Advertisement

Theoretical and Applied Climatology

, Volume 130, Issue 3–4, pp 865–877 | Cite as

Extended-range forecast for the temporal distribution of clustering tropical cyclogenesis over the western North Pacific

  • Zhiwei Zhu
  • Tim Li
  • Long Bai
  • Jianyun Gao
Original Paper

Abstract

Based on outgoing longwave radiation (OLR), an index for clustering tropical cyclogenesis (CTC) over the western North Pacific (WNP) was defined. Around 76 % of total CTC events were generated during the active phase of the CTC index, and 38 % of the total active phase was concurrent with CTC events. For its continuous property, the CTC index was used as the representative predictand for extended-range forecasting the temporal distribution of CTC events.

The predictability sources for CTC events were detected via correlation analyses of the previous 35–5-day lead atmospheric fields against the CTC index. The results showed that the geopotential height at different levels and the 200 hPa zonal wind over the global tropics possessed large predictability sources, whereas the predictability sources of other variables, e.g., OLR, zonal wind, and relatively vorticity at 850 hPa and relatively humility at 700 hPa, were mainly confined to the tropical Indian Ocean and western Pacific Ocean.

Several spatial-temporal projection model (STPM) sets were constructed to carry out the extended-range forecast for the CTC index. By combining the output of STPMs separately conducted for the two dominant modes of intraseasonal variability, e.g., the 10–30 and the 30–80 day mode, useful forecast skill could be achieved for a 30-day lead time. The combined output successfully captured both the 10–30 and 30–80 day mode at least 10 days in advance. With a relatively low rate of false alarm, the STPM achieved hits for 80 % (69 %) of 54 CTC events during 2003–2014 at the 10-day (20-day) lead time, suggesting a practical value of the STPM for real-time forecasting WNP CTC events at an extended range.

Keywords

Lead Time Western North Pacific Outgoing Longwave Radiation Forecast Skill Monsoon Trough 
These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

1 Introduction

East Asia is vulnerable to the climate hazards caused by the tropical cyclones (TCs) that form over the western North Pacific (WNP). More than a third of TCs globally are generated in this region (Gray 1968). Most WNP tropical cyclogenesis occurs at the interface of monsoon westerlies and trade easterlies during boreal summer. Because of the northwest–southeast orientation of the summer monsoon trough, WNP TCs tend to move northwestward, affecting the east coast of continental East Asia (Lander 1996). Therefore, along with summer monsoon rainfall, TCs over the WNP have disastrous impacts on the East Asian coastal region. Given increasing human population and associated development of this areas, substantial TC-related socioeconomic losses are not uncommon.

At the global scale, observational analyses indicate that TC activities posses a distinct clustering feature in terms of their temporal distribution. Between about five to 15 TCs are often observed in just 1–2 weeks, whereas in the next following 2–3 weeks hardly any TCs may appear (Gray 1979). More than one tropical cyclogenesis event in a relatively short period is referred to as clustering tropical cyclogenesis (CTC) or multiple TC events (Gao and Li 2011, 2012). Tropical cyclogenesis is largely governed by atmospheric intraseasonal oscillation (ISO) in all ocean basins. For example, Liebmann et al. (1994) found that TCs over the WNP and Indian Ocean tend to form in the ISO’s wet phase. Maloney and Hartmann (2000a, 2000b) indicated that hurricane geneses over different basins are all closely related with the wind phases of the Madden–Julian Oscillation (MJO). Ko and Hsu (2009) suggested that the ISO’s westerly phase provides a favorable background for an enhanced monsoon trough and abundant moisture for TC development, whereas the ISO’s easterly phase leads to a less favorable environment. Gao and Li (2011) suggested that the overall occurrence of multiple TC events is strongly regulated by the combined large-scale impact of the quasi-biweekly oscillation (QBWO), the MJO, and lower-frequency (90 days or longer) oscillations. Based on 30–80 day filtered outgoing longwave radiation (OLR) anomalies averaged over (5°–20° N, 120°–150° E), Cao et al. (2012) reported the number of TCs during the active phases of ISO to be fourfold greater than during its inactive phases; plus, TC formation over the WNP during the active phases of ISO can be mainly attributed to the barotropic instability of low-level zonal flow. Utilizing an MJO index reconstructed over a long period, Klotzbach and Oliver (2015) pointed out that the relationship between TC genesis and the MJO was remarkably stable over the entire period from 1905 to 2012. Kim et al. (2008) suggested that landfalling TCs in southern China, Korea, and Japan are strongly modulated by ISO.

ISO is the predictability source for extended-range forecasts (Van Den Dool and Saha 1990; Zhang 2013). Thanks to the considerable progress made in our understanding of ISO, statistical and numerical models for extended-range forecasting have developed rapidly in recently years. However, the objects of extended-range forecasts are usually precipitation and other atmospheric circulation variables (Hsu et al., 2015; Zhu et al., 2015; Zhu and Li 2016; Li et al., 2016). Although the solid relationship between ISO and tropical cyclogenesis has been extensively investigated (William and Roundy 2006), the effects are seldom devoted to generating extended-range forecasts of tropical cyclogenesis. Meanwhile, CTC episodes carry greater destructive power than a single TC; their impacts in East Asian coastal areas frequently bring about severe flooding and wind damage. Therefore, accurate forecasting the timing of CTC at 10–30-day lead times is of major socioeconomic importance in this region. Encouragingly, ISO prediction has developed rapidly, meaning a corresponding improvement in the skill of CTC extended-range forecasts is likely achievable. It has, for instance, been found that convective anomalies of waves associated with tropical cyclogenesis are detectable in analyses as early as 1 month prior to genesis (Frank and Roundy 2006). This opens up the possibility of developing statistical extended-range forecast models for CTC.

