Operational tropical cyclone (TC) maximum wind speed (intensity), size and asymmetry estimates are derived from subjective and objective analysis of satellite images at 6-hourly intervals. This research presents Pix2Pix generative adversarial networks (GANs) that directly predict complete TC 10 m wind field structure and intensity from satellite imagery. We present two models developed using complementary training datasets, each with ~ 2000 real or synthetic Infrared satellite (IR) and 10 m wind field pairs. The first utilizes Weather Research and Forecasting Model (WRF) simulations of real and idealized TCs for the wrf2wrf model. For the second model, mir2para, real satellite IR observations are paired with parametric wind fields derived from the extended best-track data set of TC intensities and wind radii.
Both models transform single-channel IR satellite observations into spatially resolved wind field reconstructions over 6° × 6° (~ 666 km × 666 km) domains at 256 × 256 pixel resolution. Cross-validation results demonstrate realistic wind field generation with mean pixel-wise RMSE of 6.7 knots for mir2para and 7.3 knots for wrf2wrf. Trained on synthetic data, wrf2wrf, performs well across domains achieving 10.4 knot RMSE when tested on 73 real IR images without additional training.
Storm outer-asymmetry is captured by mir2para, enabling diagnosis of wind radii for 34, 50 and 64 knots. When compared with operational wind radii for the same storms, mir2para achieved mean absolute errors of 12.8 nautical miles for radius of maximum wind (RMW) and 26.9 nautical miles for 34-knot wind radii across all four quadrants. More detailed storm structures including rainbands and eyewall asymmetries are captured by wrf2wrf.
Compared to existing AI and operational methodologies, this approach provides higher temporal resolution where each new TC IR satellite image (every 10-15 minutes) can be converted to TC wind field at a high spatial resolution. This approach offers significant promise for enhancing operational TC analysis and forecasting capabilities with minimal computational cost.
Copyright. All Rights Reserved.