首页 > , Vol. , Issue () : 1993-2002
热带气旋（Tropical Cyclone, TC）是影响我国的一个重要天气系统。TC强度的准确估测对台风灾害防御具有至关重要的意义。本文基于我国第二代静止气象卫星风云四号（FY-4A）多通道扫描成像辐射计（Advanced Geosynchronous Radiation Imager, AGRI）资料，建立了台风强度识别的深度卷积神经网络模型（Convolutional Neural Network, CNN），并对台风强度不同等级和台风中心最大风速进行了试验研究。结果表明，CNN模型具有良好的高维非线性处理能力和算法稳定性，能对TC强度进行有效估计，不同TC强度等级识别精度均在97%以上，近中心最大风速平均绝对误差（Mean Absolute Error, MAE）为1.75m/s，均方根误差（Root Mean Square Error, RMSE）为2.04 m/s。CNN可有效挖掘卫星TC形态的深层信息，对台风强度的定量化估测具有较高的应用前景。
Objective Tropical Cyclone (TC) is one of the most destructive meteorological disasters. The strong winds and heavy precipitation have significant impact on people"s lives, property and social and economic development. Therefore, the accuracy of TC"s path and intensity prediction is always an important direction of meteorological research. However, considering the complexity and variability of typhoon cloud pattern, the existing objective methods are usually based on statistical linear regression, and they still have deficiencies in expressing the dynamic change of complex TC cloud pattern characteristics. The deep learning algorithm has good performance in high-dimensional nonlinear modeling, which is able to identify the input mode with displacement and slight deformation accurately, and it is significant for tropical cyclone (TC) monitoring with dynamic changes over time. In order to further develop TC intensity estimation technology in the field of satellite remote sensing, this paper applied a new machine learning technology to analyze and study TC intensity of FY-4A/AGRI data from the China’s second-generation stationary meteorological satellite. Method Firstly, a deep convolution neural network (CNN) model was used to effectively distinguish and quantitatively estimate the TC intensity level and center wind speed. The images of day and night were respectively input into the CNN’s convolution sampling channel for obtaining same size spectral features and combining them together. Then, multi-layer convolution, pooling, nonlinear mapping and other operations were used to deeply mine the input characteristics, and finally the TC intensity was estimated. The experiment was divided into TC intensity classification test and quantitative estimation test of TC center maximum wind speed. The CNN model was used to convert the recognition of TC intensity into pattern recognition of satellite cloud image, which could classify and identify the TC level. Result The experiment found that whether the overall classification accuracy or the respective accuracy of day and night statistics, the recognition accuracy of TC intensity was all above 95%. Compared with k-nearest neighbor, error back-propagation neural network, multiple linear regression, support vector machine and other classical classification algorithms, it improved by 7-16 percentage points. Moreover, CNN was also superior to the classical algorithm in the respect of classification accuracy. The CNN model was consisted of two fully connected network layers (each layer has three neurons), and the TC wind speed could be quantitative estimated by prior training samples of the network parameters. Compared with the data of Tropical Cyclone 2017 Yearbook, the mean absolute error of wind speed was 1.75 m/s, and the root mean square error of wind speed was 2.04 m/s, which were lower than the corresponding errors of deviation angle variance technique (DAVT) by 85.70% and 84.38% respectively. In other words, the CNN algorithm has a high application prospect in the quantitative estimation of typhoon intensity. Conclusion In conclusion, as the first launched Chinese second generation geostationary meteorological satellite, FY-4A has its advantages of multi-channel structure; high spatial and temporal resolution. Based on these features and the techniques advantage of deep neural network and CNN’s flexible structure, this study stated an improved CNN model which is tailor made for FY-4A data. It is able to deeply mining typhoon"s morphological characteristic effectively and achieve high precision typhoon intensity estimation, which has positive research value and application prospect for the quantitative estimation of typhoon intensity.