Tropical Cyclone Wind Field Reconstruction for Hazard Estimation via Bayesian Hierarchical Modeling With Neural Network

Abstract

Tropical cyclones (TCs) are one of the biggest threats to life and property around the world. Accurate estimation of TC wind hazard requires estimation of catastrophic TCs having a very long return period spanning up to thousands of years. Since reliable TC data are available only for recently decades, stochastic modeling and simulation turned out to be an effective approach to achieve more stable hazard estimates. In common practice, hundreds of thousands of synthetic TCs are generated first, then wind fields are reconstructed along synthetic TC tracks for hazard estimation. A Bayesian hierarchical modeling approach to the reconstruction of TC wind field is proposed. A modified Rankine vortex is adopted as the wind field model, of which the four free parameters are modeled simultaneously through a multi-output neural network as a latent process of the wind field. The four parameters are finally represented, spatially and temporally, by a set of neural network weights, The Bayesian model averaging technique is used for parameter estimation and wind field reconstruction, based on a ensemble of maximum a posteriori estimates of the set of weights. Together with previously proposed algorithm for synthetic TC simulation, a two-stage scheme for TC wind hazard estimation has been formed, which is based on best-track data only and thus is highly consistent. Application of this scheme to the offshore waters in the western North Pacific basin shows inspiring performance and great flexibility for various purposes of TC wind hazard estimation.