Crucially, knowledge regarding tropical cyclogenesis remains limited, especially in terms of whether or not an observed tropical disturbance will develop into a TC (e.g., Fu et al., 2012). Thus, CTC prediction remains an open issue. Currently, the JTWC (Joint Typhoon Warning Center) and NHC (National Hurricane Center) only provide very short-range (fewer than 5 days) TC genesis probability forecasts. The ECMWF (European Center for Medium Range Weather Forecasting) issue a 12-day lead TC activity product for seven ocean basins (Vitart et al., 2012). However, extended-range CTC forecasts remain deficient in operation. Based on 51-member ECMWF 32-day ensemble forecasts, Elsberry et al. (2014) indicated that the extended-range (5–30-day) forecasting of TC events (formations and subsequent tracks) is more likely to be achieved in the WNP than in the Atlantic. Yamaguchi et al. (2015) indicated that global medium-range ensembles are capable of providing guidance on TC activity forecasts beyond 1 week. Based on a new version of the GFDL (Geophysical Fluid Dynamics Laboratory) model, Xiang et al. (2015) successfully reproduced the genesis and track of Hurricane Sandy and Super Typhoon Haiyan at an 11-day lead time, indicating the potential of using the GFDL coupled model for the extended-range forecasting of TCs over both the Atlantic Ocean and WNP. The above studies used state-of-the-art numerical models and their ensemble outputs to generate forecasts of the tropical cyclogenesis; however, none of them was based on an empirical model. CTC has a closer relationship than TCs do with ISO. Therefore, extended-range forecasting of CTC with encouraging forecast skill is likely to be achievable. Besides, compared with numerical models, statistical models offer a dramatically lower computational cost. Thus, the objective of the study reported in the present paper was to construct several extended-range statistical models to forecast CTC over the WNP.

Following this introduction, Section 2 describes the data and methodology used in the study. The predictability sources of CTC and the construction procedure of the extended-range forecast models are discussed in Section 3. In Section 4, the forecast skills of the different spatial-temporal projection models (STPMs) are assessed, and the superiority of the combined outputs from the two best STPMs which are for two independent ISO modes are presented. Section 5 provides a conclusion and discussion.

2 Data and methodology

2.1 Data

The TC best-track datasets from the JTWC were used to determine the dates of WNP TC genesis during the period 1979–2014. Specifically, we focused on the period from the 31st to the 61st pentad (May 31 to October 28), which is the most active period for WNP TC genesis. The total number of TCs (including tropical depressions) over the WNP was 819, with an average time interval of 6 days between the formation of each TC. The standard deviation (δ) of the time interval of tropical cyclogenesis was 5.4 days. Similar to the previous work of Gao and Li (2011), we classified the TC activities into three groups: (1) CTC [TC genesis with an interval of less than or equal to 3 days (the approximate value of the mean interval minus half the δ for the genesis interval)]; (2) non-clustering TCs [TC genesis with an interval of greater than or equal to 9 days (approximately the mean interval plus half the δ of the genesis interval)]; and (3) common TCs (genesis interval of between 3 and 9 days). Note that, if two or more TCs occurred successively within one pentad, they were treated as one CTC. The predictand in the present study was the temporal distribution of CTC.

The daily atmospheric variables were derived from National Centers for Environmental Prediction/Department of Energy (NCEP/DOE) Reanalysis II dataset (Kanamitsu et al., 2002), with a horizontal resolution of 2.5° × 2.5° (latitude × longitude). The daily OLR dataset was obtained from the National Oceanic and Atmospheric Administration’s polar-orbiting satellites (Liebmann and Smith 1996). A 5-day mean was applied to all the daily data to obtain the pentad-mean data. The data period spanned from 1979 to 2014. To focus on large-scale predictors, all the reanalysis data were interpolated to a 5° × 5° (latitude × longitude) horizontal resolution.

2.2 CTC index

Because CTC is not a temporally continuous phenomenon (i.e., in the context of statistical modeling, a non-continuous predictand), it is hard to predict. A priority, therefore, was to define a continuous index that could successfully represent the temporal distribution of CTC. Since CTC generates mostly over the tropical monsoon trough, the ISO of the monsoon trough is a good precursor for CTC (Cao et al., 2012). To examine which variables in which ISO band over the monsoon trough could best represent CTC, we first calculated the δ of 10–30 day, 10–60 day, 10–80 day and 30–80 day band filtered ISO signal of the 850 hPa zonal wind, 850 hPa vorticity, and OLR field averaged over the tropical monsoon trough area at (10°–30° N, 100°–180° E). Then, we calculated the ratio difference between the composite CTC number during above +0.5 δ and below −0.5 δ of the variable, as illustrated in Fig. 1. If the absolute value of the ratio difference was larger in positive sign, more CTC events were likely to occur during the positive phase of the variable. On the contrary, if the absolute value of the ratio difference was larger in negative sign, more CTC events were likely to occur in the negative phase of the variable. Finally, the normalized areal-mean variable over the maximum ratio difference region (exceeding a certain threshold) was employed to define the CTC index. Note that, in order to obtain the highest percentage of CTC events during the active phase, we slightly adjusted the threshold of ratio difference for different variables and bands. As listed in Table 1, among all variables, the OLR had the largest percentage of CTC in the active phase of the index. The percentage in the 30–80 day band (71 %) was comparable with that of the 10–30 day band (63 %), indicating that the MJO mode and QBWO mode both played critical roles in modulating the CTC events. Among the different ISO bands of OLR, the largest percentage of 76 % was obtained in the 10–80 day band, with a ratio difference threshold of −0.2. Moreover, the percentage of active phase with CTC events to the total active phase was 38 %, indicating a high possibility of CTC occurrence during the active phase of this CTC index. Thus, the OLR in 10–80 day band was selected to define the CTC index.
Fig. 1

