A Neural Network CFAR Detector Based on Local Clutter Power Estimation Under Unsupervised Learning for HFSWR

High-frequency surface wave radar (HFSWR) has been widely used for ocean monitoring in recent years. However, its target detection faces challenges in complex nonhomogeneous environments. Constant false alarm rate (CFAR) detectors degrade in performance due to inaccurate local clutter power estimation (LCPE), while the performance of data-driven intelligent radar target detectors is limited by the challenges of data annotation and poor generalization. To address the above problems in single-pulse processing, this article integrates CFAR and deep learning theories, and presents a neural network (NN) CFAR detector based on LCPE under unsupervised learning for HFSWR (ULCPE-CFAR). First, by analyzing the relationship between LCPE and CFAR in different environments, we clarify the impact of LCPE accuracy on detection performance, and then theoretically validate the feasibility of NN-based LCPE. Second, based on the above research ideas, we propose the ULCPE-CFAR. The network of ULCPE-CFAR is trained through self-supervised learning on clutter data reconstruction, followed by fine-tuning the optimal model to enhance detection performance in unknown data. During inference, ULCPE-CFAR employs the group masking strategy to mitigate background information loss, ensuring accurate LCPE. Experimental results demonstrate that the proposed ULCPE-CFAR exhibits superior performance compared to traditional detectors in complex environments, providing a reliable and practical solution for unsupervised intelligent radar detection.