Tropical Cyclone Intensity Estimation from Full-Swath 183-GHz Microwave Satellite Observations

Abstract Estimating tropical cyclone (TC) intensity from satellite observations remains challenging, and microwave (MW) imagery has been shown to improve skill relative to infrared-only approaches. This study introduces D-MINT183v2, an update to the D-MINT183 convolutional neural network that estimates current TC intensity from 183-GHz MW imagery. Unlike the original D-MINT183, which was restricted to the inner 50% of cross-track swaths to avoid limb effects, D-MINT183v2 is trained using limb-corrected full-swath MW imagery and includes scan angle as an additional predictor. A baseline model is developed for cross-track sounders (AMSU-B/MHS/ATMS) and then adapted to SmallSat sensors (TROPICS and Tomorrow.io TMS) using transfer learning. Across sensors, D-MINT183v2 maintains or improves accuracy relative to the original D-MINT183 within the inner swath and provides consistent skill across the swath, thereby substantially increasing the number of usable MW-based intensity estimates. Relative to the IR-only D-PRINT baseline, D-MINT183v2 produces more skillful intensity estimates at MW overpass times. SHAP-based attribution is used to quantify the roles of viewing geometry and channel-dependent brightness-temperature information in the retrieved intensities.