Tropical Cyclone Intensity And Structural Feature Estimation Based On Deep Learning Combined With Satellite Data

Tropical Cyclone(TC)is a common natural disaster that will threaten people’s lives and property losses.Aiming at the current problems of TC intensity estimation and structural feature recognition,we designed TC detection,TC intensity classification estimation and TC radius estimation schemes based on deep learning technology,which provides a reference for improving my country’s TC intensity forecast and wind field inversion level.It mainly includes the following three aspects of work:(1)TC detection based on TCDNet combined with infrared satellite cloud images: At present,the main methods for TC detection are based on meteorological factor analysis and machine learning feature extraction.There are few studies on automatic detection of tropical cyclones from satellite images.Based on the infrared cloud images of the tropical cyclone from FY-2 and FY-4 satellites,we have built the FY-2 satellite data set and the dual-satellite hybrid cloud image data set,which contains the infrared satellite cloud images of tropical cyclones in each generation and development stage.Then we designed a TC detection network model called TCDNet.we uses clustering to modify the anchor frame ratio on the original Faster R-CNN,adjust the anchor frame size according to the actual situation,and replace the VGGNet of Faster R-CNN with Res Net.Finally rich infrared brightness temperature characteristics of TC can be extracted from images thereby the accuracy of network detection of TC targets is improved.Experiments show that our method has a detection recall rate of more than 84% on the three tropical cyclone data sets,and the FY-2 satellite data set has reached more than 95%.This model can provide an interesting area containing tropical cyclones for the subsequent objective center positioning and intensity determination of tropical cyclones,reducing the workload of manually selecting TC targets.(2)TC intensity classification and estimation based on TCICENet combined with infrared satellite cloud images: currently TC intensity estimation model mainly uses infrared satellite images to extract the TC center,cloud band characteristics,information entropy of the cloud area brightness temperature of the cloud image,and the brightness temperature gradient of the cloud top,deviation angle and its statistical value,spiral rain band fitting and kernel symmetry and other specific factors.The methods above depend to a large extent on human subjectivity and experience,as well as existing information related to satellite data.In addition,it is difficult for meteorologists to determine whether a feature is suitable for the intensity regression of all the various TCs in different basins and development stages.Based on infrared stationary satellite images of the Northwest Pacific Basin and cascaded deep convolutional neural network(CNN),a new type of tropical cyclone intensity classification estimation model(TCICENet)is proposed in this paper.This model consists of two CNN network modules: TC intensity classification(TCIC)module and TC intensity estimation(TCIE)module.First,use the TCIC module combined with infrared satellite images to divide the TC intensity into three categories.Next,three TCIE models based on the CNN regression network are proposed,which are classified according to infrared satellite images of different intensity levels.The three TCIE models take into account the misclassification ratio of the TCIC module to improve the accuracy of TCIE determination.We use the TC from 1981-2013 as the training sample,the TC from 2014-2016 as the verification sample,and the TC from 2017-2019 as the test sample.In order to reduce the computational burden of training the TCICENet model,various input image sizes were explored.Compared with the best track,the image size of 170×170 pixels achieves the best performance.Finally,the root mean square error(RMSE)of the model is8.60 kt,and the average absolute error(MAE)is 6.67 kt.(3)TC wind radius estimation based on CNN and BiLSTM combined with infrared satellite cloud images: At present,there are relatively few real-time and objective estimation techniques for TC wind scale.First,based on infrared satellite cloud images and CNN,this chapter explores and improves the three networks of Alex Net,Res Net34,and Res Net101 for extracting TC wind radius(maximum wind radius(RMW),fresh gale radius(R34),whole gale radius(R50).),hurricane radius(R64))performance.Experiments show that different sea area data sets and different scale indicators have different suitable networks.When the number of sample is small,the model that is too complex is easy to overfit and get unsatisfactory estimation results.TC is a system that evolves over time,so timing information is effective information.We combine the TC cloud images at the previous 3 moments and the current moment into 4 moments(0-9 hours)of input timing diagrams,and propose a CNN combined with BiLSTM model,called TCIE＿BiLSTM.The model not only extracts the spatial structure features of satellite cloud images,but also considers time sequence information.The TCIE＿BiLSTM model improves the accuracy of the TC wind radius estimation compared with the TCIE model.