Illustration of how the composite difference was produced. Shading indicates the period beyond 0.5 δ

Table 1

The performance of each variable (OLR, 850 hPa vorticity, and zonal wind) in representing CTC events in the different ISO bands (10–30, 10–60, 10–80 and 30–80 day)

 

OLR

850 hPa vorticity

850 hPa zonal wind

 

10–30 day

10–60 day

10–80 day

30–80 day

10–30 day

10–60 day

10–80 day>

30–80 day

10–30 day

10–60 day

10–80 day

30–80 day

Threshold of ratio difference

<−0.2

<−0.2

<−0.2

<−0.2

>0.12

>0.15

>0.15

>0.2

>0.25

>0.25

>0.25

>0.25

Percentage of CTC events in active phase (%)

63

73

76

71

58

66

69

67

59

68

72

69

Percentage of active phases with CTC (%)

20

29

38

30

22

24

25

24

19

24

30

23

Figure 2 shows the ratio difference of CTC between the above +0.5 δ and below −0.5 δ of the 10–80 day band OLR. The largest ratio difference appeared between 5° N and 20° N in the South China Sea, across the Philippine Sea to 160° E. The grids marked with red dots were below the threshold of the ratio difference of −0.2. Therefore, the normalized areal-mean 10–80 day intraseasonal OLR over these red dotted grids were used for calculating the CTC index. We plotted Fig. 3 to directly show to what extent the CTC index was able to represent the temporal distribution of CTC events during each summer. For the training period 1979–2002, out of the total 168 CTC events, there were 133 (79 % of the total) CTC events in the active (negative) phase of the CTC index, suggesting that the CTC index was able to represent the temporal distribution of CTC events well.
Fig. 2

Ratio differences of the CTC during above +0.5 δ to below −0.5 δ. The red dots below −0.2 were used to calculate the CTC index

Fig. 3

Pentad temporal distribution (pentad 31 to pentad 61) of CTC genesis (bars) and the CTC index based on OLR (red line) for the period 1979–2002

2.3 ISO signal extraction

For the extended-range forecasting of intraseasonal variability, data filtering is an essential aspect in real-time application. Traditional band-pass filtering has a tapering problem and therefore cannot be used for real-time forecasting. Thus, we applied an “unconventional filtering” method (Hsu et al. 2015; Zhu et al. 2015), which is an alternate approach to extracting the 10–30 day, 30–80 day and 10–80 day band ISO signals, respectively. Taking the 10–80 day band as an example, firstly, because the daily data had already been transferred into pentad-mean data, the high-frequency signal (shorter than 10 days, e.g., synoptic-scale perturbation) was removed. Secondly, the climatological annual cycle was removed from these pentad-mean datasets by subtracting a climatological 18-pentad (90-day) low-pass filtered component. Thirdly, a last 8-pentad (e.g., from day −40 to day 0) running mean was removed from the anomaly field above, and therefore, the low-frequency signal of longer than 80 days was removed. Following these three steps, the 10–80 day ISO signal was obtained. Given that no future data (after day 0) were used, this approach can be applied in real-time operation. Note that the present paper also highlights the forecasting of the 10–30 day mode ISO. However, it was obviously unsuitable to use the pentad-mean data to make a prediction of 10–30 day oscillation. To resolve this problem, we used 2.5-day temporal resolution data to make the forecast. These data were calculated by the mean value of the two neighboring sets of pentad data. The “interpolated” 2.5-day resolution data were skipped when assessing the forecast skill. Therefore, the forecast skills at 5, 10, 15, 20, 25 and 30-day lead times are presented, regardless of using the pentad or 2.5-day (half pentad) temporal resolution data.

3 Model construction

3.1 Introduction to the STPMs

The STPMs were applied to the extended-range forecasting of WNP CTC events. The models were based on the extended singularity value decomposition (E-SVD) method. As is well known, SVD can extract the coupled modes of two fields. Therefore, an STPM based on E-SVD can best capture the temporally evolving coupled modes of the prior predictor and the following CTC index. For example, during the 24-year training period (1979–2002), as illustrated by the Eqs. (1) and (2) below, the large-scale predictor (and predictand) involved a data matrix X (Y) with t dimension points in time and i 1 × j 1 × n 1 (1 × n 2) points in space, in which i 1 (1) and j 1 (1) are the spatial grids and n 1 (n 2) is the number of preceding (succeeding) pentads corresponding to the forecast time point t.
$$ X\left(t,{i}_1\times {j}_1\times {n}_1\right)\approx {\displaystyle \sum_{k=1}^K{V}_k\left({i}_1\times {j}_1\times {n}_1\right)}{v}_k(t) $$
(1)
$$ Y\left(t,1\times {n}_2\right)\approx {\displaystyle \sum_{k=1}^K{U}_k\left(1\times {n}_2\right)}{u}_k(t) $$
(2)

Although K coupled modes in total were derived during the training period, only the persistent and useful modes were retained through the cross-validation (leave 1 year out) procedure. The cross-validation procedure involved the following three steps: (1) leaving 1 year of data (e.g., 31 pentads in the present study) and conducting the E-SVD analysis using the remaining 23 years of training data; (2) projecting the one left-out year’s predictor and predictand fields onto the corresponding predictor–predictand E-SVD modes derived from the remaining years, yielding a pair of 1-year projected expansion coefficients (31 pentad time series) for each of the K E-SVD modes; (3) repeating step (1) and (2) 23 times by leaving out different one-year periods of data, combining the projected expansion coefficients for the total of 24 years (31 × 24 pentads) for each coupled mode, and retaining m coupled modes with significant correlation coefficients (99 % confidence level) between their projected expansion coefficients.

