A Wind-Field-Aware Framework for Cyclone Intensity Estimation

Tropical cyclone (TC) intensity estimation is an important meteorology topic for exploring the latent relationship between the cyclone pattern in the satellite image and the maximum sustained wind (MSW) speed. Recent studies have made encouraging progress in high-dimensional classification and TC intensity estimation. Moreover, these methods mainly rely on abstract features for decision-making, which are difficult to interpret physically or be accepted by meteorologists. This paper proposes a novel TC intensity estimation framework based on ensemble learning over multisource data, including multispectral images of cyclones and wind field features showing cloud motions. The mentioned model, consisting of cyclone intensity classification (CIC) module and wind speed regression (WSR) module, is to train in an end-to-end fashion. We evaluate the model on the multispectral images (MSIs) and the atmospheric motion vectors (AMVs) both acquired by FY4 meteorological satellite of China during 2020. We observe the convergence and precision for regression to demonstrate the effectiveness and feasibility.