.1°–0.2°C, and mean SST ≥28°C (Fig. 1b).
The seasonal mean TCC is 3.8, although there is notable interannual variability, with a range from 1 to 9 TCs (Table 2). The number of severe TCs, defined as central pressure less than 970 hPa, averages between 1 and 2 per season, and it is rare for a season not to have at least one severe TC (Table 2). Whereas the TC season in the entire Australian TC region spans 1 November to 30 April, the season length is much shorter for the Coral Sea. On average, the Coral Sea season commences on 2 January and ends on 3 April (Table 2). The mean season duration is ~90 days, but is highly variable (the interquartile range is 65 days), with a minimum of 5 days and a maximum of 215 days. The season start and end dates also exhibit strong fluctuations, with 8 October (3 March) as the earliest (latest) start and 5 June (21 January) as the latest (earliest) end date. It is uncommon for TCs to form prior to the official 1 November start date of the Australian region (Table 2). However, TCs have occurred after the official end date of 30 April, with nine seasons extending well into May. An early start was found not to be associated with a greater number of TCs in a season, with a low correlation of −0.15. However, there are statistically significant relationships at 95% confidence between TCC and both season end date (r = 0.39) and season length (r = 0.41), with the more active seasons lasting longer.
Summary of TC statistics for the TC seasons 1977–2011. The numbered columns from left to right are 1) TCC, 2) severe TCC, 3) TC days, 4) season start date, 5) season end date, and 6) season length. The mean, median, standard deviation, and interquartile range of each statistic are also reported.
When a linear trend line is fit to the TCC data, a downward slope of −0.039 TCs per year is evident in the seasonal TCC (Fig. 2), largely owing to several active seasons in the mid-1980s, but it is not statistically significant, consistent with the findings of Kuleshov et al. (2010) for the 38-yr period 1969 to 2006, for their broader South Pacific domain.
Observed TCC for seasons 1977–2011 with linear trend overlain (red line). Inset shows the 95% confidence interval of the linear trend based on a BCa bootstrap with 10 000 replications.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
Observed TCC for seasons 1977–2011 with linear trend overlain (red line). Inset shows the 95% confidence interval of the linear trend based on a BCa bootstrap with 10 000 replications.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
Observed TCC for seasons 1977–2011 with linear trend overlain (red line). Inset shows the 95% confidence interval of the linear trend based on a BCa bootstrap with 10 000 replications.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
A wavelet analysis was carried out on the Coral Sea TCC time series (Fig. 3a). The wavelet power spectrum and the global power spectrum for the Coral Sea TCC both were calculated, using a Morlet wavelet, to detect possible significant signals in the time–frequency domain. The 95% confidence limit is indicated by the dashed line in Figs. 3b and 3c. Whereas the ENSO relationship with TCC is well known for the Coral Sea region, the wavelet analysis provides more detailed information about when ENSO has greater or lesser influence. For example, in Fig. 3b, a double peak is evident in the 2–6-yr period of the wavelet spectrum, from approximately 1977 to 2002, which encompasses the major El Niño events of 1982/83 and 1997/98.
(a) Wavelet analysis of the Coral Sea TCC time series for the period 1977–2011. (b) The wavelet power spectrum indicates signal strengths for the entire period, ranging from low (dark blue) to high (red) annual TCC variability. (c) High variability periods mentioned in the text are indicated by peaks in the global wavelet spectrum that highlight the power of the annual TCC variability. The TCC power is defined as (TCC)2, shown on the horizontal axis. The 95% significance interval for TCC variance is the area above the dashed black contour lines, for all plots.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
(a) Wavelet analysis of the Coral Sea TCC time series for the period 1977–2011. (b) The wavelet power spectrum indicates signal strengths for the entire period, ranging from low (dark blue) to high (red) annual TCC variability. (c) High variability periods mentioned in the text are indicated by peaks in the global wavelet spectrum that highlight the power of the annual TCC variability. The TCC power is defined as (TCC)2, shown on the horizontal axis. The 95% significance interval for TCC variance is the area above the dashed black contour lines, for all plots.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
(a) Wavelet analysis of the Coral Sea TCC time series for the period 1977–2011. (b) The wavelet power spectrum indicates signal strengths for the entire period, ranging from low (dark blue) to high (red) annual TCC variability. (c) High variability periods mentioned in the text are indicated by peaks in the global wavelet spectrum that highlight the power of the annual TCC variability. The TCC power is defined as (TCC)2, shown on the horizontal axis. The 95% significance interval for TCC variance is the area above the dashed black contour lines, for all plots.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
The global power spectrum (Fig. 3c) shows significant peaks at approximately 2 and 5 yr, also consistent with ENSO, and there is a suggestion of an interdecadal signal for the entire period, with a maximum at ~14 yr, which is not significant at the 95% level. A longer time series would be required to confirm if this interdecadal periodicity is real.