Due to the high correlation coefficients of two expansion coefficients, as indicated by Eq. (3), the independent forecasting (from time point t f ) could be made simply by multiplying the reproduced expansion coefficient of predictors (v m ) by the corresponding CTC index modes (U m ). The reproduced expansion coefficient was obtained by projecting the real-time temporally varying predictor field (X) onto the retained predictor modes (V m ), as expressed by Eq. (4).
$$ Y\left({t}_f,1\times {n}_2\right)\approx {\displaystyle \sum_{m=1}^M{U}_m\left(1\times {n}_2\right)\cdot }{v}_m\left({t}_f\right) $$
(3)
$$ {v}_m\left({t}_f\right)={\displaystyle \sum_{k=1}^{i_1\times {j}_1\times {n}_1}X\left({t}_f,{i}_1\times {j}_1\times {n}_1\right)\times }{V}_m\left({i}_1\times {j}_1\times {n}_1\right) $$
(4)

The OLR, zonal wind at 850 and 200 hPa (U850, U200), and geopotential height at 850, 500 and 200 hPa (H850, H500, H200) were used as potential predictors. Besides, vorticity at 850 hPa (Curl850) and relative humility at 700 hPa (Rhum700) were also used as potential predictors, given that the convergence of air moisture plays a role in ISO propagation (Hsu and Li 2012). Additionally, the CTC index itself was also selected as a potential predictor. The optimal ensemble was applied by selecting the best five predictors to make the ensemble forecast. Readers are referred to Zhu et al. (2015) for further detail regarding STPM construction.

Two metrics—the temporal correlation coefficient (TCC) and root-mean-square error (RMSE)—were employed to evaluate the models’ forecast performances.

3.2 Predictability sources

Before running the models, it was critical to search for the previous predictability sources of CTC events. By calculating and producing correlation coefficient maps of the previous circulation field against the CTC index, we were able to detect the predictability sources for each predictor and confirm their projection domains for each STPM. Figure 4 shows the predictability sources from the 35- to 5-day lead time for each predictor. Taking the predictor of U850 as an example, at the 35-day lead time, a northwest–southeast tilted dipole pattern appeared, with positive anomalies over the tropical Indian Ocean/Maritime Continent and negative anomalies from eastern Asia and the central/eastern Pacific to South America. From the 30- to 5-day lead time, a dipole anomaly signal gradually propagated eastward and northward. At the 5-day lead time, positive anomalies appeared over northern India, across southern China, and over the center/eastern Pacific to Central America; whereas, negative anomalies occurred over the Arabian Sea, Bay of Bengal, and Maritime Continent to the western South Pacific. Therefore, an anti-cyclonic anomaly was able to dominate the WNP in the positive phase of the CTC index. Correspondingly, the OLR, Rhum700 and Curl850 all showed consistent patterns. Note that pronounced anomaly signals of the above variables were primarily confined to the tropical western Pacific and Indian Ocean. However, the U200, H850, H500 and H200 demonstrated pronounced signals expanding to the global tropics. Compared with the predictability source from the other four predictors, the predictability sources of U200, H850, H500 and H200 were almost symmetric about the equator. Therefore, as listed in Table 2, a projection domain over the global tropics from 15° S to 15° N was chosen for H850, H500, H200 and U200; whereas, the projection domain for U850, OLR, Rhum700 and Curl850 was selected as (15° S–20° N, 60°–160° E). Note that the projection domain could not be too large (extracting the coupled SVD modes would have involved noise) or too small (some useful signals might have been excluded). This was an important principle for the choice of projection domain for each variable.
Fig. 4

The 35–5-day lead correlation pattern of each variable against the CTC index. Shading indicates statistical significance exceeding the 90 % confidence level. The red dashed line indicates the ISO signal’s propagation

Table 2

The projection domain for each predictor

Predictor

Domain

Grids

OLR

(15° S–20° N, 60°–160° E)

8 × 21

U850

(15° S–20° N, 60°–160° E)

8 × 21

U200

(15° S–15° N, 180° W–180° E)

7 × 72

H850

(15° S–15° N, 180° W–180° E)

7 × 72

H500

(15° S–15° N, 180° W–180° E)

7 × 72

H200

(15° S–15° N, 180° W–180° E)

7 × 72

Rhum700

(15° S–20° N, 60°–160° E)

8 × 21

Curl850

(15° S–20° N, 60°–160° E)

