A data-physics driven deep learning method for retrieving submesoscale wind speed fields in far-sea typhoons

Abstract

Submesoscale processes in far-sea typhoons represent a critical pathway for the transfer of mesoscale kinetic energy to small-scale turbulent dissipation. The associated wind fields also constitute a primary source of initial conditions and real-time assimilation data for contemporary mesoscale meteorological models. Traditional land-based and ocean-based observation methods struggle to provide complete information on submesoscale wind speed fields over the deep ocean. This study proposes a data-physics driven deep learning retrieval method based on satellite infrared remote sensing data in conjunction with fluid dynamics equations. First, a dynamic grid adaptation method for the multiscale spatiotemporal matching of typhoon wind fields is developed using multi-channel infrared remote sensing data from the Himawari-8 meteorological satellite and the ERA5 global reanalysis wind field data. Then, based on the established spatiotemporally matched database of horizontal wind speed fields for far-sea typhoons in the Northwest Pacific, a data-physics driven deep convolutional neural network model is constructed by embedding loss functions with the two-dimensional Navier-Stokes equations and the continuity equation in polar coordinates. This approach ultimately achieves accurate retrieval of the submesoscale horizontal wind speed fields in far-sea typhoons. The study demonstrates that, compared to the data driven models which only identify typhoon center locations and vortex directions, the data-physics driven model promotes the convergence of high-order iterations, effectively reconstructs radial gradient variations in the typhoon horizontal wind speeds, and significantly improves retrieval accuracy in the high-wind eyewall region. The recommended distribution for the physics driven sample points is at r/R = 0.1, 0.2, 0.3, 0.4, 0.7, and 1.0, with a physical constraint intensity coefficient α = 100. Under these conditions, the model achieves RMSEs of 0.91 m/s and 0.86 m/s for eyewall and global horizontal wind speed retrieval respectively, with BIASs of −0.27 m/s and 0.26 m/s.