b. Relationship between TCC and SSTA
Exploration of the correlations between seasonal TCC and SSTA were made using traditional ENSO measures (e.g., Niño-3.4 and Niño-4), local SSTA, and variants of ENSO defined in section 2, such as the EMI and the TNI. There was a large negative correlation between the May–July (MJJ) EMI and the TCC (r = −0.64), well before the official start of the Australian TC season on 1 November. The spatial pattern of correlations between TCC and Pacific SSTA, for the period 1977–2011, revealed a tripole consistent with EMI, characterized by large correlations in box A, flanked by correlations of similar magnitude and opposite sign in boxes B and C. Owing to the magnitudes and patterns of the correlations, it was decided to examine all possible combinations of the three Modoki SST boxes.
By varying the regression weights between −2 and +2 in increments of 0.05 and plotting the weights as ratios of b3/b1 (x axis; Fig. 4) and b2/b1 (y axis; Fig. 4), contours of SSE are shown in Fig. 4. The small white circle in Fig. 4 represents the ratio of weights for the EMI-weighted SSE, whereas the star is the ratio of weights for the optimized regression SSE (no longer the EMI, but rather A + B + C). The weights for the EMI are most suboptimal in March–May (MAM), but are much closer in the three subsequent time periods MJJ, July–September (JAS), and September–November (SON).
SSE values (shaded) as a function of the ratio of weights used in the model TCC prediction = b0 + b1[SSTA]A + b2[SSTA]B + b3[SSTA]C for months MAM, MJJ, JAS, and SON. The stars denote the optimized ratio of weights, and circles show the ratio of weights used in the EMI (i.e., b3/b1 = −0.5, b2/b1 = −0.5).
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
SSE values (shaded) as a function of the ratio of weights used in the model TCC prediction = b0 + b1[SSTA]A + b2[SSTA]B + b3[SSTA]C for months MAM, MJJ, JAS, and SON. The stars denote the optimized ratio of weights, and circles show the ratio of weights used in the EMI (i.e., b3/b1 = −0.5, b2/b1 = −0.5).
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
SSE values (shaded) as a function of the ratio of weights used in the model TCC prediction = b0 + b1[SSTA]A + b2[SSTA]B + b3[SSTA]C for months MAM, MJJ, JAS, and SON. The stars denote the optimized ratio of weights, and circles show the ratio of weights used in the EMI (i.e., b3/b1 = −0.5, b2/b1 = −0.5).
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
Each of the 35 LOOCV requires creating sets, with each set composed of 34 yr of training data and 1 yr of testing data. These training data provide a source of information about the closeness of the fit to the TCC by the predictors. Of the 38 models tested (Table 1), some models did not perform well (as indicated by small correlation values, not significant at 
Training data correlations between fitted TCC and observed TCC for each 3-month period for all models. Any 95% confidence interval about the median correlation that does not cross zero is identified by daggers (††). Models expected to generalize best based on criteria stated in the text are indicated by bold font.
Of the 35 EMI-based models examined with the training data, four satisfy criteria (i) and (ii): A + B, A + AB, A + BC, and A + ABC. Additionally, the TNI-based model, Niño-[1 + 2] + Niño-4 satisfies these criteria. These five models have a commonality as they are composed of multiple SST regions in the tropical Pacific Ocean.
As LOOCV holds out a single year for testing, statistics based on these testing data are used to explain the predictive capability of the models. Application of LOOCV to all 38 models showed that A + BC, A + ABC, and A + B + AB are consistently the most significant (99% level) for all five sets of three consecutive month prediction windows (AMJ through ASO) (Table 4). Another commonly used model performance measure, the root-mean-square error (RMSE), was used also to assess the accuracy of the models. Table 5 shows that, after LOOCV, the most accurate models are the same as those obtained from the correlation analysis (Table 4). Therefore, a total of six EMI-based models were investigated further, along with the TNI-based model, Niño-3.4 and Niño-4, making a total of nine models.
