**4. Dynamical climate models**

**3.1. The ENSO indices and linear discriminant analysis model**

240 Recent Developments in Tropical Cyclone Dynamics, Prediction, and Detection

The Southern Oscillation Index (SOI) and sea surface temperature anomalies (SSTAs) in Niño3.4 and Niño4 regions (NIÑO3.4 and NIÑO4) are commonly used indices in defining

It has been demonstrated by Kuleshov et al. [2, 7] and Ramsay et al. [14] that in the Australian region a strong correlation (about −0.7) exists between the annual number of TCs and the NIÑO4 and NIÑO3.4 indices averaged over 3 months preceding the onset of a Southern Hemisphere TC season (August-September-October). In the eastern South Pacific Ocean, better correlation of the annual TC number with ENSO indices was found for the NIÑO3.4 and the SOI [7]. Based on these findings, the NIÑO3.4 and the SOI indices have been selected by the Australian Bureau of Meteorology for use in operational LDA statistical TC-ENSO model for

A multivariate ENSO index has been developed at the NCC with the aim to integrate both atmospheric and oceanic responses in one index [2]. It is based on the first principal component of monthly Darwin mean sea level pressure (MSLP), Tahiti MSLP and the NIÑO3, NIÑO3.4 and NIÑO4 SST indices [2, 7]. This standardised monthly anomaly index is usually denoted as the 5VAR index. Further examining correlation of ENSO indices with TC occurrences in the Australian region, Kuleshov et al. [15] found that the 5VAR performs better than the SOI and NIÑO3.4 demonstrating the strongest monthly (−0.67, pre-season September), bi-monthly (−0.67, August and September) and tri-monthly correlation (−0.66, July, August and Septem-

Incorporating into statistical model a decreasing trend in TC activity over the Australian region in recent years [2, 16] and using 5VAR, SOI and NIÑO3.4 indices as predictors, Kuleshov et al. [15] demonstrated potential for improving skill of the LDA operational model. Brief description of the developed statistical model is presented in the Appendix. Cross-validation employed to assess the models' performance demonstrated that the models which used the preseason July-August-September SOI and September 5VAR indices and the time trend as the predictors [15] demonstrated increased skill in TC seasonal forecasting compared with

Recently, application of advanced statistical methodologies for seasonal prediction of TCs has been explored. It has been demonstrated that improvement in prediction skill compared to the LDA model can be achieved using support vector regression (SVR) models, exploring new environmental indicators and non-linear relationships between them [9]. Detailed description of the developed SVR models for the Australian and South Pacific Ocean regions could be found in [17] and its brief description is presented in the Appendix. Analysis of the results of the SVR models shows that the Dipole Mode Index, the 5VAR index and the SOI are the most frequently used indices selected for TC seasonal forecasting in the Australian and South Pacific

ENSO phases which describe oceanic and atmospheric responses, respectively.

seasonal prediction of TCs in both the Australian and the South Pacific regions.

ber).

regions.

currently used LDA model [7].

**3.2. Support vector regression (SVR) models**

Dynamical climate modelling is an alternative to statistical modelling. Early analysis has revealed that the dynamical seasonal prediction system Predictive Ocean Atmosphere Model for Australia (POAMA) has skill in the prediction of ENSO which modulates TC activity in the SH [2].

As a part of the 'Climate Change and Southern Hemisphere Tropical Cyclones' International Initiative, an evaluation of performance of dynamical climate models for TC seasonal prediction was conducted under the PACCSAP program. The Australian Bureau of Meteorology and the Japan Meteorological Agency/Meteorological Research Institute (JMA/MRI) have developed systems to provide predictions of TC activity based on their dynamical models.

The two agencies each have their own coupled seasonal forecast model comprised of a number of ensemble members and a 30-year hindcast period. The JMA/MRI-CGCM is used by the JMA/ MRI; the Bureau uses POAMA. Each agency has employed a different TC identification and tracking procedure to determine the number of TCs produced by their model within each ensemble member in each year of the hindcast.

In this section, the ability of each model to produce an environment consistent with observations is examined in the context of environmental parameters related to TC genesis. The TC tracking methods are presented and their performance when applied to the respective model hindcast ensembles is evaluated.

