A Physics-Informed, Rotation-Equivariant Model with Multiscale Fusion for Multitask Tropical Cyclone Analysis

Multitask tropical cyclone (TC) analysis remains challenging, despite recent advances in deep learning for automatically extracting structural features from multispectral geostationary satellite imagery. Most state-of-the-art models overlook two fundamental principles: the rotational symmetry inherent in TC evolution and the physical consistency governing atmospheric dynamics. This research proposes a physics-guided hierarchical fusion transformer that jointly performs maximum sustained wind speed (MSW) regression, minimum sea level pressure (MSLP) regression, and TC grade classification. The proposed framework incorporates a hierarchical equivariant encoder based on E2-steerable convolutions to capture multiscale, rotation-equivariant features, together with a cascaded cross-attention module that aggregates short-term temporal information and extracts spatiotemporal evolution patterns. Physics-informed loss functions are introduced to enforce physical consistency between MSW, MSLP, and TC intensity grades, while an uncertainty-based weighting strategy is employed to dynamically balance the multitask objectives during training. The model is trained and evaluated using IBTrACS best track data and the Japanese Meteorological Agency’s Himawari-8 / Himawari-9 infrared satellite images. Experiments on the 2023–24 test set yielded MAE errors of 6.38 (kt) for MSW and 5.58 (hPa) for MSLP, and these results consistently demonstrate improved performance over the accompanying benchmark deep learning models, including the VideoSwinTransformer (V-Swin), 3D residual network with 18 layers (ResNet3D-18), and ResNet-50. Further comparisons with the current advanced Dvorak technique (ADT) for 42 TCs in the western North Pacific demonstrate MAE and RMSE reductions of 21.63% and 20.78%, respectively, highlighting the potential of the proposed model for operational TC analysis. Significance Statement This paper aims to improve the accuracy and reliability of tropical cyclone (TC) intensity estimation from geostationary satellite imagery by proposing a physics-guided, hierarchical fusion transformer that explicitly captures the inherent rotational symmetry of TCs and the physical relationships among key intensity-related features.