LOOCV correlations for each 3-month period for all models. Any confidence interval on the median correlation that does not cross zero at the (1-
RMSE after LOOCV for each 3-month prediction period for all models. Any confidence interval on the median difference between model RMSE and climatology RMSE that does not cross zero at the (1-
Although the official start date of the Australian TC season is 1 November, the Coral Sea region mean start date is 3 January; therefore, the predictors for SON and October–December (OND) often will be required. Figure 5a shows the correlation between the predicted and observed TCC for all the models emerging from both the training and testing data, and three reference models representing the optimized EMI and the two Niño regions (i.e., A + B + C, Niño-3.4, and Niño-4) plotted against eight 3-month periods (MAM through OND). It is clear from Figs. 5a and 6a that all models based on multiple SST regions have superior performance when compared with those based on single SST regions, such as Niño-3.4 or Niño-4. For example, the correlation between Niño-3.4 and TCC in MJJ is less than 0.1, whereas A + AB has a correlation exceeding 0.5 for the same 3-month window. Furthermore, Fig. 5a suggests that four models—A + BC, A + ABC, the optimized EMI (A + B + C) and the optimized TNI (Niño [1 + 2]+Niño 4)—are equally as good at predicting TCC based on SST in June–August (JJA). This finding is expected, as the models incorporate similar SST regions [Fig. 1(a)].
(a) Correlations between observed TCC and fitted TCC using training data from nine models, plotted against 3-month prediction period. (b) As in (a), but for predicted TCC following LOOCV.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
(a) Correlations between observed TCC and fitted TCC using training data from nine models, plotted against 3-month prediction period. (b) As in (a), but for predicted TCC following LOOCV.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
(a) Correlations between observed TCC and fitted TCC using training data from nine models, plotted against 3-month prediction period. (b) As in (a), but for predicted TCC following LOOCV.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
(a) RMSE values for fitted TCC using training data from nine models, plotted against 3-month prediction period. (b) As in (a), but for predicted TCC following LOOCV.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
(a) RMSE values for fitted TCC using training data from nine models, plotted against 3-month prediction period. (b) As in (a), but for predicted TCC following LOOCV.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
(a) RMSE values for fitted TCC using training data from nine models, plotted against 3-month prediction period. (b) As in (a), but for predicted TCC following LOOCV.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
Each model in Fig. 5a is cross-validated and the correlation between the prediction and testing data shown in Fig. 5b. Comparing the generalization performance of the models in Fig. 5b to their counterparts in Fig. 5a documents the decreased correlation for each model when the testing data are applied. The large drop in correlations from the best models using the training data (r = 0.65) to the best models using the testing data (r = 0.51) translates to an even greater decrease in the TCC variance explained (down from 42% to 26%, which is a 38% decrease). This is an example of cross-validation shrinkage, as described in Wilks (2011, p. 254). Similarly, the RMSE for these models increase after LOOCV and the ranking of the accuracy of the models remains unchanged (Figs. 6a,b). This finding has implications for seasonal TC predictions in other basins, emphasizing the need for cross-validation in model evaluation. For the interested reader, the coefficient weights from the regression (for JJA) are provided in Table 6 for the nine models shown in Figs. 5 and 6. Note that the coefficient weights are based on SST and not SSTA, and therefore the weights for the multiplicative terms are substantially smaller than the coefficient weights for single regions.
Coefficient weights for the nine models shown in Figs. 5 and 6, during the prediction period JJA.
Figures 5b and 6b can be summarized as follows: predictions based on Niño-3.4 and Niño-4 are poorly correlated with TCC over the specified five 3-month overlapping periods (AMJ to ASO). Of the remaining models, A + B + AB is now rejected because of the poor correlations in three of the five months (MJJ to JAS). Of the last five models, three perform comparably best. In order of decreasing correlation (and increasing RMSE), these three are A + BC, A + ABC, and Niño-[1 + 2] + Niño-4. Three other models (A + B, A + AB, A + B + C) perform less well.
4. Discussion
Numerous seasonal TC prediction schemes use Niño regions as predictors (e.g., Niño-3.4 is used by the Australian Bureau of Meteorology for its predictions (http://www.bom.gov.au/climate/ahead/tc.shtml#tabs=About-the-outlook) and the Guy Carpenter Asia–Pacific Climate Impact Centre uses Niño 4 (Liu and Chan 2012). Despite the weak positive correlations between Niño-3.4 and TCC in MAM and AMJ (Fig. 5a), Niño-3.4 is found to be negatively correlated with TCC when cross-validated (Fig. 5b), but the relationship is not stable, as the correlation becomes positive in MJJ. Niño-4 has larger correlations with TCC than Niño-3.4, but they remain lower than models based on the relative warming or cooling of the central Pacific compared to the eastern Pacific and, hence, neither the Niño-3.4- nor the Niño-4-based models are competitive. In contrast, the cross-validated multiple Niño region model, Niño-[1 + 2] + Niño 4 (the optimized TNI model), has correlations of ~0.5 in four of the five 3-month windows (Table 4, Fig. 5b). Under LOOCV, the optimized TNI model correlation is exceeded only by the TCC correlations in models A + BC and A + ABC, because they capture much of the same variance in SST as does the EMI. Moreover, Niño-1 + 2 and Niño-4 are geographical subsets of [SSTA]B and [SSTA]A, respectively (Fig. 1a).
