SOI and Saturated Equivalent Potential Temperature Based Forecast of Australian Region Tropical Cyclone Activity Using a Poisson Regression Model Approach

Poisson regression model forecasts of tropical cyclone genesis in the Australian-southwest Pacific Ocean region (6°-20°S, 105°-170°E) are presented for the 2004/2005 season (November-May). The probable genesis activity of Eastern (Coral Sea) subregion (6°-20°S, 145°-170°E) tropical cyclones is also forecast for the season. Following the conventional definition, tropical cyclone genesis is identified to occur at the time and location where the wind speed first exceeds 34 knots (17.5 ms^-1). For the purposes of this work, cyclone genesis occurrence points have been determined from the Australian-southwest Pacific Ocean region tropical storm data (obtained from the Australian Bureau of Meteorology (BoM) via http://www.bom.gov.au/climate/how/) and binned into monthly (November-May) cells in time and 2° latitude ´ 5° longitude boxes in space across the region (see McDonnell and Holbrook 2004a).

There has been considerable observational and statistical research investigating links between the Southern Oscillation index (SOI) and tropical cyclone frequency in the Australian-southwest Pacific Ocean region (e.g., Solow and Nicholls 1990; Nicholls 1992). Neville Nicholls from BoM has routinely conducted a SOI-based forecast of seasonal tropical cyclone activity, using a linear regression modelling approach on first differences between the October lead SOI and tropical cyclone numbers (e.g., Nicholls 1999). This linear statistical model has been shown to perform very well in the Australian region. Its limitation is that it provides basin-wide, seasonal forecasts.

Extending Nicholls’ earlier work, McDonnell and Holbrook (2004b) have shown that using a Poisson regression model, developed with the September lead SOI as the predictor variable, skilful hindcasts and arguably slightly-improved forecasts of the number of expected occurrences of tropical cyclogenesis in both the Australian-southwest Pacific Ocean region (as a whole) and also in the Coral Sea subregion, are provided for individual seasons. For the purposes of our analyses, the model(s) were developed and cross-validated using observations over the 33-year seasonal record 1960/61-1992/93. The correlation coefficient between the cross-validated hindcasts and observed cyclones formed is 0.5 (significant at the 99% level). The root-mean-square error (rmse) for the model cross-validated hindcasts is a 22% improvement over climatology (i.e., the number of tropical cyclones predicted using only month, latitude and longitude as predictors). Independent forecasts of 10 tropical cyclone seasons 1993/94-2002/03 using this model showed greater overall skill than the linear first differences SOI model (McDonnell and Holbrook 2004b).

McDonnell and Holbrook (2004b) have further shown that by combining spatial and temporal variations of the saturated equivalent potential temperature gradient (EPT) (out of six spatially-varying oceanic and atmospheric predictors, the EPT was by far the most valuable predictor variable) with the space-invariant SOI in a Poisson regression model, additional (and significant) cross-validated tropical cyclone genesis hindcast skill is afforded by the 2° x 5° grid-scale EPT information. The temperature fields at 1000 mb and 500 mb used to calculate the EPT (see McDonnell and Holbrook 2004a) were obtained from the National Center for Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR) Reanalysis Project monthly mean subsets (http://dss.ucar.edu/datasets/ds090.2/). In physical terms, the EPT depends on the vertical temperature profile and is a measure of the potential for cumulonimbus convection from a lapse-rate stability viewpoint. Specifically, McDonnell and Holbrook (2004b) demonstrate that Poisson regression models of tropical cyclogenesis developed with a combination of the September lead SOI and the September lead EPT plus the EPT 5-1 month lead gradient throughout the tropical cyclone season, result in an increased correlation coefficient of 0.6 between cyclogenesis observations and model hindcasts, with a rmse 34% better than climatology. The temporally- and spatially-varying EPT (at the 2° latitude x 5° longitude grid scale) offers the potential for probabilistic forecasting of not only the number of tropical cyclones formed, but also in what part of the season and in what location they might form. Note: the current model configuration is only reliable at the broader subregional (defined as the Western (105°-125°E), Northern (125°-145°E) and Eastern (145°-170°E) subregions) and subseasonal (defined as early (November-December), mid (January-March) and late (April-May) season) scales or larger as defined in McDonnell and Holbrook (2004b).