#### **4.1. A comparison of dynamical seasonal tropical cyclone predictions for the Australian and Western Pacific regions**

#### *4.1.1. TC tracking in POAMA*

Dynamical model POAMA [18] is comprised of 30-member ensemble and 31-year hindcast (1980–2010). Realisations initialised on 1 October each year (i.e. prior to start of the Southern Hemisphere TC season) are used to evaluate model's performance. Each realisation provides 9 months of daily global atmospheric environmental fields at approximately 2.5˚ × 2.5˚ resolution.

TCs are identified and tracked using 'Okubo-Weiss-Zeta Parameter' (OWZP) scheme [19]. In brief, regions of low deformation vorticity (large OWZP) at 850- and 500-hPa levels which are vertically coherent and are sustained for an appreciable duration (at least 48 hours) are identified. Where such regions occur in presence of small vertical wind shear and large lower tropospheric humidity, local environment is considered conducive to imminent TC genesis or TC maintenance.

OWZP scheme is applied to 9-month daily POAMA data for each ensemble member and year individually. Statistics for TC-like disturbances for each member are then averaged together to give ensemble mean statistics for each year.

#### *4.1.2. TC tracking in JMA/MRI-CGCM*

The seasonal JMA/MRI-CGCM [20] is comprised of a 10-member ensemble, and the same 31 year hindcast period as POAMA is utilised in this analysis. The realisation for the forecast period beginning 1 November each year is used; each realisation provides daily global atmospheric environmental fields covering the November-April period at approximately 1.875˚ × 1.875˚ resolution.

TCs are identified and tracked using a method similar to that outlined by [21], whereby grid points with 850 hPa relative vorticity less than a threshold value are identified and the sealevel pressure minimum within the surrounding grid points is denoted as the centre of the possible TC. A warm-core is required of these possible TCs: the thickness between 500 and 200 hPa must exceed that of the local environment, and wind speed at 850 hPa must be greater than that at 200 hPa. A full description of this method can be found in [22].

For a possible TC to be considered, it must last longer than 2 days and be equatorward of 30˚S.

This scheme is applied to daily JMA/MRI-CGCM data for the November-April period for each ensemble member and year individually. Statistics for TC-like disturbances for each member are then averaged together to give ensemble mean statistics for each year.

The TC identification and tracking method is basin dependent. The method is applied to the global atmospheric fields using a variety of 850 hPa relative vorticity thresholds. For each basin, the observed climatological number of TCs is compared to the ensemble mean hindcast climatological value for each low-level vorticity threshold; the closest value is then selected for each basin.

#### **4.2. Model environment**

The ability of the models to represent the large-scale environment in which the TCs form has been demonstrated for 850 hPa relative vorticity vertical and troposphere-deep (850–200 hPa) vertical wind shear [23]. The results discussed here use a threshold value of 4.5 × 10−5 s−1 for the Australian region and 7.5 × 10−5 s−1 for the South Pacific region.

During November-December-January (NDJ), both models realistically capture the variability in low-level vorticity near the equator in the western Pacific with correlation values exceeding 0.8 is places. In the tropical Australian region the models do much less well.

The model drift associated with longer lead times is clearly evident in the February-March-April (FMA) season plots, with correlation values of low-level vorticity reduced from NDJ for both models.

Similar statements apply for both models in terms of the vertical wind shear.

#### **4.3. Seasonal TC prediction**

Ensemble mean variability in the number of TCs in the SH TC season (November-April; NDJFMA) is shown in **Figures 8** and **9**, along with the observed number of TCs from the BoM SH TC dataset [7].

*4.1.2. TC tracking in JMA/MRI-CGCM*

242 Recent Developments in Tropical Cyclone Dynamics, Prediction, and Detection

1.875˚ × 1.875˚ resolution.

each basin.

both models.

SH TC dataset [7].

