Abstract: In radar adaptive target detection under sea clutter background, the accurate estimation of the clutter covariance matrix is critical to detection performance. Traditional covariance matrix estimation methods typically rely on specific statistical models (e.g., compound Gaussian models). When the actual clutter environment mismatches the assumed model, especially under heterogeneous sample conditions, the estimation accuracy degrades significantly, leading to detection performance loss. To address this issue, this paper proposes a data-driven covariance matrix estimation algorithm based on a Doppler-Focused Complex-Valued Covariance Estimation Network (DF-CVCENet). The algorithm constructs a three-channel input comprising primary data, secondary data, and prior information, and employs a complex-valued residual dense network to fully exploit the amplitude and phase characteristics of clutter data. Addressing the Doppler characteristics of sea clutter, a Doppler focusing strategy is introduced during the training stage: based on the known clutter Doppler frequency and clutter-to-noise ratio, a sample weight function is designed to weight the loss function, guiding the network to focus more on feature learning in clutter-dominated regions, thereby improving covariance estimation accuracy in critical areas. The estimated covariance matrix is then incorporated into the Adaptive Normalized Matched Filter (ANMF) for target detection. Validated through simulated data and IPIX radar measured data, and compared against the standard complex-valued network (CVCENet) and the covariance estimation algorithm based on Generalized Inner Product (GIP) sample selection, the results demonstrate that DF-CVCENet effectively overcomes estimation biases caused by insufficient samples under heterogeneous clutter environments. Its detection probability significantly outperforms the compared methods, verifying the effectiveness and superiority of the proposed algorithm.