Accurate tropical cyclone (TC) forecasts depend critically on realistic representations of storm structure so that high-resolution numerical model forecasts can be correctly initialized and the multifaceted hazards associated with TCs can be effectively communicated. Operational forecasters routinely assess surface wind metrics such as the maximum sustained 10-m wind and quadrant wind radii, yet TCs are inherently three-dimensional (3D) systems whose full kinematic structure must be accurately depicted to achieve skillful forecasts. A key resource for determining this 3D structure is the tail Doppler radar (TDR) aboard NOAA’s WP-3D (P-3) and G-IV aircraft, which provides high-resolution wind and reflectivity data within roughly 50 km of the flight track at a horizontal and vertical grid spacing of 2.0 and 0.5 km, respectively. However, TDR data coverage is limited by aircraft availability, mission logistics, and the distribution of hydrometeors, which can result in large spatial and temporal data gaps.
To address these limitations, a suite of machine learning models was trained on historical TDR analyses to learn the statistical relationships between TC 3D wind structures and coincident geostationary satellite observations, which are globally available, as well as other operationally available diagnostics pertaining to the TC's environment and intensity. The machine learning suite, referred to as the Tropical Cyclone Synthetic Wind Analysis using Regression Models (TC-SWARM), produces 3D wind field estimates with skill comparable to current uncertainties in best-track intensity estimates. This presentation will describe the suite of machine learning models that comprise TC-SWARM and evaluate the performance of TC-SWARM for cases drawn from an independent testing data set. Examples of TC-SWARM’s ability to fill in observational data gaps by reconstructing cases in the training data set will also be illustrated.
Copyright. All Rights Reserved.