**4.3. Seasonal TC prediction**

**4.2. Model environment**

The seasonal JMA/MRI-CGCM [20] is comprised of a 10-member ensemble, and the same 31 year hindcast period as POAMA is utilised in this analysis. The realisation for the forecast period beginning 1 November each year is used; each realisation provides daily global atmospheric environmental fields covering the November-April period at approximately

TCs are identified and tracked using a method similar to that outlined by [21], whereby grid points with 850 hPa relative vorticity less than a threshold value are identified and the sealevel pressure minimum within the surrounding grid points is denoted as the centre of the possible TC. A warm-core is required of these possible TCs: the thickness between 500 and 200 hPa must exceed that of the local environment, and wind speed at 850 hPa must be greater

For a possible TC to be considered, it must last longer than 2 days and be equatorward of 30˚S. This scheme is applied to daily JMA/MRI-CGCM data for the November-April period for each ensemble member and year individually. Statistics for TC-like disturbances for each member

The TC identification and tracking method is basin dependent. The method is applied to the global atmospheric fields using a variety of 850 hPa relative vorticity thresholds. For each basin, the observed climatological number of TCs is compared to the ensemble mean hindcast climatological value for each low-level vorticity threshold; the closest value is then selected for

The ability of the models to represent the large-scale environment in which the TCs form has been demonstrated for 850 hPa relative vorticity vertical and troposphere-deep (850–200 hPa) vertical wind shear [23]. The results discussed here use a threshold value of 4.5 × 10−5 s−1 for

During November-December-January (NDJ), both models realistically capture the variability in low-level vorticity near the equator in the western Pacific with correlation values exceeding

The model drift associated with longer lead times is clearly evident in the February-March-April (FMA) season plots, with correlation values of low-level vorticity reduced from NDJ for

Ensemble mean variability in the number of TCs in the SH TC season (November-April; NDJFMA) is shown in **Figures 8** and **9**, along with the observed number of TCs from the BoM

than that at 200 hPa. A full description of this method can be found in [22].

are then averaged together to give ensemble mean statistics for each year.

the Australian region and 7.5 × 10−5 s−1 for the South Pacific region.

0.8 is places. In the tropical Australian region the models do much less well.

Similar statements apply for both models in terms of the vertical wind shear.

**Figure 8.** Time series of annual (NDJFMA) number of TCs in POAMA for the Australian (top panel) and South Pacific (bottom panel) regions. TC numbers are shown for observations (solid black line) and ensemble mean (dashed grey line). Ensemble members are shown as coloured circles.

**Figure 9.** Time series of annual (NDJFMA) number of TCs in JMA/MRI-CGCM for the Australian (top panel) and South Pacific (bottom panel) regions. TC numbers are shown for observations (solid black line) and ensemble mean (dashed grey line). Ensemble members are shown as grey circles.

In the Australian region POAMA underestimates the number of TCs throughout the hindcast period, suggesting a deficiency in the model's ability to produce TC-like disturbances in this region. This is not the case in the western South Pacific. In both basins, some of the inter-annual variability is captured by POAMA, yielding a correlation values with observations of ~0.55.

**Figure 10.** Climatological number of TCs as a function of month for observations (solid black line), POAMA (blue dashed line) and JMA/MRI-CGCM (red dashed line) in the Australian (top panel) and South Pacific Ocean (bottom panel) regions.

By design JMA/MRI-CGCM yields annual totals of TCs for NDJFMA close to climatology for both basins and in the Australian region a similar degree of variability to POAMA is captured, demonstrated by a correlation value of 0.48. In the South Pacific JMA/MRI-CGCM fairs less well at capturing the variability.

Ensemble mean monthly TC climatologies for each basin and model are shown compared with observations in **Figure 10**.

Both models capture the monthly variability in the Australian region well, although neither model represents the peak value correctly. In the South Pacific, POAMA performs well, however, JMA/MRI-CGCM peaks a month too early and drops off too quickly.

In summary, POAMA and JMA/MRI-CGCM both represent the large-scale environment relevant to TCs reasonably well, although possible deficiencies exist in the Australian region. The monthly TC climatologies in both models are reasonably realistic. Both models capture some of the inter-annual variability in the Australian region, although POAMA performs better in the South Pacific. Probabilistic NDJFMA TC number predictions both models, evaluated over the 31-year hindcast, show skill over random chance.

With further development of dynamical climate models and improving of their skill it is expected that both statistical and dynamical models will be used in operational TC seasonal prediction in the Australian and South Pacific regions, to complement each other.