Figure 7 shows time series of observed versus predicted TCC for the best model found here (A + BC) and also for the widely used ENSO index (Niño-3.4) during MJJ, providing a substantial TCC prediction lead time. The time series of TCC, predicted by A + BC in MJJ, captures much of the interannual variability of the observed TCC (Fig. 7a), including the active seasons of 1983 and 1998. The model prediction from the LOOCV time series is similar to the model prediction for the multiple time series generated from the training data. The most notable difference between the training and testing data occurs in the extreme season of 1983 (Fig. 7a), as expected.
(a) Time series of observed TCC (dark blue line) and predicted TCC following LOOCV (red line) from model A + BC for the MJJ predictability window. The thin light blue lines show the 35 training models. The mean TCC of ~3.8 is indicated by the dashed line. (b) As in (a), but for prediction based on Niño-3.4 SSTA.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
(a) Time series of observed TCC (dark blue line) and predicted TCC following LOOCV (red line) from model A + BC for the MJJ predictability window. The thin light blue lines show the 35 training models. The mean TCC of ~3.8 is indicated by the dashed line. (b) As in (a), but for prediction based on Niño-3.4 SSTA.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
(a) Time series of observed TCC (dark blue line) and predicted TCC following LOOCV (red line) from model A + BC for the MJJ predictability window. The thin light blue lines show the 35 training models. The mean TCC of ~3.8 is indicated by the dashed line. (b) As in (a), but for prediction based on Niño-3.4 SSTA.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
The physical relationships underpinning the accuracy of model A+BC were explored by investigating the Pacific-wide SST patterns—both in the preseason period (JJA) and during the TC season [December–February (DJF)]—in addition to well-known large-scale atmospheric parameters favorable for TC formation, obtained from the Interim European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-Interim; Dee et al. 2011). More active or less active TCC seasons were often preceded by either Modoki-like (tripole) or Trans-Niño-like (dipole) SSTA patterns during the JJA period. Examples of such patterns are shown for two of the most active seasons of 1983 and 1998 (Figs. 8a,d) and two of the least active seasons of 1994 and 2004 (Figs. 9a,d). The tripole pattern in JJA appears to be a precursor to substantial warm or cool anomalies occurring in box A during the ensuing TC season (DJF), as evident in Figs. 8b,d and 9b,d. The cool (warm) anomalies in box A are associated with high (low) relative humidity (RH) anomalies at 600 hPa over the Coral Sea TC genesis region (Figs. 8c,f and 9c,f), which in turn provide favorable (unfavorable) atmospheric conditions for TC formation. Further, a correlation of −0.62 was obtained between SSTA in box A and 600-hPa RH anomaly, averaged over the Coral Sea genesis region, shown in Fig. 1b, for 1979–2011, suggesting that this relationship is more general than just the illustrative four seasons shown here. Other atmospheric parameters known to modulate TCC, such as 850-hPa wind anomalies and 850–200-hPa layer shear anomalies, were explored also in relation to box A SST and TCC, but no consistent relationship was found. The strong influence of midlevel RH on TCC in the Coral Sea basin is consistent with Camargo et al. (2007), who showed that 600-hPa RH is the dominant factor in the Genesis Potential Index (GPI) for that region.
Examples of two active seasons, 1983 and 1998. (a),(b) The SSTA patterns for JJA (preseason period) 1983 and DJF (during TC season) 1983/84, respectively. (c) Relative humidity anomalies from ERA-Interim reanalysis at 600 hPa for DJF 1983/84. The subregions are as in Fig. 1, excluding the Niño regions. The signs and magnitudes of the anomalies are given by the color bars below (a) and (c). (d)–(f) As in (a)–(c), but for the 1998 season.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
Examples of two active seasons, 1983 and 1998. (a),(b) The SSTA patterns for JJA (preseason period) 1983 and DJF (during TC season) 1983/84, respectively. (c) Relative humidity anomalies from ERA-Interim reanalysis at 600 hPa for DJF 1983/84. The subregions are as in Fig. 1, excluding the Niño regions. The signs and magnitudes of the anomalies are given by the color bars below (a) and (c). (d)–(f) As in (a)–(c), but for the 1998 season.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
Examples of two active seasons, 1983 and 1998. (a),(b) The SSTA patterns for JJA (preseason period) 1983 and DJF (during TC season) 1983/84, respectively. (c) Relative humidity anomalies from ERA-Interim reanalysis at 600 hPa for DJF 1983/84. The subregions are as in Fig. 1, excluding the Niño regions. The signs and magnitudes of the anomalies are given by the color bars below (a) and (c). (d)–(f) As in (a)–(c), but for the 1998 season.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
