Research On Typhoon Cloud Image Prediction With Synergistic GAN And Self-attention Spatiotemporal LSTM

China is a country prone to typhoons,which are often accompanied by strong winds and rainstorms.The disasters triggered by typhoons pose a great threat to people's daily travel and safety of life and property.Therefore,how to scientifically and effectively provide typhoon warning services is of great significance to the construction of China's meteorological disaster defense system.At present,the commonly used methods in typhoon forecasting include numerical forecast,statistical forecast and statistical dynamic forecast.These three methods all rely on a lot of prior knowledge and artificial parameter setting,which is easy to cause problems such as low accuracy,short timeliness and poor generalization of the prediction.Recently,with the extensive layout of meteorological satellites and the rapid development of remote sensing technology,the global satellite cloud image big data has seen an explosive growth.More and more scholars try to apply deep learning technology to the processing and analysis of satellite cloud images,opening up a new field of meteorological visualization prediction research.Inspired by this,this paper carried out a study on typhoon cloud image prediction based on deep learning,focusing on solving the shortcomings of existing traditional methods and methods based on deep learning,in order to further improve the accuracy,refinement and intelligence level of typhoon cloud image prediction.The main work and innovations of this article are summarized as follows:1)Aiming at the problems of short timeliness,large error and poor generalization of traditional methods,a typhoon cloud image prediction model based on self-attention spatiotemporal LSTM is proposed.Based on ConvLSTM,the model introduces spatial memory cells and attention mechanism into its internal unit,which enables the model to extract temporal information and spatial features at the same time,and further enhances the unified modeling of short-term trend and long-term dependence;residual connection is added to the external frame to promote the fusion and sharing of spatial features of different coarse and fine granularity.The experimental results show that the model can accurately predict the morphological changes and trajectories of typhoons,and is superior to the existing models in various evaluation indexes.2)Aiming at the problems of poor image authenticity and lack of clarity based on deep learning methods,a typhoon cloud image prediction network based on GAN and self-attention spatiotemporal LSTM is proposed.The model draws on the game idea of the GAN network,introduces the space-time 3D discrimination module and the cloud image prediction module for adversarial training,and continuously optimizes and corrects the imaging quality of the predicted cloud image.In addition,the overall performance of the prediction network is further enhanced by constructing spatiotemporal attention modules and modifying the loss function.Experiments have proved that the network can maintain the overall visual sense of the predicted cloud image while ensuring the accuracy of long-term prediction.The generated cloud image is both accurate and realistic,which better meets the actual typhoon prediction business needs.In summary,the method proposed in this article is another innovative application of deep learning in the field of meteorological satellite cloud images,which provides new ideas and solutions for building a precise,refined and intelligent typhoon prediction system.And the network has strong generalization and can be better applied to other weather forecasting tasks such as radar echo extrapolation,strong convective weather warning,etc.,and has good landing application value and industrialization prospects.