8 × 21

CTC index

Itself

1 × 1

4 Model performance

The ISO is comprised of the QBWO (10–30 day) mode and the MJO (30–80 day) mode, and the WNP TC genesis is strongly modulated by these two modes (Gao and Li 2011; Wen et al. 2012). Figure 5 shows the power spectra of the CTC index during the training period 1979–2002. As we can see, the CTC index featured these two independent ISO modes, with one peak at around 40–60 days and the other at around 15–30 days. As indicated in our previous work (Zhu and Li 2016), although the predictive skill of the STPM for summer rainfall anomalies over China can reach a lead time of 25 days, the 10–30 day mode (e.g., the QBWO) of the rainfall signal is difficult to reproduce. In the present study, we attempted to solve this deficiency. Using different model structures and temporal resolutions, we ran several STPMs that highlighted different ISO modes. The STPMs could be classified into three groups: the first group used the 10–80 day band datasets and covered the entire ISO signal; the second group was based on the 10–30 day band dataset and focused mainly on the QBWO mode of the ISO signal; and the third group was for the 30–80 day band signal, which highlighted the MJO mode. Each group was then further divided into four types, based on different model structures. The first type used the preceding six pentads’ predictors to forecast the succeeding six pentads’ predictand. To produce a 5–30-day lead time (six pentads) forecast, only one set of this type of STPM was needed. We named this type of STPM 10-80d_(6-6). The second type used the preceding three pentads’ predictors to forecast the succeeding three pentads’ predictand. To produce a 5–30-day lead time (six pentads) forecast, two sets of this type of STPM were required. Using the same three pentads’ predictors, one set of this type of STPM forecasted the first three pentads predictand, and another set forecasted the second three pentads predictand. This type of STPM was named 10-80d_(3-3). The third type used the preceding three half-pentads (2.5-day temporal resolution) predictors to forecast the succeeding three half-pentads predictand. To produce a 5–30-day lead time (six pentads) forecast, three sets of STPMs were needed. The first set aimed for forecasts with 5-, 7,5-, 10-day lead times; the second set for forecasts with 15-, 17.5-, 20-day lead times; and the third set for 25-, 27.5-, 30-day lead times. This type of STPM was named 10-80d_(3-3)_2.5d. The fourth type used the preceding two pentads predictors to forecast the succeeding two pentads predictand. This required three sets of STPMs to produce a 5–30-day lead time forecast and was named 10-80d_(2-2). Similarly, the second group (based on 10–30 day band data) and third group (based on 30–80 day band data) of STPMs were run in the same manner. Detailed information regarding each type of STPM is provided in Table 3.
Fig. 5

Power spectrum of the observed CTC index for the period 1979–2014. The red dashed line is the threshold of the 95 % confidence level

Table 3

TCC and RMSE (in brackets) skill for the different types of STPM

ISO band

Model structure

Sets of STPMs

Temporal resolution

5-day lead TCC (RMSE)

10-day lead TCC (RMSE)

15-day lead TCC (RMSE)

20-day lead TCC (RMSE)

25-day lead TCC (RMSE)

30-day lead TCC (RMSE)

5–30-day lead weighted skilla

10–80 day

6–6

1

5 day

0.50 (0.97)

0.45 (1.02)

0.43 (1.04)

0.40 (1.07)

0.34 (1.13)

0.27 (1.18)

0.36 (1.10)

3–3

2

5 day

0.53 (0.95)

0.42 (1.06)

0.39 (1.08)

0.38 (1.09)

0.36 (1.12)

0.30 (1.16)

0.36 (1.11)

3–3

3

2.5 day

0.52 (0.94)

0.39 (1.04)

0.36 (1.08)

0.37 (1.09)

0.36 (1.12)

0.31 (1.15)

0.36 (1.10)

2–2

3

5 day

0.52 (0.95)

0.38 (1.05)

0.35 (1.08)

0.36 (1.10)

0.34 (1.13)

0.29 (1.17)

0.34 (1.11)

10–30 day

6–6

1

5 day

0.43 (1.00)

0.36 (1.06)

0.32 (1.10)

0.27 (1.14)

0.24 (1.18)

0.23 (1.17)

0.27 (1.14)

3–3

2

5 day

0.55 (0.89)

0.41 (1.01)

0.32 (1.06)

0.24 (1.15)

0.23 (1.17)

0.22 (1.12)

0.27 (1.11)

3–3

3

2.5 day

0.59 (0.84)

0.47 (0.93)

0.32 (1.11)

0.24 (1.15)

0.23 (1.16)

0.22 (1.13)

0.28 (1.11)

2–2

3

5 day

0.56 (0.83)

0.45 (0.95)

0.32 (1.12)

0.23 (1.17)

0.23 (1.17)

0.23 (1.18)

0.28 (1.13)

30–80 day

6–6

1

5 day

0.66 (0.80)

0.58 (0.90)

0.55 (0.94)

0.52 (0.97)

0.45 (1.04)

0.36 (1.13)

0.40 (1.01)

3–3

2

5 day

0.71 (0.74)

0.62 (0.84)

0.54 (0.93)

0.54 (0.94)

0.51 (0.98)

0.43 (1.06)

0.45 (0.96)

3–3

3

2.5 day

0.65 (0.81)

0.53 (0.94)

0.54 (0.93)

0.53 (0.93)

0.51 (0.97)

0.45 (1.03)

0.44 (0.96)

2–2

3

5 day

0.68 (0.72)

0.58 (0.80)

0.54 (0.97)

0.53 (0.10)

0.47 (1.04)

0.39 (1.12)

0.45 (1.01)

10–80 day

Best 10–30 day plus best 30–80 day

0.56 (0.82)

0.42 (0.91)

0.41 (0.95)

0.41 (1.01)

0.40 (1.03)

0.34 (1.06)

0.39 (1.00)

a5–30-day weighted skill formula: 5d × 1/21 + 10d × 2/21 + 15d × 3/21 + 20d × 4/21 + 25d × 5/21 + 30d × 6/21

Based on Bretherton et al. (1999), the effective sample size (Ne) is estimated as Ne = N (1 − r12)/(1 + r12), where r1 is the lag (−1) auto-correlation and N is the original sample size, which was 372 [31 pentads per year over 12 years (2003–2014)] in our case. The lag (−1) autocorrelation (r1) was around 0.7 for the 30–80 day band CTC index, 0.44 for the 10–80 day band, and 0.26 for the 10–30 day band. Therefore, the effective sample sizes were 127, 251, and 325, respectively. Based on the effective sample size, a TCC of above 0.23, 0.16, and 0.14 could be considered as a useful skill threshold (significant at the 99 % confidence level) for the 30–80 day, 10–80 day and 10–30 day bands, respectively.

