Simulation of sea surface temperature retrieval based on GF-5 thermal infrared data

Abstract

A Tropical Cyclone (TC) is one of the most destructive meteorological disasters. The strong winds and heavy precipitation have significant effect on people's lives, property, and social and economic development. Therefore, the accuracy of the path and intensity prediction of TCs is always an important consideration in meteorological research. However, considering the complexity and variability of typhoon cloud patterns, the existing objective methods are usually based on statistical linear regression. Moreover, they still have deficiencies in expressing the dynamic changes of the complex characteristics of TC cloud patterns. The deep learning algorithm performs well in high-dimensional nonlinear modeling and accurately identifies the input mode with displacement and slight deformation. This algorithm finds significance in TC monitoring with dynamic changes over time. To develop TC intensity estimation technology further in the field of satellite remote sensing, this study applied a new machine-learning technology to analyze and to study the TC intensity of FY-4A/AGRI data from China's second-generation stationary meteorological satellite. First, a deep Convolution Neural Network (CNN) model was used to distinguish effectively and estimate quantitatively the TC intensity level and center wind speed. The images of day and night were placed into the convolution sampling channel of the CNN to obtain and combine same-size spectral features. Then, multilayer convolution, pooling, nonlinear mapping, and other operations were used to mine the input characteristics deeply. Finally, the TC intensity was estimated. The experiment was divided into the TC intensity classification test and the quantitative estimation test of the TC center maximum wind speed. The CNN model was used to convert the recognition of the TC intensity into the pattern recognition of satellite cloud images, which could classify and identify the TC level. The experiment found that the recognition accuracy of the TC intensity was all above 95% regardless of the overall classification accuracy or the respective accuracy of day and night statistics. Compared with k-nearest neighbor, error back-propagation neural network, multiple linear regression, support vector machine, and other classical classification algorithms, it improves by 7-16 percentage points. Moreover, the CNN is also superior to the classical algorithm in terms of classification accuracy. The CNN model comprises two fully connected network layers (each layer has three neurons). The TC wind speed can be quantitatively estimated by prior training samples of the network parameters. Compared with the data of Tropical Cyclone 2017 Yearbook, the mean absolute error of the wind speed was 1.75 m/s, and the root mean square error of the 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%. Thus, the CNN algorithm has a high application prospect in the quantitative estimation of typhoon intensity. As the first second-generation Chinese geostationary meteorological satellite to be launched, FY-4A has its advantages of multichannel structure and high spatial and temporal resolution. On the basis of these features, the advantages of the techniques of the deep neural network, and the flexible structure of CNN, this study proposes an improved CNN model that is tailor-made for FY-4A data. The model has the capacity to mine the morphological characteristic of typhoons deeply and effectively and achieve high-precision typhoon intensity estimation. This model has positive research value and application prospect for the quantitative estimation of typhoon intensity.