Research On Cloud Movement Forecasting Method Based On Deep Learning

Spatiotemporal series data is that the data recorded at each time are images with spatial information.The prediction of such data is also called video prediction,that is,the prediction of future frames is based on the known frame sequence of video.Spatiotemporal series forecasting has important applications in robot,automatic driving,weather forecast and other fields,so the research on spatiotemporal series forecasting is of great significance.This paper focuses on the forecasting of cloud movement.The data used in this paper are the satellite cloud images obtained from the observation of FY-4A.Through the forecasting of the future movement and change of cloud,we can provide more reference information for weather forecast,which has practical research significance.Be cause the satellite cloud image contains a lot of non-cloud irrelevant information,the research process of this paper is to study the cloud detection problem of the satellite cloud image first,and then study the cloud movement forecasting problem on the obtained detection results.To solve the problem of cloud detection in satellite cloud image,this paper proposes a cloud detection algorithm based on dilated convolution and data distillation.Firstly,because of the various shapes and sizes of the cloud,by adding dilated convolution to the model,we can expand the receptive field without increasing the amount of computation,reduce the use of the max pooling layer,and save more information in the feature map.Combined with the fluid characteristics of the cloud area,the model can learn the multi-scale characteristics of the cloud area and improve the detection accuracy of the small cloud area by using the module composed of dilated convolution with different sampling rates.Then,this paper proposes a method of unlabeled data screening based on ensemble learning,which uses the unlabeled data after screening and data distillation to expand the training samples.Experimental results show that the model based on dilated convolution can achieve more accurate cloud detection results,and then the detection accuracy can be further improved by data distillation method.To solve the problem of cloud movement forecasting of satellite cloud image,this paper proposes a forecasting method of cloud movement based on RAST-LSTM.The algorithm adds attention mechanism and residual unit to ST-LSTM unit,in which attention mechanism can better allev iate the fuzzy problem of multi-frame forecasting of the model,and can supplement more details in the forecasting results;residual unit can improve the learning ability when the depth of the model increases,and further improve the accuracy of the model.Finally,the forecasting model of seq2 seq structure is designed based on RAST-LSTM.The experimental results show that the forecasting method based on RAST-LSTM proposed in this paper has a good performance in the clarity of the forecasting results and model accuracy.