Tropical cyclone size estimation based on deep learning using infrared and microwave satellite data

Abstract

Tropical cyclone (TC) size is an important parameter for estimating TC risks such as wind damage, rainfall distribution, and storm surge. Satellite observation data are the primary data used to estimate TC size. Traditional methods of TC size estimation rely on a priori knowledge of the meteorological domain and emerging deep learning-based methods do not consider the considerable blurring and background noise in TC cloud systems and the application of multisource observation data. In this paper, we propose TC-Resnet, a deep learning-based model that estimates 34-kt wind radii (R34, commonly used as a measure of TC size) objectively by combining infrared and microwave satellite data. We regarded the resnet-50 model as the basic framework and embedded a convolution layer with a 5 × 5 convolution kernel on the shortcut branch in its residual block for downsampling to avoid the information loss problem of the original model. We also introduced a combined channel-spatial dual attention mechanism to suppress the background noise of TC cloud systems. In an R34 estimation experiment based on a global TC dataset containing 2003–2017 data, TC-Resnet outperformed existing methods of TC size estimation, obtaining a mean absolute error of 11.287 nmi and a Pearson correlation coefficient of 0.907.