Trough Line Automatically Recognition Based On Semantic Segmentation Network

The trough line is the line connecting the point of maximum cyclonic curvature in the isobaric low-pressure trough area.The air-pressure in the trough area is lower than both sides,the horizontal convergence of air flow is the strongest,and the weather near the trough changes obviously.This feature makes it an important consideration in weather forecasting.Usually,the trough line needs to be drawn for the analysis of upper-air weather map.At present,this work is mainly completed manually,and the drawing results exist errors due to human subjective influence,which takes a lot of human and time resources.Therefore,the automatic identification of trough line is of great significance to prediction.Starting from the definition of trough line,scholars have proposed a variety of algorithms to extract trough line from wind or geopotential height,but these recognition methods based on manually set rules have limitations.The trough line types in different seasons and regions are different,and this kind of method mainly depends on manual setting rules,which can hardly cover all types,resulting in low generalization ability of the method.To solve the above problems,this paper studies the automatic trough recognition method based on deep learning.The main work of this paper is as follows:(1)In view of the low generalization ability of the traditional rule-based automatic trough recognition method,an automatic trough recognition method based on fully supervised neural network is proposed.The model uses UNet of U-shaped encoder decoder structure as the basic structure to realize the accurate positioning of the recognition target by continuously fusing the shallow and deep features.To extract the abstract semantic features of trough and correlate the features between different channels,Xception network with deep separable convolution as the basic unit is used to replace the encoder part of the original UNet.In addition,a squeeze and excitation(SE)module with attention mechanism is added after each ordinary convolution in the decoder part to improve the recognition accuracy of the trough area by increasing the weight of the trough area.The experiments were carried on the meteorological elements data set composed of wind,geopotential height and temperature,and the results show that the recognition accuracy of this method on the test data set can reach more than80%.We also compared the metrics of several other excellent network and traditional trough automatic recognition methods in the field of semantic segmentation.The results show that the performance of the proposed model is better than other methods.(2)The recognition performance of convolutional neural network(CNN)based methods at the trough edge is poor,and this kind of methods need a lot of labeled data,resulting in a waste of human and time resources.In addition,there are a lot of subjective errors in the label,resulting in the deviation of the recognition results.An automatic trough recognition method based on multi-task cross-pseudo supervised network is proposed.This method uses a large amount of unlabeled data and a small amount of labeled data to train two networks with the same structure but different initialization methods.It can extract a large amount of useful information existed in the unlabeled data,and use the cross-pseudo supervised loss to control the prediction of the two baseline models to be consistent,To maintain the consistency of the two networks and optimize the loss,two tasks of trough edge recognition and trough area recognition are performed in each sub network,which can improve the recognition accuracy at the trough edge.Finally,experiments show that adding rough edge information prediction is conducive to improve the overall recognition accuracy of trough area,and the multi-task cross-pseudo supervision loss can get better results than other commonly used semi-supervised methods.