Abstract:Weather radar systems operating continuously over long periods are often affected by equipment failures, external interference, and other factors, which may lead to abnormal echoes and limit their effectiveness in monitoring, warning, and forecasting severe weather.In this study, two types of abnormal echoes, radial and annular echoes, were manually selected and labeled from historical radar mosaic data. A weather radar abnormal echo dataset for radar mosaics, WRAED, containing 20 000 images, was then constructed using data augmentation techniques.An improved model was developed based on the semantic segmentation model DeepLabV3+. A simplified ResNet50 architecture was adopted as the backbone network to reduce computational cost. A simple attention mechanism, SimAM, was added to the backbone feature extraction layers to help the model focus on important echo features. In addition, an intermediate layer was introduced into the decoder to incorporate more detailed and contextual background information.The improved model, named DeepLab-ARER, was designed for abnormal echo recognition in radar mosaics. Experimental results show that DeepLab-ARER achieves good performance in identifying abnormal echoes, with a mean pixel accuracy of 96.75% and a mean intersection over union of 93.95%, representing a clear improvement over DeepLabV3+.The DeepLab-ARER model provides effective technical support for the automatic identification of abnormal echoes in radar mosaic.