A TENSOR NETWORK FOR TROPICAL CYCLONE WIND SPEED ESTIMATION

It is challenging to estimate wind speed of tropical cyclones directly using remote sensing image patterns. This paper approaches the task in two major steps: cyclone category estimation and wind speed regression. A novel framework based on Tensor Convolutional Neural Network (Tensor CNN) is proposed to solve the problem. Not only does the framework combine Tensor analysis for dimensionality reduction and deep neural networks for pattern recognition, the Tensor CNN also provides a unitary and concise mathematical representation form of the two significant models. The proposed framework is able to categorize cyclones by classification based on the Tensor CNN as well as exploit the estimated categories and predict the wind speed by a successive regression model. Experiments are conducted on multispectral imagery acquired by the FY-4 Satellite. Results show that the framework outperforms several classic models in cyclone category estimation and one state-of-the-art method, Deviation Angle Variance Technique (DAVT), in wind speed regression.