As the 5-1 month lead gradient reflects the rate of change in the EPT prior to (and throughout) the tropical cyclone season, we do not use it for our pre-season seasonal forecast here. We also note that the September lead SOI Poisson regression coefficient proved to be insignificant when used together with the September lead EPT predictor in the September-only lead combined-predictor Poisson regression model. Hence, we provide our up-front Poisson regression model seasonal forecasts using two models with predictors of (i) the September lead SOI, and (ii) the September lead EPT. Figure 1 shows the observed seasonal number of tropical cyclones formed and the cross-validated hindcasts for the years 1960/61-1992/93. Also shown are independent forecasts (hindcasts) of the 10 tropical cyclone seasons 1993/94-2002/03 using both forecast models. The Poisson regression coefficients and standard errors used for the 2004/05 season forecasts are shown in Table 1.

The monthly-averaged SOI for September 2004 is -2.8. Using our (i) SOI predictor Poisson regression model, we forecast a total of 8 (7.9) tropical cyclones to form in the Australian-southwest Pacific Ocean region during the 2004/2005 season. This is close to the long-term average of 7.4 tropical cyclones per season. Using our (ii) EPT predictor Poisson regression model, we forecast a just-below average total of 7 (6.8) tropical cyclones to form in this same region and season.

For each Poisson regression model forecast, expected occurrences are determined within every 2° latitude ´ 5° longitude cell in the region. McDonnell and Holbrook (2004b) found that the temporal variability of the seasonal number of tropical cyclone genesis occurrences in the Eastern (Coral Sea) subregion, where El Niño-Southern Oscillation (and hence SOI) influences may be expected to be strongest, is hindcast with the most skill. Our Poisson regression models, using predictors of (i) SOI, and (ii) EPT, generates forecasts of 3 (2.8) and 2 (2.4) tropical cyclone genesis occurrences, respectively, for the Coral Sea subregion in the upcoming 2004/2005 season. Note, that for both the entire region and Coral Sea subregion forecasts, the spatial EPT predictor variable causes the Poisson regression models to generate more conservative forecasts of the total number of tropical cyclones that are expected to form across the region in the upcoming season.

The present Poisson regression formulation does not currently provide useful forecasts at the 2° ´ 5° grid scale due to the simple representation of space in the model. Nevertheless, these Poisson regression models provide skilful aggregates of expected occurrences on subregional scales, and in time.

It must be noted that the forecasts reported here are experimental in nature. The reader is advised that the methods and forecasts are subject to future change and improvement.

References:

McDonnell, K. A. and N. J. Holbrook, 2004a: A Poisson regression model of tropical cyclogenesis for the Australian-southwest Pacific Ocean region. Wea. Forecasting, 19, 440-455.

McDonnell, K. A. and N. J. Holbrook, 2004b: A Poisson regression model approach to predicting tropical cyclogenesis in the Australian/southwest Pacific Ocean region using the SOI and saturated equivalent potential temperature gradient as predictors. Geophys. Res. Lett., 31, L20110, doi:10.1029/2004GL020843.

Nicholls, N., 1992: Recent performance of a method for forecasting Australian seasonal tropical cyclone activity, Aust. Meteor. Mag., 40, 105-110.

Nicholls, N., 1999: SOI-based forecast of Australian region tropical cyclone activity, Experimental Long-Lead Forecast Bulletin, 8(4), 71-72.

Table 1. Poisson regression coefficients (β_j) and standard errors (SE) estimated over the period 1960/61-2002/03, for both the September lead SOI model and the September lead EPT model.

Parameter	SOI model β_j	SE	EPT model β_j	SE
Intercept	-4.72	0.08	-4.47	0.08
SOI_sep	0.17	0.06	-	-
EPT_sep	-	-	0.47	0.06
Month	-0.42	0.06	-0.42	0.06
Cyclone year	0.07	0.06	0.08	0.06
Latitude	-0.47	0.06	-0.72	0.07
Longitude	-0.28	0.05	-0.31	0.06

Figure 1. The observed (solid line) and cross-validated hindcast (dashed line) seasonal numbers of Australian-southwest Pacific Ocean tropical cyclones formed during the period 1960/61-199293 using the (a) SOI, and (b) EPT predictor Poisson regression models. Independent tropical cyclone genesis observations (o) for 1993/94-2002/03 are also forecast (x) using both models.