Abstract:During the prolonged operation of weather radar systems, equipment malfunctions and external interferences frequently result in abnormal radar echoes. These anomalies impair the radar’s effectiveness in monitoring, warning, and forecasting severe weather events. To address this issue, this study manually labeled radial and circular types of abnormal echoes based on historical data, and utilized data augmentation techniques to construct a radar mosaic dataset of abnormal echoes consisting of 20,000 images. Building upon the advanced semantic segmentation model DeepLabv3+, we introduced several optimizations: first, a streamlined ResNet50 backbone with reduced parameters was adopted; second, the SimAM attention mechanism to enhance feature extraction and emphasize critical features was incorporated; third, intermediate layers in the decoder path are added to integrate additional details and contextual information through layer fusion. Consequently, the DeepLab-ARER model for recognition of abnormal radar echo in radar mosaic data was developed. Experimental evaluations demonstrated that DeepLab-ARER achieved superior performance in identifying both types of abnormal echoes, with a mean pixel accuracy (MPA) of 96.75% and a mean intersection-over-union (MIOU) of 93.95%, representing significant improvements over the original DeepLabv3+ model. This research provides robust technical solutions for the automatic recognition of abnormal echoes in radar mosaic data and establishes a solid foundation for enhancing the quality of radar mosaic data in operational applications.