Tropical Cyclone Intensity Estimation Using Multidimensional Convolutional Neural Network From Multichannel Satellite Imagery

Abstract

Estimating tropical cyclone (TC) intensity is the first step in the processes of monitoring and predicting destructive TC disasters. Due to the dilemma of meteorological methods, accurate estimation of TC intensity is a long-term challenge. In recent years, while deep learning methods have been applied to TC intensity estimation, most of them fail to make full use of multichannel satellite imageries to consider the three-dimensional (3-D) structure of TC. In this letter, we propose a novel deep learning model (3DAttentionTCNet) to overcome this shortcoming. The model can automatically extract 3-D environment information related to TC intensity from multichannel satellite observation imageries such as infrared (IR), water vapor (WV), and passive microwave rainrate (PMW) satellite imageries by 3-D convolution. In addition, we employ the convolutional block attention module (CBAM) to simulate visual attention for strengthening the model’s attention to core cloud structure and important channels. The experimental results show that the root-mean-square error (RMSE) of the proposed model is 9.48 kts, which is improved by 25% compared to that of the advanced Dvorak technique (ADT) and by 9.2% over that of the traditional deep learning method of TC intensity estimation.