Typhoon Activity Identification And Prediction Research Based On Satellite Cloud Images

Typhoons are a very destructive natural disaster,and China is also located in the northwest Pacific Ocean typhoon high incidence area,with an average of seven typhoons landing on the southeast coast of China each year,causing significant economic losses and casualties in China.In engineering,wind field information is the basis for building wind effect analysis,but most of the existing building wind load codes describe the wind field for benign winds.Typhoons,on the other hand,are anomalous winds,and their wind profiles differ significantly from those of benign winds.In order to solve the problem that the typhoon wind pressure field is assumed to be constant along the height,a three-dimensional typhoon wind pressure field model needs to be established,and the intensity and path of the typhoon is the basis for establishing a three-dimensional barometric pressure field model.At present,typhoon intensity and path predictions are mainly provided by meteorological stations,and there are limitations in scale and computational resources in practical engineering applications.In this paper,based on meteorological data,deep learning techniques are used to build a prediction model for typhoon intensity and path.Firstly,the problem of typhoon prediction is transformed into a time-series prediction problem;then the one-dimensional data prediction is transformed into an image prediction,and the key parameters are identified through the image features of the next moment,so as to achieve typhoon prediction.The main contents and conclusions of the thesis are as follows:The main work in this study involves typhoon parameter identification,typhoon parameter prediction and typhoon generation prediction;technically,text-based and image-based prediction can be classified according to the type of data;and end-to-end and non-end-to-end according to the model structure.End-to-end refers to the mapping of deep learning models directly from raw data to output.Based on the end-to-end idea,this study improves the Conv LSTM model to achieve typhoon parameter prediction for group maps.And the non-end-to-end technique proposed in this study refers to first predicting the satellite cloud map at the next moment using spatio-temporal prediction models(Conv LSTM and SA-Conv LSTM),and then identifying the target using target recognition models.Using non-end-to-end techniques this study proposes a PQTC scheme and an FSTC scheme;the PQTC scheme not only successfully predicts typhoon intensity for the next 24 hours,but also effectively detects rapid typhoon intensification;the FSTC scheme not only solves the data timeliness problem,but also successfully predicts typhoon generation for the next 48 hours.In addition,a loss function applicable to image prediction was designed in this study,which successfully solved the problems such as edge noise metric defects.In order to validate the reasonableness of the model,this study uses case studies,visualisation and comparison with traditional methods to successfully verify that the model has a certain degree of reliability.In high-rise buildings,where wind loads are the main controlling loads,this study will contribute to the full life-cycle simulation of typhoons and coastal wind hazard assessment.