Application of the Mask R-CNN model to cold front identification in Eurasia

Abstract

Cold fronts often bring catastrophic weather events, which are exacerbated under global warming. Thus, the automatic and objective identification of cold fronts will be helpful for accurate forecasting and comprehensive analysis of cold fronts. Recently, machine learning methods have been applied to meteorological study. In this study, a cold front identification method based on the deep learning model Mask R-CNN is proposed to automatically identify cold fronts from massive data. The Mask R-CNN method shows high accuracy after the comparison with traditional methods and is effective for identifying the cold fronts in both continuous time and extreme precipitation events. Based on the obtained cold-front samples, we conduct some statistical analysis. The results show that the frequency of cold front is unevenly distributed over Eurasia, with the highest in the Daxing'anling region and the mid-latitude storm axis, especially in winter. The method and results presented in this study may have some implications for the application of deep learning models in weather system identification.