Table 3 lists the TCC and RMSE skills for the different STPMs. Note that the TCC skills for the 5–30-day lead time forecasts of the different STPMs all exceeded the 99 % confidence level, suggesting useful forecast skills of up to 30 days in advance could be achieved. For the 10–80 day band STPM group, the predictive skills (both TCC and RMSE) differed little among the different types of STPM in this group. From the lead time of 5 to 30 days, the TCC decreased from about 0.52 to about 0.30, whereas the RMSE increased from around 0.95 to around 1.16. The weighted skills for the 5–30-day lead time (based on the 5–30-day lead weighted formula shown at the foot of the table) were around 0.36 (for the TCC) and 1.10 (for the RMSE). The predictive skill was therefore very encouraging. However, when we calculated the power spectra of the forecasted CTC index, as shown in Fig. 6, it was found that the 10-80d_(6-6) STPM could not capture the 10–30 day mode of ISO, regardless of the 10- or 20-day lead forecast. The other types of 10–80 day bands all showed similar results (figures not shown). The double peaks for the observed CTC index were replaced by a single peak for the 10- and 20-day lead forecasted CTC index.
Fig. 6

Power spectrum of the observed (left panels) and 10-day (middle panels) and 20-day (right panels) lead forecasted CTC index produced by different STPMs for the period 2003–2014. The red dashed line is the threshold of the 95 % confidence level

For the 10–30 day band STPM group, the 10-30d_(3-3) and 10-30d_(2-2) STPMs showed comparable forecast skill, both in terms of their TCCs and RMSEs. The 10-30d_(6-6) STPMs demonstrated relatively lower skill in terms of their TCCs and RMSEs for 5–15-day lead times but higher skill after a lead time of 20 days. This indicated that the model structure did indeed influence the model skill for different lead times. The more pentads the model concatenated, the higher the skill it was able to obtain at longer lead times. The 10-30d_(3-3)_2.5d STPM showed the best skill, with a relatively higher weighted TCC (0.28) and lower weighted RMSE (0.11) (Table 3). Furthermore, aside from the relatively higher TCC and RMSE skill, the power spectra in Fig. 6 showed that the 10-30d_(3-3)_2.5d STPM was able to reproduce the 10–30 day peak, while the other two types of 10–30 day STPM could not; instead, they reproduced an unrealistic 40–60 day peak, even at a lead time of 10 days. For the 30–80 day STPM group, the 30-80d_(3-3) STPM showed the best skill, with a weighted TCC of 0.45 and a weighted RMSE of 0.96. All the STPMs in this group reproduced similar 30–80 day peaks as the 30-80d_(3-3) STPM as shown in Fig. 6.

Based on the power spectrum of the CTC index forecasted by the three groups of STPMs, it is clear that the 10–80 day STPM was unable to reproduce the 10–30 day mode of the ISO signal; whereas only the 10–30 day band STPM with 2.5-day temporal resolution could reasonably reproduce the 10–30 day mode of the ISO signal with a relatively higher TCC and lower RMSE. Therefore, for the purpose of predicting both 30–80 day and 10–30 day band signals of the CTC index, our strategy was to combine the forecast outputs from the best STPMs in the 10–30 day and 30–80 day groups to produce the final forecast. As shown in Fig. 7, by simply summing the outputs from the 10-30d_(3-3)_2.5d and 30-80d_(3-3) STPMs, the combined results reproduced quite realistic 10–30 day and 30–80 day modes of the CTC index, at least 10 days in advance. And as indicated in Table 3, compared with the 10–80 day band STPM, the weighted TCC of the combined outputs did indeed increase, and the weighted RMSE reduced, suggesting a beneficial outcome from this combination strategy.
Fig. 7

As in Fig. 6 but for the combined outputs of the 10-30d_(3-3)_2.5d and 30-80d_(3-3) STPMs

Figure 8 shows the 10- and 20-day lead forecasts along with the observed CTC index for the independent forecast period from 2003 to 2014. For the 10-day lead time, we can see clearly that the forecasted CTC index from the combined outputs (blue) was much closer to the observation, while the CTC index forecasted by the 10-80d_(6-6) STPM (red) seemed smoother, missing the higher frequency of the QBWO mode signal. For the 20-day lead time, the forecasted CTC from the combined outputs also missed some of the 10–30 day mode of the ISO signal, but the TCC and RMSE still showed better skill compared with the forecasted result from the 10-80d_(6-6) STPM.
Fig. 8

The observed (black) and 10-day lead (upper panels) and 20-day lead (lower panels) forecasted CTC index produced by the 10-80d_(6-6) STPM (red) and combined outputs (blue) of the 10-30d_(3-3)_2.5d and 30-80d_(3-3) STPMs for the period 2003–2014

Using the forecasted CTC index, we were able to calculate the forecast skill for the temporal distribution of CTC events. Table 4 shows the “hits,” “false alarms,” and “misses” for the forecasted CTC from the combined STPM outputs. The hits indicate the number of CTC events that occurred in the active phase of the forecasted CTC index; the false alarms indicate the number of CTC events when the forecasted CTC index remained in its negative phase for two or more pentads but no CTC occurred, and the misses indicate the number of CTC events that occurred in the inactive phase of the forecasted CTC index. During the independent forecast period, 49 out of 54 observed CTC events occurred in the active phase of the CTC index, suggesting the index was able to represent the CTC events during the independent forecast period well. With increasing lead time, the hits (misses) decreased (increased). The STPM still produced hits for at least around 68 % of the total number of CTC events before the lead time of 20 days; whereas, after this lead time, hits became relatively low. At the lead time of 30 days, the STPM could only produce hits for 57 % of the total number of CTC events. It is worth noting, however, that the number of false alarms remained at around 16 for all the different lead times. The false alarm rate of the STPM was around 30 % of the total number of CTC events, which is less than the ensemble forecast of numerical models (Vitart et al., 2012; Tsai and Elsberry 2013).
Table 4

The hits, false alarms, and misses of the observed and 5–30-day lead forecasted CTC from combined STPM outputs

 

Observed

5-day lead

10-day lead

15-day lead

20-day lead

25-day lead

30-day lead

Hits

49

46

43

39

37

33

31

False alarms

14

15

15

16

17

17

18

Misses

5

8

11

15

17

21

24

5 Conclusion and discussion

In the present study, based on normalized areal-mean OLR, a WNP CTC index was defined. For the period 1979–2002, around 76 % of the total number of CTC events occurred during the active (negative) phase of the CTC index. Therefore, the CTC index was proposed as a continuous predictand which, to a large extent, reflected the temporal distribution of CTC events.

