2026AJ....171...94C

Radio telescopes are susceptible to radio frequency interference (RFI) during observations, which introduces noise and artificial signals into astronomical data. Failure to properly process RFI-contaminated data can severely compromise data reliability and even lead to erroneous scientific conclusions. Consequently, RFI identification and mitigation have become critical scientific challenges in radio astronomy. This study proposes an RFI recognition and data cleaning method based on a deep learning image classification algorithm (RFI-MobileNetV2, RFI-MN). By performing feature extraction and preliminary classification on visual astronomical data, the method aims to effectively mitigate RFI, improve the data signal-to-noise ratio, and provide more reliable data support for astronomical research. Using the lightweight convolutional neural network MobileNetV2 as the core architecture, the model achieves efficient extraction and classification of diverse RFI features, thereby providing an effective solution for subsequent RFI suppression tasks. For the experiments, a dataset containing multiple RFI types was constructed as a training sample. During model optimization, performance was significantly enhanced through the integration of an attention mechanism, modification of the activation function in inverted residuals, and multiscale feature fusion. Evaluation metrics including precision, recall, and F1 score were employed to verify the scheme's effectiveness through baseline training, inference, multimodel comparison, and ablation experiments. The results demonstrate that the optimized RFI-MN achieves performance exceeding 94%, substantially outperforming other comparative models and providing an effective solution for RFI-affected astronomical data preprocessing.