Examples of two less active seasons, 1994 and 2004. (a),(b) The SSTA patterns for JJA (preseason period) 1994 and DJF (during TC season) 1994/95, respectively. (c) Relative humidity anomalies from ERA-Interim reanalysis at 600 hPa for DJF 1994/95. The subregions are as in Fig. 1, excluding the Niño regions. The signs and magnitudes of the anomalies are given by the color bars below (a) and (c). (d)–(f) As in (a)–(c), but for the 2004 season.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
Examples of two less active seasons, 1994 and 2004. (a),(b) The SSTA patterns for JJA (preseason period) 1994 and DJF (during TC season) 1994/95, respectively. (c) Relative humidity anomalies from ERA-Interim reanalysis at 600 hPa for DJF 1994/95. The subregions are as in Fig. 1, excluding the Niño regions. The signs and magnitudes of the anomalies are given by the color bars below (a) and (c). (d)–(f) As in (a)–(c), but for the 2004 season.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
Examples of two less active seasons, 1994 and 2004. (a),(b) The SSTA patterns for JJA (preseason period) 1994 and DJF (during TC season) 1994/95, respectively. (c) Relative humidity anomalies from ERA-Interim reanalysis at 600 hPa for DJF 1994/95. The subregions are as in Fig. 1, excluding the Niño regions. The signs and magnitudes of the anomalies are given by the color bars below (a) and (c). (d)–(f) As in (a)–(c), but for the 2004 season.
Citation: Journal of Climate 27, 22; 10.1175/JCLI-D-14-00017.1
Predictions based on Niño-3.4 using data for the MJJ prediction period are shown in Fig. 7b. Unlike the predictions for A + BC in Fig. 7a, which have large deviations about the mean, those for Niño-3.4 (Fig. 7b) are only weakly dispersed about the climatological mean of 3.8 TCs per season. Therefore, using Niño-3.4 as the sole ENSO-based seasonal predictor of Coral Sea TCC is not recommended, since it provides only negligible skill measured relative to climatology. Similarly, Niño-4, while superior to Niño-3.4, also is not recommended as a sole ENSO-based seasonal predictor of Coral Sea TCC. Finally, whereas it is possible to increase seasonal forecast skill for the region by incorporating a number of atmospheric predictors (e.g., Werner and Holbrook 2011), here we have examined the role of just one predictor, the ENSO-related SST, which is a key component of many TC seasonal prediction schemes.
5. Conclusions
Analyses carried out in this study found that the Coral Sea TC basin (0°–30°S, 142.5°–170°E) averages 3.8 TCs per season, although this number fluctuates considerably from year to year, ranging from as few as 1 TC (1980 and 2004) to as many as 9 TCs (1983). The average season starts on 2 January and, although some seasons have started as early as October, there is no significant statistical correlation between an early start date and a more active season.
There is a well-established relationship between traditional ENSO measures (e.g., Niño-3.4, Niño-4) and Australian/southwest Pacific TC activity on the interannual time scale, with enhanced TC activity occurring during La Niña and suppressed activity during El Niño (e.g., Nicholls 1979, 1984, 1985; Solow and Nicholls 1990; Evans and Allan 1992; Basher and Zheng 1995; Ramsay et al. 2008; Chand and Walsh 2009; Vincent et al. 2011). Moreover, it has been recognized that at least two predictors are needed to describe the evolution of ENSO (e.g., Trenberth and Stepaniak 2001). More recently, the El Niño Modoki (Ashok et al. 2007) has received considerable attention, given its relevance to regional climate modeling, including Australian precipitation anomalies (e.g., Taschetto and England 2009; Cai and Cowan 2009; Risbey et al. 2009). Modoki ENSO events also have been shown to affect TC frequency in the North Atlantic and northwest Pacific basins (e.g., Kim et al. 2009; Kim et al. 2011; Chen and Tam 2010), but its influence on southwest Pacific TCs has been less explored (Chand et al. 2013a,b; Wijnands et al. 2014).
Motivated by the Modoki ENSO studies cited above, a regression-based prediction system is created, constructed from those regions, and optimized to establish a highly significant statistical relationship between SST and the seasonal TCC in the Coral Sea. The predictions show significant correlations at lead times up to 6 months prior to the mean onset of the Coral Sea TC season. Models using multiple predictors, thereby accounting for the relative warming or cooling of the central Pacific Ocean, in contrast with the eastern and western Pacific Ocean, perform much better than those models based on individual SST indices, such as Niño-3.4 and Niño-4. Models that use optimized variants of the El Niño Modoki index to generate predictions of TCC give the highest correlations with observed TCC using the training dataset (Fig. 5a) but, because that analysis is not cross-validated, the correlation magnitude is inflated because of overfitting. In general, a failure to cross-validate leads to an ~40% overestimate of the TCC variance explained by the SST regions. Therefore, it is strongly recommended that all statistical modeling of TCC apply some form of cross-validation to document the climate signal that generalizes.
