Strong Convective Weather Warning Algorithm Based On Satellite Image Sequence

Meteorological research shows that strong convection is often the direct cause of natural disasters such as rainstorms,tornadoes,hail,etc.These natural disasters will cause great inconvenience to people's life and production behaviors,and even endanger people's life safety in severe cases.If we can predict the development of convective cloud systems,we can avoid the losses caused by natural disasters to a certain extent.We can observe the change process of convection from the satellite cloud image sequence.In order to predict the future development of convective clouds,we can first predict the subsequent changes of the satellite cloud image,and then detect the strong convection based on the predicted satellite cloud image.This process can be abstracted into two problems: video image sequence prediction and image sequence segmentation.The existing video prediction algorithms and image segmentation algorithms have a wide range of application scenarios,and there is no specific model algorithm designed for the characteristics of satellite data.In the video prediction problem,the current mainstream method is often based on the seq2 seq structure.A major disadvantage of this structure is that the information obtained by the decoder only comes from the encoder's encoding of the input,and the effect presented is to increase the image with the number of prediction steps.Increasingly blurred.In image segmentation,the current methods are mostly for single-frame images and the target that needs to be segmented is a rigid body.These methods do not take into account the effect of temporally continuous picture sequences on the segmentation effect and the case that the target is fluid.In this paper,a recursive model based on feature pyramid is proposed to address the shortcomings of the seq2 seq prediction framework.The design of the feature pyramid is mainly to fuse features of different granularities and to restore the image more accurately.In order to make better use of multi-scale features,the model in this paper also uses a multi-scale objective optimization scheme and designs the corresponding loss function.The design of the recursive architecture refers to the idea of a one-way language model,mainly to solve the problem of image blur when predicting multiple steps.The prediction results of the model in this paper are clearer than the current mainstream methods,and the HSS index has increased by 0.71 percentage point,and the CSI index has increased by 0.51 percentage point.A strong convective cloud is a special cloud system in a satellite cloud image.It is a fluid,that is,it undergoes large-scale deformation during the change process.In general,accurate detection of convective clouds requires a combination of satellite cloud image changes,so this paper proposes to use the Seq U-Net model to solve this problem.Seq U-Net introduces temporal features to the image segmentation problem.The model uses a convolution-long-term short-term memory network to extract temporal features.Compared with the previous model,the results of the model in this paper have an improvement of 0.54 and 0.21 percentage points on the POD and CSI indexes,respectively.