DeepLab Network for Meteorological Trough Line Recognition

Abstract

A meteorological trough line recognition method is proposed in this paper, in which a DeepLab network that adopts an encoder-decoder architecture is utilized to classify each point in the meteorological grid data into two categories: trough point or not, and then the trough area with the strongest horizontal convergence in the low-pressure area will be identified. The meteorological elements data related to the formation of trough includes the air pressure, the wind velocity and the temperature on 500hp, while the labels are marked with trough lines manually, they are used to train the network model. The proposed method first uses the Deeplab model to recognize the trough area from the meteorological elements data and then extracts the trough line from the trough area by skeleton line extraction algorithm. To evaluate our proposed method, the quantitative experiments were conducted and the results show us that the precission rate of proposed method performances better than the traditional method.