Using LOOCV, the two best models (based on both correlation and RMSE) are found to be A + BC and A + ABC, demonstrating that the interaction between the SSTA in Modoki box B (eastern Pacific Ocean) and those in Modoki box C (western tropical Pacific Ocean) is important. These best models, A + BC and A + ABC, are statistically significant at the 99% level for the five 3-month periods considered in this study. A third model that employs multiple ENSO boxes, Niño-[1 + 2] + Niño 4 (the optimized TNI), performs comparably during JJA, but is less accurate in the other seven prediction periods (Figs. 5b and 6b). All three models incorporate the critical SST gradient between the central and eastern Pacific, and capture the observed interannual TCC variability in the Coral Sea. To gain some physical understanding of the best statistical models, several atmospheric parameters known to modulate TCC were explored, including 850-hPa winds, 850–200-hPa wind shear, and midlevel relative humidity. Of these, only midlevel relative humidity revealed a consistent relationship with both TCC and central Pacific (Modoki box A) SSTA, such that it increased during central Pacific cooling events, and decreased during warming events. Future work will include a focus on nonlinear Coral Sea TCC predictors.
Acknowledgments
We thank Andrew Ballinger for stimulating discussions on the topic and for his assistance with part of the statistical code, and Alexandre Fierro for providing the wavelet analysis. The first author (HAR) acknowledges funding from the ARC Centre of Excellence for Climate System Science. MBR and LML acknowledge the support of the Cooperative Institute for Mesoscale Meteorological Studies.
REFERENCES
Ashok, K., S. K. Behera, S. A. Rao, H. Weng, and T. Yamagata, 2007: El Niño Modoki and its possible teleconnection. J. Geophys. Res., 112, C11007, doi:10.1029/2006JC003798.
Basher, R. E., and X. Zheng, 1995: Tropical cyclones in the southwest Pacific: Spatial patterns and relationships to Southern Oscillation and sea surface temperature. J. Climate, 8, 1249–1260, doi:10.1175/1520-0442(1995)008<1249:TCITSP>2.0.CO;2.
Buckley, B. W., L. M. Leslie, and M. S. Speer, 2003: The impact of observational technology on climate database quality: Tropical cyclones in the Tasman Sea. J. Climate, 16, 2640–2645, doi:10.1175/1520-0442(2003)016<2640:TIOOTO>2.0.CO;2.
Cai, W., and T. Cowan, 2009: La Niña Modoki impacts Australia autumn rainfall variability. Geophys. Res. Lett., 36, L12805, doi:10.1029/2009GL037885.
Camargo, S. J., K. A. Emanuel, and A. H. Sobel, 2007: Use of a genesis potential index to diagnose ENSO effects on tropical cyclone genesis. J. Climate, 20, 4819–4834, doi:10.1175/JCLI4282.1.
Chand, S. S., and K. J. E. Walsh, 2009: Tropical cyclone activity in the Fiji region: Spatial patterns and relationship to large-scale circulation. J. Climate, 22, 3877–3893, doi:10.1175/2009JCLI2880.1.
Chand, S. S., K. J. E. Walsh, and J. C. L. Chan, 2010: A Bayesian regression approach to seasonal prediction of tropical cyclones affecting the Fiji region. J. Climate, 23, 3425–3445, doi:10.1175/2010JCLI3521.1.
Chand, S. S., J. McBride, K. Tory, M. Wheeler, and K. J. E. Walsh, 2013a: Impact of different ENSO regimes on southwest Pacific tropical cyclones. J. Climate, 26, 600–608, doi:10.1175/JCLI-D-12-00114.1.
Chand, S. S., K. Tory, J. McBride, M. Wheeler, R. Dare, and K. J. E. Walsh, 2013b: The different impact of positive-neutral and negative-neutral ENSO regimes on Australian tropical cyclones. J. Climate, 26, 8008–8016, doi:10.1175/JCLI-D-12-00769.1.
Chen, G., and C.-Y. Tam, 2010: Different impacts of two kinds of Pacific Ocean warming on tropical cyclone frequency over the western North Pacific. Geophys. Res. Lett., 37, L01803, doi:10.1029/2009GL041708.
Dee, D. P., and Coauthors, 2011: The ERA-Interim reanalysis: Configuration and performance of the data assimilation system. Quart. J. Roy. Meteor. Soc., 137, 553–597, doi:10.1002/qj.828.
