Research On Ground-based Cloud Classification Based On Transfer Deep Learning And Multimodalit

Cloud is a common natural phenomenon,and the accurate identification of its category is of great significance to weather forecast,satellite communication and other fields.With the rapid development of ground-based cloud observation equipment,ground-based cloud is a popular research direction in cloud classification.Since there are many types of ground-based clouds,the shapes of different types are similar and constantly changing,so the accurate identification of ground-based clouds has always been a great challenge.In recent years,deep learning method has been applied in the field of ground-based cloud classification,but the classification results are not ideal,and more research is needed for the application of ground-based cloud automatic classification service.In order to further improve the effect of deep learning in ground-based cloud classification,this paper introduces transfer learning and multi-modal information into deep learning,and uses transfer learning and convolutional neural network to carry out technical verification of ground-based cloud classification and recognition.Through network model,multi-feature mining and fusion,parameter design and other key technology research,The classification model of deep learning ground cloud suitable for large ground cloud database is established,and the classification software of ground cloud based on deep learning technology is designed.The research content is mainly divided into the following four parts:(1)In order to understand and study the application of deep learning and transfer learning in ground-based cloud classification,the features of ground-based cloud in convolutional neural network are visualized first,and then the advantages of transfer learning are verified by experimental comparison between transfer learning and non-transfer learning.Finally,in order to verify the applicability of the transfer deep learning technology in the classification of ground-based cloud,the common machine learning methods and deep learning methods are introduced respectively,and the superiority of the transfer deep learning is verified by experiments.(2)Aiming at the low classification accuracy of ground-based cloud,a ground-based cloud classification method based on migration deep learning Dense Net121 network is proposed.Firstly,a variety of classical networks are verified,Dense Net121 network with the best performance is selected,and then the network model is improved,and the top-level structure of shallow fusion branch and 4-layer network is designed and added.The transfer learning is used for training and fine-tuning.Finally,the high accuracy of this model is 93.43%on the ground-based cloud data set.It provides the theoretical and technical basis for the business application of ground-based cloud classification.(3)In addition to the visual features of ground-based clouds,the auxiliary features of ground-based clouds,namely the multi-modal information of ground-based clouds,play an important role in the classification of ground-based clouds.Based on the research on the mining and fusion of ground-based cloud multi-mode feature information,a ground-based cloud classification method based on multi-mode deep learning network(two-flow multi-mode Multi-layer Fusion network,DMMFN)is proposed.For the first time,the multi-mode information is transferred to different subnetworks separately,and heterogeneous feature fusion is carried out at the feature layer.Finally,the model has a high accuracy of85.70% on the multi-modal ground-based cloud data set.(4)Design and compile the ground-based cloud automatic classification software,load the first two experimental training models into the software,realize the ground-based cloud automatic classification function applicable to different application scenarios,and the performance test of the software,the test results show that the software recognition speed and accuracy is high,in line with the application standards of ground-based cloud automatic recognition software.