Prior Knowledge-Aided Contrastive Learning for HFSWR Target Detection

Abstract

High-frequency surface-wave radar (HFSWR) is widely used for coastal surveillance. Traditional detectors with embedded statistical priors achieve stable performance in homogeneous environments but are limited in complex and time-varying environments, while data-driven supervised approaches require costly large-scale labeled datasets despite their superior performance. A practical strategy to overcome the label scarcity is to generate labels from traditional detectors, yet the inevitable noisy labels degrade supervised methods. To address this, based on the common contrastive mechanism underlying constant false alarm rate (CFAR) detection and contrastive learning (CL), this article proposes a neural network detector trained via prior knowledge-guided CL. First, motivated by the Neyman–Pearson criterion, we highlight the essential requirements for radar detection in a reduced-dimensional feature space: information preservation and class separation. Second, using distributionally robust optimization and feature space analysis, we theoretically validate CL as a robust framework that meets both requirements under label noise. Third, to further improve separability, we propose a prior-guided contrastive learning framework that leverages category priors from CFAR-generated pseudolabels to guide network training. Experimental results demonstrate that the proposed method outperforms the comparative detection methods and achieves over 96% of the detection probability of models trained with manually annotated labels, validating its effectiveness in practical HFSWR scenarios.