AF‐UNet: Intelligent Recognition Model for Atmospheric Fronts Based on Adaptive Fuser

Atmospheric fronts are crucial for weather forecasting as they are often associated with intense weather phenomena. Current machine learning methods for identifying atmospheric fronts directly input multiple meteorological element data into networks for training and prediction, without considering feature conflicts among these elements caused by special weather conditions or geographical factors. Such conflicts can increase training difficulty and reduce identification accuracy. To address this, this study proposes the adaptive fuser (AF)‐UNet model, which integrates an adaptive fusion module (AFM) to resolve feature conflicts in multielement meteorological data and employs a dual loss function during training to mitigate severe class imbalance in labels. Experimental results demonstrate that AF‐UNet outperforms existing methods in accuracy and robustness, achieving a critical success index (CSI) of 69.8%, probability of detection of 79.8%, and accuracy of 84.6%. Our work tackles the challenge of feature conflicts in multielement meteorological data that impair network training and recognition accuracy in current machine learning approaches, providing an effective solution to significantly improve atmospheric front identification.