Research On Satellite Cloud Image Classification Based On CNN-LSTM Network

Different cloud types reflect different atmospheric conditions and development changes,foretell the future weather.For example,the vertical development clouds are usually associated with hazardous weather.The meteorological satellites has the advantages of continues,high spatial resolution,and wide coverage,which can provides more spectral features.Therefore,the automatic identification of the satellite cloud is one of the research hotspots in the field of remote sensing.Early cloud classification methods usually rely on the features of spectral channels for classification.However,the same cloud type with different spectral features and the different cloud types with same spectral features,so the classification performance is not ideal.Consider using the convolutional neural network(CNN)to automatically extract and classify the features of the space-spectrum joint information,and then compare the cloud image classification results of the classical cloud classification algorithm combined with the spectral information to verify the effectiveness of the satellite cloud image space-spectrum combination.In addition,in order to make full use of the space-spectrum joint information of cloud images,this paper constructs a CNN-LSTM network cloud classification model,and analyzes the classification results with other network models.In this paper,the research on the classification method of satellite cloud images is as follows:(1)Firstly,by the introduction of the physic properties of each spectral channel of Himawari-8 satellite,the type cloud,and the performance of the cloud categories is determined,the cloud classification scheme is determined,and the appropriate spectral characteristics are selected for the pixel cloud.And then combined with the spatial characteristics of the joint 7*7 neighborhood,establish an space-spectral set of combined cloud samples.(2)Secondly,based on the CNN network classification principle,the advantages of the model is extracted the spatial features.The CNN network that is suitable for satellite cloud image sample database is classified.And compared the satellite cloud image of the classical cloud classification method,verify the effectiveness of the combination of space spectral information.(3)Finally,the deficiencies of the CNN network cloud classification results is found that there is a case where there is a similar cloud spatial structure similar to the similarity.To this end,the CNN-LSTM network classification model is constructed,and perform the multi-view feature joint classification on satellite cloud images of space spectrum information to provide more effective classification information.Solve the problem of cloud block spatial structures,thereby enhancing cloud classification performance.And the model has an application prospects in the classification of satellite observation data.