Development of Statistical-dynamical Tools for Tropical Cyclone Intensity Prediction in the Southwest Indian Ocean

Despite a great deal of effort spent over the last decade in research and operational modeling perspectives, intensification processes are not yet sufficiently understood to allow improvements in intensity forecasts to equal those in track forecasts at 24-72 h lead time (DeMaria et al. 2014).
At La Réunion RSMC, operational forecast errors computed during the 2001-2016 period for tropical systems over the southwest Indian Ocean (SWIO) indicate a mean position error of about 120 km at 24 h with a downward time trend. Operational intensity forecast errors are significantly larger for RI events at short lead times (10.8 m/s versus 4.9 m/s for non-RI events at 24 h); however, RI events (intensity increase >= 30 kt in 24 h) are characterized by significantly lower track errors at 12-, 24- and 48-h lead time than non-RI events, possibly due to a better location of the storm center at high intensities.

To offer further guidance to the SWIO practical intensity forecasts and better anticipate rapid intensity changes, the dominant large-scale factors governing the intensity changes of tropical systems in the SWIO have been identified. The predictors have been identified using both Era-Interim reanalysis fields at 0.75° latitude-longitude resolution and the final analysis of best-track data produced by La Réunion RSMC for all overwater storms of tropical characteristics from 1999 to 2016. As a key prerequisite, a proper SWIO threshold for rapid intensification has been defined at the 95% level of 24-h intensity change following the methodology used for Atlantic systems which enables basin inter-comparisons. Also, the statistical relationship linking the observed maximum intensity of a tropical system to the sea surface temperature (SST) has been examined to derive the first empirical MPI formula over the SWIO basin; it is tested as a potential predictor for TC intensity changes.

Based on the examination of a total of 26 potential predictors, statistical-dynamical tools of the same ilk as those developed in other basins have thus been designed to predict TC intensity change and/or the probability of RI at short range. Two tools will be presented: (i) a multiple linear regression model for TC intensity change at 24 h based on a multivariate adaptive regression splines (MARS) technique that models nonlinearities and interactions between variables, and (ii) a decision tree to anticipate RI periods in the next 24 h.