Tropical Cyclone Intensity Estimation with a Soft Label Boosted Network

Abstract

Tropical cyclone (TC) intensity refers to maximum sustained wind (MSW) speed near the center of a cyclone. TC intensity estimation can provide early warnings for coastal areas to avoid economic damage and life casualty. Recently, deep learning for remote sensing images has been applied to TC intensity estimation and enabled accurate MSW regression. In this paper, we first construct a new cyclone dataset, namely FY4A-TC, using the multispectral images (MSIs) of 81 cyclones captured by China's FY4A meteorological satellite from 2018–2021. Then we propose a Convolutional Neural Network boosted by Soft Labels (CNN-SL) to estimate TC intensity. The CNN is designed for MSW regression. Specifically, we superimpose a novel soft-label regularizer on the regression loss to increase estimation accuracy. The soft labels are generated from cyclone intensity categories following Gaussian distributions to provide additional information for supervision. To facilitate wind speed estimation, we also propose a series of preprocessing and postprocessing procedures, including MSW smoothing that utilizes temporal relevance of TCs to increase estimation accuracy. Experimental results on the FY4A-TC dataset show that CNN-SL outperforms several state-of-the-art methods for TC intensity estimation.