A Semisupervised Approach Using the Adaptive Threshold Mechanism and Boundary-aware Learning for Radio-frequency Interference Segmentation

In the task of radio-frequency interference (RFI) detection, the scarcity of high-quality labeled samples limits the performance of supervised models, while the severe class imbalance between background and RFI makes semisupervised methods prone to confirmation bias during pseudolabel generation, resulting in the weakening of foreground regions and blurred boundaries. To address these issues, this paper proposes a semisupervised segmentation method optimized for class imbalance. First, we introduce an adaptive threshold mechanism that dynamically calculates independent thresholds for RFI and background classes based on convergence difficulty and accumulated predictions. This alleviates the defects of fixed thresholds, which tend to bias the model toward the background class and fail to reliably select unlabeled foreground samples. Second, a consistency-weighted fusion strategy is designed to integrate predictions from statistical rule-based methods and deep neural networks, further enhancing the reliability of pseudolabels. Finally, a boundary-aware module is introduced to explicitly reinforce feature modeling in edge regions. This reduces the boundary ambiguity caused by oversmoothing in sparse foreground regions. Experimental results demonstrate that the proposed method significantly outperforms Only-Sup training across various labeling ratios. The average F1 and F2 scores improve by approximately 9% each, and the Boundary F1 improves by approximately 10%. Under an extremely low labeling ratio of 1/64, the F1 score of the proposed method increases from 64.67% to 70%, the F2 score increases from 60.28% to 69.34%, and the Boundary F1 increases from 70.68% to 83.63%.