TC-RadiiFormer: Current-Time Quadrant Wind-Radius Estimation from Multispectral Satellite Sequences

The extent of tropical-cyclone winds is essential for assessing marine and coastal hazards, but surface observations remain sparse over the open ocean. Geostationary satellites continuously observe cloud systems, offering complementary spectral and temporal information for estimating wind radii. We present TC-RadiiFormer to jointly estimate the current radii of 34, 50 and 64 kt winds (R34, R50 and R64) in four geographic quadrants from multispectral satellite sequences. The model combines interactions among spectral channels with historical observations at their actual acquisition times, then decodes geographically aligned, ordered radius distributions. Training used 2017–2022 observations, model selection used 2023, and evaluation used 2024 final best-track wind radii across three basins. Three-seed positive-radius mean absolute errors (MAEs) were 13.164, 8.384 and 5.731 nautical miles for R34, R50 and R64. MAE averaged over the three thresholds was 27.1%, 25.4% and 13.0% lower than ResNet-50, Swin V2-T and Swin V2-T with a temporal Transformer, respectively. Matched comparisons supported contributions from multispectral observations, historical information and actual acquisition timing. Spatial and temporal case analyses connected these aggregate improvements to estimates of wind extent and its evolution. These results support joint use of spectral and temporal satellite information for geographically resolved wind-radius estimation.