An automatic trough line identification method based on improved UNet

Abstract

A trough is an elongated region of relatively low atmospheric pressure. Automatic analysis and recognition of trough lines in upper weather charts are challenging and of great significance to weather analysis. The existing automatic identification methods mainly depend on manually setting rules which cannot cover all trough line types and have low generalization ability. This paper proposes an automatic trough line identification method based on an improved model which can extract the trough line from meteorological element data at 500 hPa. The model adopts the UNet of a U-shaped encoder and decoder as basic structure, which is designed to enable precise localization by continuously combining low-level and high-level features. To extract abstract semantic features of the trough, the Xception, which takes depthwise separable convolution as basic unit, is adopted to replace the encoder of the original UNet. In addition, the Squeeze and Excitation (SE) module with an attention mechanism is added after each ordinary convolution in the decoder part to improve the recognition accuracy by increasing the weighting of the trough area. The experiments are conducted on a meteorological dataset and the results show that the recognition accuracy with our proposed method on the testing dataset can reach over 80%. We also compare our results to several other types of networks and traditional automatic identification methods, which demonstrates that the performance of the proposed network is superior to other methods.