Application of remotely sensed inner-core rainfall and surface latent heat flux in typhoon intensity forecast

Despite improvements in statistical and dynamic models in recent years, the prediction of tropical cyclone (TC) intensity still lags that of track forecasting. Recent advances in satellite remote sensing coupled with artificial intelligence techniques offer us an opportunity to improve the forecasting skill of typhoon intensity.

In this study rapid intensification (RI) of TCs is defined as over-water minimum central pressure fall in excess of 20 mb over a 24-h period. Composite analysis shows satellite-based surface latent heat flux (SLHF) and inner-core rain rate (IRR) are related to rapid intensifying TCs over the western North Pacific, suggesting SLHF and IRR have the potential to add value to TC intensity forecasting.

Several linear regression models and neural network models are developed for the intensity prediction of western North Pacific TC at 24-h, 48-h, and 72-h intervals. The datasets include Japan Meteorological Agency (JMA) Regional Specialized Meteorological Center Tokyo (RSMC Tokyo) best track data, the National Centers for Environmental Prediction (NCEP) Global Forecasting System Final analysis, the Tropical Rainfall Measuring Mission (TRMM) Microwave Imager sea surface temperature (SST), the Objectively Analyzed Air-sea Fluxes (OAflux) SLHF and TRMM Multisatellite Precipitation Analysis (TMPA) rain rate data. The models include climatology and persistence (CLIPER), a model based on Statistical Typhoon Intensity Prediction System (STIPS), which serves as the BASE model, and a model of STIPS with additional satellite estimates of IRR and SLHF (STIPER). A revised equation of TC maximum potential intensity (MPI) is derived using TMI Optimally Interpolated Sea Surface Temperature data (OISST) with higher temporal and spatial resolutions. Analysis of the resulting models indicates that the STIPER model reduces the mean absolute intensity forecast error by 6% for TC intensity forecasts out to 72 h compared to the CLIPER and BASE. Neural network models with the same predictors as STIPER can provide up to 28% error reduction compared to STIPER. The largest improvement is the intensity forecasts of the rapidly intensifying and rapidly decaying TCs.