Global Statistical Predictions of Tropical Cyclones Intensity: Regional Contrasts in most Efficient Atmospheric Predictors and Role of air-sea Coupling

We have developed a global statistical model for tropical cyclones (TCs) intensity hindcasts, based on a multiple linear regression that relates TCs characteristics and their surrounding large-scale atmospheric environmental parameters to TCs intensity change. Our results reveal that the skill of TCs intensity predictions strongly depends on the TC region, intensity and phase of the life cycle. The relative importance of the atmospheric predictors also varies regionally. We further show that using climatological rather than daily atmospheric environmental parameters does not significantly degrade the skill, suggesting that the climatological spatial distribution of those predictors constrain TCs intensity more than their temporal variability.

We then explore the effect of accounting for TCs air-sea coupling processes on the intensity hindcast skill. Including oceanic parameters in the linear prediction scheme does not improve the TCs intensity prediction. Given the non-linear nature of the relation between the TC characteristics and TC-induced cooling, we further develop a non-linear statistical prediction scheme based on Artificial Neural Network. In contrast to the linear model, including oceanic variables in the neural network scheme considerably improves the prediction skill, with a similar skill improvement to that of the most skilful large-scale atmospheric parameters. Using a proxy of the upper thermocline depth results in a far better improvement than the commonly used Ocean Heat Content. The consequences of these findings for TCs intensity statistical prediction scheme are discussed.