Deep Learning Assimilation of SAR Data with WRF Parameterizations for Improved Tropical Cyclone Wind Forecasts

This study evaluates the impacts of parameterization schemes in the Weather Research and Forecasting (WRF) model and synthetic aperture radar (SAR) wind assimilation using a deep learning 3D convolutional neural network (3D-CNN) on tropical cyclone (TC) forecasts. Among 8 physics suites, Scheme 6 achieves the best overall performance, with the smallest wind-speed errors root-mean-square error (RMSE) = 2.32 m s⁻¹, the best reproduction of maximum wind speed and radius against the International Best Track Archive for Climate Stewardship (IBTrACS), and accurate tracking of TC rapid intensification and track. Notably, Scheme 2 yields a structurally realistic vortex but suffers from large statistical bias, highlighting the need to combine qualitative morphology with quantitative validation. Performance differences arise from nonlinear interactions among microphysics, cumulus, and radiation processes. Assimilating SAR-retrieved surface winds using 3D-CNN reduced the RMSE by 33.8% in eyewall maximum wind speed compared with IBTrACS, while the correlation coefficient (Cor) increased from 0.82 to 0.91. The wind correction is concentrated in the inner core, with a bias reduction of 15–20 m s⁻¹, while the outer wind field remains unchanged. SAR assimilation also refines both TC intensity and radial structure, including the radius of maximum wind (RMW) and the 34-, 50-, and 64-kt wind radii (R34, R50, and R64), with RMSE reductions of 23.8%, 24.7%, and 33.7%, respectively. Systematic biases are observed during two physically distinct scenarios: extreme intensity (≥ 50 m s⁻¹ or ≤ 920 hPa), likely due to deficiencies in fully resolving nonlinear sea-state, sea spray, and air–sea coupling processes, and fast TC translation (≥ 14 m s⁻¹), due to under-represented asymmetric wind fields from the vector addition of translational flow. These biases may be mitigated by enhanced eyewall sampling and explicit translation-vector inputs. Overall, combining the optimal physics suite (Scheme 6) in WRF with SAR data assimilation using deep learning is robust and effective in improving TC intensity and structure forecasts, which is critical for TC early warning and hazard mitigation.