Improved Tropical Cyclone Size Estimation Model Using Deep Learning and Multisource Data Fusion

Abstract

Tropical cyclone (TC) size, characterized by azimuthally averaged wind radii at specific wind speed thresholds, is a crucial indicator for assessing TC wind hazards. However, existing satellite-based TC size estimation methods often rely on limited multisource observations and insufficient spatiotemporal information, thereby limiting estimation accuracy. To address this issue, this study develops a TC size estimation residual network (TCSE-ResNet). The model introduces two key innovations: 1) joint spatiotemporal encoding, which converts five spatiotemporal features into 2-D spatial fields for integrated spatiotemporal learning and 2) an enhanced model architecture for improved feature representation, extraction, and fusion. For model training and evaluation, an eight-channel TC size estimation dataset is constructed by combining remote sensing images with TC best-track data. The proposed model estimates four TC size parameters: radius of maximum wind (RMW), 64-knot Wind Radius (R64), 50-knot Wind Radius (R50), and 34-knot Wind Radius (R34), achieving mean absolute errors (MAEs) of 3.5, 4.4, 6.4, and 9.4 nmi. Compared with the three-channel model, the eight-channel model achieved an average relative MAE reduction of 33.1%, including reductions of 37.1%, 41.8%, and 39.0% for R64, R50, and R34, respectively. The squeeze-and-excitation (SE) module analysis indicates that the model assigns the highest attention weights to satellite-retrieved precipitation and short-term temporal information. The spatial attention analysis shows that the model focuses on complementary spatial regions across the three satellite variables, providing insight into the observational features emphasized by the model during TC size estimation.