Analysis And Research Of Storm Detection Algorithm Based On Deep Learning

In recent years,with the improvement of computer technology and the acceleration of social informatization,the requirements of weather forecasting in various industries have also improved.Especially the storm,which directly affects the bad convective weather of people's travel and outdoor work,brings a lot of inconvenience to people's lives.In addition to timely and accurate forecasts,accurate and real-time identification of the location of the storm is also a very research-oriented part.Current storm recognition algorithms are mostly based on traditional numerical methods for radar echo data.These methods search and merge from the various dimensions of spatial radar mosaic data in the identification process,the process is cumbersome,and the intermediate process is too easy to generate multiple errors,which seriously affects the final recognition result of the storm.In addition,the traditional storm identification method only cares about the continuity between spatial data during the detection process,and the gradient that best reflects the overall storm is not taken seriously.This easily causes errors in the merger and screening process of storm cells,affecting the final result.Considering the similarity between spatially structured radar mosaic data and image data,Convolution Neural Network(CNN)is used to detect storms in radar mosaic data.After studying the object detection model based on deep learning,comprehensive consideration of the shortcomings of traditional storm detection algorithms,This paper proposes a Deep Learning Based SD-CNN Storm Detection Model(Storm Detection Based on Convolutional Neural Network,SD-CNN).The SD-CNN model uses the RPN(Region Proposal Network)sub-network to extract the candidate storm regions,and then uses the RPC(Region Proposal Classification)sub-network to re-classify the candidate storm regions,and finally obtains the results of the storm detection.The experimental results show that the SD-CNN model is superior to the traditional storm detection model in accuracy and speed.