An Intelligent Identification Method of Fronts Based on Air Mass Labels

DING Xinya; LI Qian; WANG Tianying; ZHANG Liang; LIU Yudi; ZHANG Yunpeng; HUANG Bing; FENG Xiao

doi:10.16032/j.issn.1004-4965.2024.086

Abstract：Current machine learning methods for the automatic identification of fronts often face challenges due to a serious imbalance in the proportion of frontal grid points versus non-frontal grid points in the training labels.This imbalance can lead to biased recognition results in favor of the non-frontal category.Moreover,the input of multiple meteorological elements may result in data feature conflicts or poor quality data due to special weather conditions and geographic variations,leading to mismatches between input data and the network.Consequently,this affects the training process and recognition accuracy.To address these issues,we proposed a method that trains the AMA-UNet model for the intelligent identification of fronts based on air mass labels.This approach used multiple meteorological parameters from the ERA5 dataset provided by the European Centre for Medium-Range Weather Forecasts as network inputs,while generating air mass labels from the front dataset provided by the Weather Prediction Center in the U.S.This effectively mitigated the imbalance between non-frontal and frontal categories.Furthermore,the adapter in the AMA-UNet architecture was utilized to resolve the mismatch between the input data and the network,which facilitated network training and improved the comprehensive performance of the network.Experiments show that the use of air masses as labels improved evaluation metrics by approximately 5%compared to networks trained directly using fronts as labels.Moreover,incorporating adapters yielded an average improvement of about 3%across multiple evaluation metrics.This method demonstrates significant enhancements in all evaluation indexes compared with other methods.