The predictability sources for CTC were detected via lag correlation analyses between the CTC index and previous atmospheric fields at lead times of 35–5-day. The results indicated that the geopotential height at different levels and 200 hPa zonal wind had the predictability sources over the global tropics, whereas the predictability sources of other variables, such as OLR and 850 hPa zonal wind, were mainly confined to the tropical Indian Ocean and western Pacific Ocean.

Using the projection domain of each predictor, several sets of STPM were constructed to carry out extended-range forecasts of the temporal distribution of WNP CTC. Because the 10–80 day band CTC index featured double-peak power spectra at both 15–30 and 30–80 days, three groups of STPM with different model structures and temporal resolutions were designed to forecast different ISO bands (e.g., 10–30 day and 30–80 day bands, respectively). It was found that the STPM in the 10–30 day band with 2.5-day temporal resolution was able to capture the 10–30 day mode of ISO. Based on the combined outputs from the best STPMs in the 10–30 day and 30–80 day bands, for the independent forecast period of 2003–2014, the CTC index was captured well from lead times of 5 to 30 days, with relatively high predictive skill. Both the 10–30 day and 30–80 day mode of the CTC index were reproduced well at a lead time of at least 10 days. The STPM produced hits for 80 and 69 % of the total 54 CTC events in the 10- and 20-day lead time forecasts; plus, the false alarm rate was about 28 and 31 % with respect to the total number of CTC events, suggesting an encouraging skill level for extended-range forecasting.

The STPM for CTC events proposed herein was not designed for an accurate extended-range forecasting of the accurate location, track, and landfall of CTC events. Reliably forecasting the timing of CTC events at an extended range is of great practical value for disaster mitigation. If we can be alerted about upcoming CTC events at longer lead times, more effective measures to minimize damage through early decision-making may be possible, or, on the other hand, we can seek to maximize any benefits stemming from the climate impacts of CTC events.

On a final note, the deficiency of the STPM in capturing the 10–30 day mode of ISO might be solved by using a different STPM structure or a higher temporal resolution with an extra attention paid to the 10–30 day mode of ISO. This may facilitate future improvements in the extended-range forecasting of both the MJO and QBWO modes of ISO.

Notes

Acknowledgments

The authors would like to thank the two anonymous reviewers for their constructive comments. This work was supported by the Natural Science Foundation of Jiangsu Province (BK20140046), the China National 973 project (2015CB453200), the Special Fund for Meteorological Scientific Research of the Public Sector (Grant no. GYHY201306032), the National Natural Science Foundation of China (41575052/41630423), the key project of the Fujian Provincial Department of Science and Technology (2011Y0008), and the PAPD (Priority Academic Program Development) of Jiangsu Higher Education institutions. This paper is SOEST contribution number 9817, IPRC contribution number 1210 and ESMC contribution number 125.

