Sea Clutter Covariance Matrix Estimation and Adaptive Detection Algorithm Based on Doppler Focusing Complex-valued Convolutional Neural Network

In radar adaptive detection under sea clutter backgrounds, traditional covariance matrix estimation methods are dependent on prior statistical models. Model mismatches are easily caused under non-homogeneous and small-sample conditions, and the detection performance is thus deteriorated. To address this issue, a data-driven covariance matrix estimation algorithm based on Doppler focusing complex-valued convolutional neural network is proposed in this paper. Firstly, a three-channel input structure integrating cells under test, auxiliary cells and prior information is established, and multi-dimensional amplitude-phase features of sea clutter are extracted by a complex-valued residual dense network. Subsequently, to enhance the estimation accuracy of key frequency bands, a Doppler focusing strategy is adopted during the training process. A sample weight function is designed based on clutter Doppler frequencies and clutter-to-noise ratios to realize frequency-domain weighting of the loss function, and feature analysis of clutter-dominated regions is guided for the network. Finally, the estimated covariance matrix is imported into the adaptive normalized matched filter for detection judgment. Validations based on simulated data and intelligent pixel processing X-band radar data demonstrate that compared with traditional complex-valued networks and generalized inner product-based sample screening methods, estimation errors induced by insufficient samples are effectively mitigated by the proposed algorithm in non-homogeneous scenarios, and the target detection probability is greatly improved.