Using Deep Learning Algorithms To Estimate Atlantic Hurricane Intensity From Satellite Infrared Images

Tropical cyclone(hurricane)is one of the most frequent and serious natural disasters in the world.Promptly grasping the information of its center location,intensity and structure is of great significance for the tropical cyclone forecasting,disaster prevention and management.With the development of artificial intelligence,various deep learning methods have shown strong abilities in solving nonlinear problems,image recognition and other aspects.However,there are still few studies on the application of these methods in hurricane intensity estimation.In addition,the main reason that restricts the accuracy of hurricane intensity estimation is the lack of real observational data,while in the Atlantic Ocean,there is an abundance of aircraft observation data which can be used for verification.Therefore,this study intends to carry out the research on hurricane intensity estimation method based on the infrared observation data of geostationary satellites with high spatial and temporal resolution in the Atlantic Ocean,intending to explore the potential of deep learning methods in hurricane intensity estimation,so as to provide reference for forecasters to understand the actual situation.Specific research contents and main conclusions are as follows:Firstly,37 characteristic factors within 1.25° radius of hurricane center are extracted from the satellite infrared brightness temperature data,and the model is established by using stepwise regression method.A total of 18 factors are selected to reflect hurricane intensity by the size and location of convective clouds and the smoothness of the cloud system.According to the error distributions of different levels,the intensity estimation equation is linearly modified,the root mean square error(RMSE)between the final estimation result and the best track data is 13.16 kt,and the root mean square error comparing with the aircraft reconnaissance data is15.87 kt.Then,in the establishment of the deep neural network model,the automatically extracted features are combined with the human-defined features.that is,the hurricane image is first classified into two categories(the wind speed is classified with 64 kt as the dividing line),and then the intensity is estimated respectively.The classification results on the test samples reached 94.2%.the RMSE between the estimated intensity based on the classification results and the best track data is 12.64 kt,and the RMSE comparing with the aircraft reconnaissance data is 14.63 kt,the accuracy is improved by 4%-7.8% than using the linear method.Finally,the research of automatic feature extraction of hurricane cloud images and intensity estimation using convolutional neural network is carried out.Taking8°×8° image as the best input,the optimal model is obtained after operations such as convolution and pooling.At the same time,18 h time smoothing is introduced.The RMSE between the estimated results and the best track data is 10.59 kt,and the RMSE between the estimated results and the aircraft reconnaissance data is 11.35 kt,which improved the accuracy by 16.2%-22.4% compared with the previous two methods.Meanwhile,the accuracy of this method(10.59kt)is 9.3% higher than using ADT,it is also better than other methods for estimating hurricane intensity in the Atlantic Ocean.In conclusion,the stepwise regression method combined with the artificially defined characteristic factors and the deep neural network method can effectively estimate the hurricane intensity,indicating that the artificial defined characteristic factors have a significant impact on the hurricane intensity.However,the estimation accuracy of the deep neural network method is higher than that of the stepwise regression method,which indicates that the nonlinear method is better than the linear method in hurricane intensity estimation.Moreover,the accuracy of the convolutional neural network method is higher than that of the deep neural network,this may be because the automatically extracted feature factors are more comprehensive than the artificial defined feature factors,which reflects the advantages of the image recognition method in hurricane intensity estimation.