References

  1. Bretherton CS, Widmann M, Dymnikov VP, Wallace JM, Bladé I (1999) The effective number of spatial degrees of freedom of a time-varying field. J Clim 12:1990–2009CrossRefGoogle Scholar
  2. Cao X, Huang P, Chen G, Chen W (2012) Modulation of western North Pacific tropical cyclone genesis by intraseasonal oscillation of the ITCZ: a statistical analysis. Adv Atmos Sci 29(4):744–754CrossRefGoogle Scholar
  3. Elsberry RL, Tsai HC, Jordan MS, Vitart F (2014) Extended-range forecasts of Atlantic tropical cyclone events during 2012 using the ECMWF 32-day ensemble predictions. Weather Forecast doi. doi: 10.1175/WAF-D-13-00104.1 Google Scholar
  4. Frank WM, Roundy PE (2006) The role of tropical waves in tropical cyclogenesis. Mon Wea Rev 134:2397–2417. doi: 10.1175/MWR3204.1 CrossRefGoogle Scholar
  5. Fu B, Li T, Peng M, Weng F (2007) Analysis of tropical cyclogenesis in the western North Pacific for 2000 and 2001. Weather Forecast 22:763–780CrossRefGoogle Scholar
  6. Gao JY, Li T (2011) Factors controlling multiple tropical cyclone events in the western North Pacific. Mon Wea Rev 139:885–894. doi: 10.1175/2010MWR3340.1 CrossRefGoogle Scholar
  7. Gao JY, Li T (2012) Interannual variation of multiple tropical cyclone events in the western North Pacific. Adv Atmos Sci 29(6):1279–1291. doi: 10.1007/s00376-012-1031-1 CrossRefGoogle Scholar
  8. Gray WM (1968) Global view of the origin of tropical disturbances and storms. Mon Wea Rev 96:669–700CrossRefGoogle Scholar
  9. Gray WM (1979) Hurricanes: their formation, structure, and likely role in the tropical circulation. In: Shaw DB (ed) Meteorology over the tropical oceans. Roy Meteor Soc, LondonGoogle Scholar
  10. Hsu PC, Li T (2012) Role of the boundary layer moisture asymmetry in causing the eastward propagation of the Madden–Julian Oscillation. J Clim 25:4914–4931CrossRefGoogle Scholar
  11. Hsu PC, Li T, You LJ, Gao JY, Ren HL (2015) A spatial-temporal projection method for 10-30-day forecast of heavy rainfall in Southern China. Clim Dynam 44:1227–1244CrossRefGoogle Scholar
  12. Kanamitsu M, Bisuzaki EW, Woollen J, Yang SK, Hnilo JJ, Fiorino M, Potter GL (2002) NCEP–DOE AMIP-II reanalysis (R-2. B Am Meteorol Soc 83:1631–1643CrossRefGoogle Scholar
  13. Kim KH, Ho CH, Kim HS, Sui CH, Ki PS (2008) Systematic variation of summertime tropical cyclone activity in the western North Pacific in relation to the Madden–Julian Oscillation. J Clim 21:1171–1191. doi: 10.1175/2007JCLI1493.1 CrossRefGoogle Scholar
  14. Klotzbach PJ, Oliver ECJ (2015) Variations in global tropical cyclone activity and the Madden-Julian Oscillation since the mid-twentieth century. Geophys Res Lett 42. doi: 10.1002/2015GL063966
  15. Ko KC, Hsu HH (2009) ISO modulation on the sub-monthly wave pattern and recurving tropical cyclones in the tropical western North Pacific. J Clim 22:582–599. doi: 10.1175/2008JCLI2282.1 CrossRefGoogle Scholar
  16. Lander MA (1996) Specific tropical cyclone track types and unusual tropical cyclone motions associated with a reverse-oriented monsoon trough in the western North Pacific. Weather Forecast 11:170–186CrossRefGoogle Scholar
  17. Li WK, Hsu PC, He JH, Zhu ZW, Zhang WJ (2016) Extended-range forecast of spring rainfall in southern China based on the Madden–Julian Oscillation. Meteorog Atmos Phys 128:331–345CrossRefGoogle Scholar
  18. Liebmann B, Smith CA (1996) Description of a complete (interpolated) outgoing longwave radiation dataset. B Am Meteorol Soc 77:1275–1277Google Scholar
  19. Liebmann B, Hendon HH, Glick JD (1994) The relationship between tropical cyclones of the western Pacific and Indian Oceans and the Madden-Julian oscillation. J Meteorol Soc Jpn 72:401–412CrossRefGoogle Scholar
  20. Maloney ED, Hartmann DL (2000a) Modulation of eastern North Pacific hurricanes by the Madden–Julian oscillation. J Clim 13:1451–1460. doi: 10.1175/1520-0442(2000)013<1451:MOENPH>2.0.CO;2 CrossRefGoogle Scholar
  21. Maloney ED, Hartmann DL (2000b) Modulation of hurricane activity in the Gulf of Mexico by the Madden-Julian oscillation. Science 287:2002CrossRefGoogle Scholar
  22. Tsai HC, Elsberry RL (2013) Opportunities and challenges for extended-range predictions of tropical cyclone impacts on hydrological predictions. J Hydrometeorol 506:42–54. doi: 10.1016/j.jhydrol.2012.12.025 Google Scholar
  23. Van Den Dool HM, Saha S (1990) Frequency dependence in forecast skill. Mon Wea Rev 118:128–137CrossRefGoogle Scholar
  24. Vitart F, Prates F, Bonet A, Sahin C (2012) New tropical cyclone products on the web. ECMWF Newsletter No 130:17–23Google Scholar
  25. Wen M, Li T, Zhang R, Qi Y (2010) Structure and origin of the quasi-biweekly oscillation over the tropical Indian Ocean in boreal spring. J Atmos Sci 67:1965–1982CrossRefGoogle Scholar
  26. William MF, Roundy PE (2006) The role of tropical waves in tropical cyclogenesis. Mon Wea Rev 134:2397–2417. doi: 10.1175/MWR3204.1 CrossRefGoogle Scholar
  27. Xiang BQ, Lin SJ, Zhao M, Zhang S, Vecchi G, Li T, Jiang X, Harris L, Chen JH (2015) Beyond weather time-scale prediction for Hurricane Sandy and Super Typhoon Haiyan in a global climate model. Mon Wea Rev 143:524–535. doi: 10.1175/MWR-D-14-00227.1 CrossRefGoogle Scholar
  28. Yamaguchi M, Vitart F, Simon T, Linus M, Russell LE, Grant E, Masayuki K, Tetsuo N (2015) Global distribution of the skill of tropical cyclone activity forecasts on short- to medium-range time scales. Wea Forecasting 30:1695–1709CrossRefGoogle Scholar
  29. Zhang C (2013) Madden-Julian oscillation: bridging weather and climate. B Am Meteorol Soc 94:1849–1870CrossRefGoogle Scholar
  30. Zhu ZW, Li T (2016) The statistical extended-range (10-30-day) forecast of summer rainfall anomalies over the entire China. Clim Dynam DOI. doi: 10.1007/s00382-016-3070-2 Google Scholar
  31. Zhu ZW, Li T, Hsu PC, He JH (2015) A spatial-temporal projection model for extended-range forecast in the tropics. Clim Dynam 45:1085–1098CrossRefGoogle Scholar

Copyright information

© Springer-Verlag Wien 2016

Authors and Affiliations

  1. 1.Key Laboratory of Meteorological Disaster, Ministry of Education (KLME)/Joint International Research Laboratory of Climate and Environment Change (ILCEC)/Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters (CIC-FEMD)Nanjing University of Information Science and TechnologyNanjingChina
  2. 2.International Pacific Research Center and Department of Atmospheric SciencesUniversity of Hawaii at ManoaHonoluluUSA
  3. 3.Fujian Climate CenterChina Meteorological AdministrationFuzhouChina

Personalised recommendations