Diamond, H. J., A. M. Lorrey, K. R. Knapp, and D. H. Levinson, 2012: Development of an enhanced tropical cyclone tracks database for the southwest Pacific from 1840 to 2010. Int. J. Climatol., 32, 2240–2250, doi:10.1002/joc.2412.
Done, T. J., 1992: Effects of tropical cyclone waves on ecological and geomorphological structures on the Great Barrier Reef. Cont. Shelf Res., 12, 859–872, doi:10.1016/0278-4343(92)90048-O.
Dowdy, A. J., L. Qi, D. Jones, H. Ramsay, R. Fawcett, and Y. Kuleshov, 2012: Tropical cyclone climatology of the South Pacific Ocean and its relationship to El Niño–Southern Oscillation. J. Climate, 25, 6108–6122, doi:10.1175/JCLI-D-11-00647.1.
Efron, B., and R. J. Tibshirani, 1993: An Introduction to the Bootstrap. Monogr. Stat. Appl. Probab., Vol. 57, Chapman and Hall, 456 pp.
Evans, J. L., and R. J. Allan, 1992: El Niño/Southern Oscillation modification to the structure of the monsoon and tropical cyclone activity in the Australasian region. Int. J. Climatol., 12, 611–623, doi:10.1002/joc.3370120607.
Folland, C. K., J. A. Renwick, M. J. Salinger, and A. B. Mullan, 2002: Relative influence of the interdecadal Pacific oscillation and ENSO on the South Pacific convergence zone. Geophys. Res. Lett., 29, 1643, doi:10.1029/2001GL014201.
Harper, B. A., S. A. Stroud, M. McCormack, and S. West, 2008: A review of historical tropical cyclone intensity in northwestern Australia and implications for climate change trend analysis. Aust. Meteor. Mag., 57, 121–141.
Hastings, P. A., 1990: Southern Oscillation influences on tropical cyclone activity in the Australian/southwest Pacific region. Int. J. Climatol., 10, 291–298, doi:10.1002/joc.3370100306.
Kim, H.-M., P. J. Webster, and J. A. Curry, 2009: Impact of shifting patterns of Pacific Ocean warming events on North Atlantic tropical cyclone. Science, 325, 77–80, doi:10.1126/science.1174062.
Kim, H.-M., P. J. Webster, and J. A. Curry, 2011: Modulation of North Pacific tropical cyclone activity by three phases of ENSO. J. Climate, 24, 1839–1849, doi:10.1175/2010JCLI3939.1.
Klotzbach, P. J., 2011: A simplified Atlantic basic seasonal hurricane prediction scheme from 1 August. Geophys. Res. Lett., 38, L16710, doi:10.1029/2011GL048603.
Knapp, K. R., M. C. Kruk, D. H. Levinson, H. J. Diamond, and C. J. Neumann, 2010: The International Best Track Archive for Climate Stewardship (IBTrACS). Bull. Amer. Meteor. Soc., 91, 363–376, doi:10.1175/2009BAMS2755.1.
Kossin, J. P., T. L. Olander, and K. R. Knapp, 2013: Trend analysis with a new global record of tropical cyclone intensity. J. Climate, 26, 9960–9976, doi:10.1175/JCLI-D-13-00262.1.
Kug, J.-S., F.-F. Jin, and S.-I. An, 2009: Two types of El Niño events: Cold tongue El Niño and warm pool El Niño. J. Climate, 22, 1499–1515, doi:10.1175/2008JCLI2624.1.
Kuleshov, Y., L. Qi, R. Fawcett, and D. Jones, 2008: On tropical cyclone activity in the Southern Hemisphere: Trends and the ENSO connection. Geophys. Res. Lett., 35, L14S08, doi:10.1029/2007GL032983.
Kuleshov, Y., R. Fawcett, L. Qi, B. Trewin, D. Jones, J. McBride, and H. Ramsay, 2010: Trends in tropical cyclones in the South Indian Ocean and the South Pacific Ocean. J. Geophys. Res., 115, D01101, doi:10.1029/2009JD012372.
Larkin, N. K., and D. E. Harrison, 2005: Global seasonal temperature and precipitation anomalies during El Niño autumn and winter. Geophys. Res. Lett., 32, L16705, doi:10.1029/2005GL022860.
Larson, S., S.-K. Lee, C. Wang, E.-S. Chung, and D. Enfield, 2012: Impacts of non-canonical El Niño patterns on Atlantic hurricane activity. Geophys. Res. Lett., 39, L14706, doi:10.1029/2012GL052595.
Liu, K. S., and J. C. L. Chan, 2012: Interannual variation of Southern Hemisphere tropical cyclone activity and seasonal forecast of tropical cyclone number in the Australian region. Int. J. Climatol., 32, 190–202, doi:10.1002/joc.2259.
