Clutter Covariance Matrix Estimation Based on the CNN and Whitening Metric for Adaptive Detection

Naixin Kang, Weijian Liu, Zheran Shang, Jun Liu, Xiaotao Huang, Jianjun Ge

Abstract: In this article, we address the problem of clutter covariance matrix estimation for radar adaptive detection. Traditional estimation methods are usually based on specific models. However, performance will experience degradation in the presence of model mismatch, which occurs commonly in reality. Therefore, we resort to the data-driven deep learning method and construct a network based on the convolutional neural network (CNN) to estimate the clutter covariance matrix. Besides, due to the unavailable ground truth of the covariance matrix of measured data, simulated data are usually applied for training as a compromise. We design a loss function according to the whitening metric, which makes it possible to train the network directly by measured data. Compared with traditional covariance matrix estimators, the proposed network estimator has higher whitening ability. Moreover, we exploit the obtained covariance matrix estimations to an adaptive detector to evaluate the detection performance. Results with the Intelligent Pixel (IPIX) datasets show that the detector applying the network covariance matrix estimator gains a higher probability of detection (PD).