Research On Forecasting Method Of Temporal And Spatial Sequence Typhoon Images Based On Improved LSTM Model

The storm surge disaster caused by typhoons causes a large number of economic losses and casualties in the coastal areas of the world every year.Therefore,obtaining accurate and real-time typhoon forecast information to provide decision support for disaster prevention departments'early warning and forecasting has become an urgent and important research topic.At present,typhoon prediction research at home and abroad is divided into two categories,one is based on traditional methods such as numerical prediction simulation and statistics-based typhoon prediction^[1][2].This method requires a large amount of typhoon observation data,and requires a sufficient understanding of the typhoon's internal structural changes,key physical processes and mutation mechanisms^[3].In addition,the prediction results of the typhoon process are affected by the typhoon model structure,topography,model initial field and boundary conditions^[4].The other is a strategy based on the combination of remote sensing satellite cloud image data set and deep learning.With the breakthrough and development of satellite remote sensing technology,the acquisition of remote sensing satellite images with wide coverage and high spatial resolution provides a large amount of real-time data set resources for deep learning that has attracted much attention from academia and industry.At this stage,the research on typhoon mainly focuses on typhoon intensity and typhoon path.Based on deep learning and time series typhoon images,the research on predicting real-time typhoon images is relatively fragmented.The complete cycle of remote sensing satellite typhoon images from generation to extinction is a typical long time interval time series data.The attention mechanism has been applied in the field of natural language processing,but there are also differences between the connection between language context and the connection mechanism before and after the physical phenomena of meteorological disasters.At present,there is no research on applying the attention mechanism to the prediction problem of meteorological physics time series.On the other hand,to capture spatial correlation and dig out its temporal and spatial information.Based on this,this paper researches the method of typhoon satellite image prediction and typhoon disaster grade prediction based on deep learning time-space sequence.The research results are mainly divided into three aspects:(1)For the typhoon remote sensing image data set with time series characteristics,a typhoon image prediction and generation model Seq Typhoon based on time series is constructed.Remote sensing satellites provide typhoon satellite images of the entire life cycle of a typhoon.The typhoon images of the entire life cycle have the characteristics of time series.The main object of the cyclic neural network is the data with the characteristics of time series,which can effectively mine the time series information and the time series relationship in the data.Based on the above ideas,the specific construction of the Seq Typhoon model is as follows:First,perform gray image histogram equalization preprocessing on the typhoon image sequence data set to achieve image enhancement;then,four consecutive historical moment typhoon remote sensing image sequences with a time interval of 6 hours After being resampled,they are used as the input of the encoder part of the Seq Typhoon model.The main feature of the Attention mechanism is to update the semantic vector output by the encoder in a targeted manner;finally,predict the typhoon image in the next 6 hours,12 hours or 18hours based on the semantic vector and the LSTM-based decoder.From the root mean square error results and the loss function graph,it can be concluded that the Seq Typhoon model converges effectively.Image information entropy and typhoon disaster grade prediction results are multi-faceted evaluation standards,which also prove the quality of the predicted typhoon image.At the same time,the influence of different time intervals,different prediction durations and different spatial resolutions between adjacent images on the prediction of typhoon images is studied.Experimental results show that the most important factor in typhoon images is the time interval between adjacent typhoon cloud images,followed by prediction duration and spatial resolution.(2)For the typhoon remote image data set with temporal and spatial characteristics,a multi-modal model TS?Conv LSTM for typhoon image prediction and generation based on temporal and spatial sequence is constructed.The life cycle of a typhoon,from generation to extinction,not only has obvious time series characteristics,the position of the typhoon eye on the typhoon image,the change of typhoon trajectory and other information indicate that it also has spatial information.Therefore,the TS?Conv LSTM model is innovatively proposed,which is based on convolution and deconvolution neural networks to improve LSTM(Long Short-Term Memory).This model also uses the Seq2Seq framework based on the Attention mechanism to increase the spatial correlation of typhoon images extracted by convolutional neural networks.The specific method is to change the operation of the gate(forget gate,input gate,output gate)in the LSTM cell in the encoder to convolution calculation,that is,convolution and LSTM cells are merged to form a conv LSTM cell.Deconvolution is an up-sampling learning model that can generate images of the same size as the input data.The decoder uses the fusion of deconvolution and LSTM cells to form trans Conv LSTM cells,trains and predicts and generates typhoon images in the future.Experimental results show that the time series-based multi-modal fusion TS?Conv LSTM model is better than the time series-based Seq Typhoon model.(3)Construct a typhoon disaster grade prediction model corresponding to the typhoon image.In order to further verify the results of typhoon image prediction,in addition to using evaluation indicators such as root mean square error and image information entropy,this paper also builds a typhoon disaster grade prediction model.That is,for typhoon images generated based on time series and time-space sequence model predictions,a deep convolutional neural network is constructed to predict the typhoon disaster level corresponding to the typhoon image.The typhoon image can correctly predict the accuracy of the corresponding typhoon disaster level,as one of the evaluation indicators for evaluating the typhoon image prediction model.