Research On Satellite Cloud Image Reconstruction And Prediction Based On Deep Learning

Satellite cloud image plays an important role in weather analysis and forecast.However,the spatial resolution of satellite cloud images can not meet the growing needs of meteorological monitoring.Using super-resolution(SR)method to improve the resolution is helpful to identify and locate the weather system.In addition,the prediction of cloud movement and evolution information in satellite cloud images is helpful to achieve more accurate weather forecast.However,the complexity of prediction methods limits the spatial resolution of the satellite cloud image.The satellite cloud image SR methods can get higher resolution predicted cloud image.Based on the deep learning method,this paper studies the SR and prediction of satellite cloud images.The main conclusions are as follows:(1)In order to solve the SR task of satellite cloud images,a new multi-path network model based on the joint training of back projection network and local residual network is proposed,which is called SRCloud Net.The multi-path structure enables SRCloud Net to extract features from different aspects of the image to obtain more accurate reconstruction results.In this paper,the satellite cloud image data set for SR research is constructed,which is called NCCI data set.NCCI data set includes natural color satellite cloud image data set and satellite contrail image data set.Considering the special linear shape,contrail image is used to test SRCloud Net.The reconstruction image of conventional SR methods is blurred and lost high-frequency details.Extensive experiments show that high-quality satellite cloud image and high-resolution reconstruction image can be obtained based on SRCloud Net.(2)In order to solve the prediction of satellite cloud image,an infrared cloud image prediction network OF-LSTM based on spatiotemporal long and short term memory network is proposed,which is called OF-LSTM.OF-LSTM can further learns and transmits the effective features in the input information through the proposed optical flow information module.A satellite infrared cloud image data set suitable for deep learning training is constructed.The experiments show that OF-LSTM can accurately predict the moving path and shape change of cloud cluster.Finally,the SR algorithm of satellite cloud image is used to get higher resolution cloud image.