Nicholls, N., 1979: A possible method for predicting seasonal tropical cyclone activity in the Australian region. Mon. Wea. Rev., 107, 1221–1224, doi:10.1175/1520-0493(1979)107<1221:APMFPS>2.0.CO;2.
Nicholls, N., 1984: The Southern Oscillation, sea-surface temperature, and internannual fluctuations in Australian tropical cyclone activity. J. Climatol., 4, 661–670, doi:10.1002/joc.3370040609.
Nicholls, N., 1985: Predictability of interannual variations in Australian seasonal tropical cyclone activity. Mon. Wea. Rev., 113, 1144–1149, doi:10.1175/1520-0493(1985)113<1144:POIVOA>2.0.CO;2.
Nicholls, N., C. W. Landsea, and J. Gill, 1998: Recent trends in Australian tropical cyclone activity. Meteor. Atmos. Phys., 65, 197–205, doi:10.1007/BF01030788.
Ramsay, H. A., L. M. Leslie, P. J. Lamb, M. B. Richman, and M. Leplastrier, 2008: Interannual variability of tropical cyclones in the Australian region: Role of large-scale environment. J. Climate, 21, 1083–1103, doi:10.1175/2007JCLI1970.1.
Ramsay, H. A., S. J. Camargo, and D. Kim, 2012: Cluster analysis of tropical cyclone tracks in the Southern Hemisphere. Climate Dyn., 39, 897–917, doi:10.1007/s00382-011-1225-8.
Rayner, N. A., D. E. Parker, E. B. Horton, C. K. Folland, L. V. Alexander, D. P. Rowell, E. C. Kent, and A. Kaplan, 2003: Global analyses of sea surface temperature, sea ice, and night marine air temperature since the late nineteenth century. J. Geophys. Res., 108, 4407, doi:10.1029/2002JD002670.
Revell, C. G., and S. W. Goulter, 1986a: South Pacific tropical cyclones and the Southern Oscillation. Mon. Wea. Rev., 114, 1138–1145, doi:10.1175/1520-0493(1986)114<1138:SPTCAT>2.0.CO;2.
Revell, C. G., and S. W. Goulter, 1986b: Lagged relations between the Southern Oscillation and numbers of tropical cyclones in the South Pacific region. Mon. Wea. Rev., 114, 2669–2670, doi:10.1175/1520-0493(1986)114<2669:LRBTSO>2.0.CO;2.
Risbey, J. S., M. J. Pook, P. C. McIntosh, M. C. Wheeler, and H. H. Hendon, 2009: On the remote drivers of rainfall variability in Australia. Mon. Wea. Rev., 137, 3233–3253, doi:10.1175/2009MWR2861.1.
Solow, A., and N. Nicholls, 1990: The relationship between the Southern Oscillation and tropical cyclone frequency in the Australian region. J. Climate, 3, 1097–1101, doi:10.1175/1520-0442(1990)003<1097:TRBTSO>2.0.CO;2.
Taschetto, A. S., and M. H. England, 2009: El Niño Modoki impacts on Australian rainfall. J. Climate, 22, 3167–3174, doi:10.1175/2008JCLI2589.1.
Trenberth, K. E., 1976: Spatial and temporal variations of the Southern Oscillation. Quart. J. Roy. Meteor. Soc., 102, 639–653, doi:10.1002/qj.49710243310.
Trenberth, K. E., and D. P. Stepaniak, 2001: Indices of El Niño evolution. J. Climate, 14, 1697–1701, doi:10.1175/1520-0442(2001)014<1697:LIOENO>2.0.CO;2.
Vincent, E. M., M. Lengaigne, C. E. Menkes, N. C. Jourdain, P. Marchesiello, and G. Madec, 2011: Interannual variability of the South Pacific convergence zone and implications for tropical cyclone genesis. Climate Dyn., 36, 1881–1896, doi:10.1007/s00382-009-0716-3.
Werner, A., and N. J. Holbrook, 2011: A Bayesian forecast model for Australian region tropical cyclone formation. J. Climate, 24, 6114–6131, doi:10.1175/2011JCLI4231.1.
Wijnands, J. S., K. Shelton, and Y. Kuleshov, 2014: Improving the operational methodology of tropical cyclone seasonal prediction in the Australian and the South Pacific Ocean regions. Adv. Meteor., 2014, 838746, doi:10.1155/2014/838746.
Wilks, D. S., 2011: Statistical Methods in the Atmospheric Sciences. 3rd ed. Academic Press, 676 pp.
Yu, J. Y., and H. Y. Kao, 2007: Decadal changes of ENSO persistence barrier in SST and ocean heat content indices: 1958–2001. J. Geophys. Res., 112, D13106, doi:10.1029/2006JD007654.
