Self-Learning Parameter Estimation of K-Distributed Clutter Using GRU Network

Robust and accurate parameter estimation of K-distribution clutter plays an import role in target detection for marine radars. The estimation performance of existing estimators is limited by only two moments or percentiles. To break this limit, a self-learning estimator using gate recurrent unit (GRU) network is proposed to estimate shape parameter by a well-designed feature vector composed of ten moment ratios and thirteen percentile ratios. It is theoretically proved that the feature vector is independent of scale parameter. Then, the scale parameter is determined by a shape-parameter-dependent percentile. Finally, it is verified by simulated data and measured data that the proposed estimator is suitable for the complicated and various clutter environments.