A Novel Deep Learning Framework for Tropical Cyclone Intensity Estimation Using FY-4 Satellite Imagery

Abstract

Tropical cyclone (TC) intensity estimation is a challenging task. Using the multispectral images (MSIs) captured by China's FY-4 Satellite, this paper addresses the issue by proposing a novel deep learning framework, combining a Coupled Convolutional Neural Network (Coupled CNN) for intensity categorization and Class-wise Regressors for wind speed estimation. The Coupled CNN is constructed by two parallel CNNs with specially-designed Transformation layers to process the MSIs with different spectral dimensions. The Coupled CNN could reduce the dimensions of input MSIs and generate classification probabilities as low-dimensional discriminative features to be fed to the successive Regressors. By sharing the weights between the two CNNs, it is also able to reduce the disparity of accuracy and avoid overfitting. Without being influenced by the curse of dimensionality, the Class-wise Regressors achieve high estimation accuracy by making full use of the reliable results of the Coupled CNN. The experimental results show that the proposed framework is able to perform TC intensity estimation with high intensity classification accuracy and low wind speed estimation errors for MSIs of different spectral dimensions.