Cyclone Intensity Estimation using Multispectral Imagery from the FY-4 Satellite

Tropical cyclone (TC) intensity estimation is vital to disastrous weather forecasting. In this paper, the task is approached as a classification problem, regarding the cyclone intensity levels as the class labels. Multispectral Imagery (MSI) captured by a recently launched satellite, No. 4 meteorological satellite (FY-4) of China, is used as the raw data for classification. To solve the problem, this paper proposes a machine learning framework with three major parts: useable band determination, band-wise classification and fusion. The framework is compatible with arbitrary classifiers for the band-wise classification. Since some band images acquired during night hours contain little useful information, a selector is designed and placed before each band classifier. Moreover, majority voting, a very efficient method, is employed to fuse the band-wise classification results. Experiments demonstrate that Multiple Logistic Regression (MLR), Support Vector Machine (SVM) and Back-Propagation Neural Network (BPNN), each in turn used as the band-wise classifiers, can yield high accuracy in labelling the TC intensity. The results also show the usefulness of the FY-4 data and the potentials of machine learning for automatic and accurate TC intensity estimation.