The Prediction Algorithm Of Time Series Satellite Nephogram Based On Generative Adversarial Network

Satellite nephogram,as one of the important source of weather information,in the severe weather detection and prediction plays an increasingly important role,such as ocean,desert,plateau areas lack of meteorological observation stations,satellite nephogram information,make up for the disadvantages of conventional detection data,play an important role to improve the forecast accuracy.To predict the movement and change of the cloud image under a certain time period can enhance the practicability of satellite cloud image data.At present,the challenges of cloud image prediction have two aspects:the observed cloud changes are mostly non-stationary,nonlinear and irregular;It is difficult to use one model to accurately describe the varying degrees of different cloud systems in the same area.In view of the above problems,this paper conducts prediction research on the future time image obtained by FY-4 based on deep learning.The main research content is divided into the following four parts:(1)Build a neural network to predict the future moment of the cloud image.The FY-4 satellite provided a large number of water vapor nephogram with a time interval of 1 small,and created a time series data set after grouping,cutting and orthogonal combination of 720 images in a month.Experiments show that the prediction model of the basic generative antagonistic network can predict the image of the next moment in the future more accurately,and the multi-layer GAN network prediction improves the prediction accuracy and prolongates the prediction timeliness.(2)Improve the model objective optimization function based on sample image features.The differences of features between the training samples and the predicted samples were analyzed,and the gray co-occurrence matrix eigenvalue information,such as contrast,entropy value and shape information,was calculated,and the model structure was improved accordingly.The improved multi-scale prediction model based on residuals changes its function from trajectory prediction of moving concerns to cloud map prediction which is more suitable for texture shape features.The experimental results show that the similarity between the predicted image and the real image is very high.(3)Realize super-resolution reconstruction of the predicted results.In order to solve the problem of low resolution and lack of detail in the predicted images,a super-resolution generative antagonistic network model was built to restore finer texture details.The experimental results show that the super-resolution operation has little effect on the prediction results of the basic GAN prediction model and the GLCM constrained prediction model,but it can improve the prediction accuracy of the multi-layer GAN network.(4)Evaluate the prediction accuracy and super-resolution accuracy.The similarity between the predicted image and the real image was analyzed by using two characteristic values of PSNR and SSIM.Firstly,the accuracy of the prediction of the three networks for the first time in the future is compared.Secondly,the timeliness of the prediction ability of the model in the subsequent prediction process of the three networks is compared.Finally,the influence of the super-resolution process on the prediction results is analyzed.