Physics and statistics co-enhanced Gaussian process regression for efficient modeling of the conducting target's RCS

In this paper, a physics and statistics co-enhanced Gaussian process regression (GPR) for efficient and accurate radar cross section (RCS) modeling of conducting targets. This study introduces two key innovations. First, we develop an advanced covariance function, termed physical optics-spectral mixture (POSM) covariance function, based on the physical optics (PO) in physics and the spectral mixture method (SMM) in statistics to improve the accuracy and applicability of GPR in modeling the target's RCS. Second, we propose an empirical spectral density-based initialization method for the POSM covariance function, enabling GPR faster converge during training. Experiments conducted with simulated data (involving the NASA almond model, the SLICY model, and a scale-down missile model) and measured data (obtained from the physical model of the missile) demonstrate the superiority of the proposed GPR. It achieves up to an 78.72% reduction in RMSE for simulated data and 69.68% for measured data compared with other alternative covariance function-based GPRs. In terms of efficiency, the training time is reduced by more than 33%, and the well-trained GPR can model the target's RCS in near-real-time (within 0.06 seconds), indicating great potential of our GPR for practical applications in RCS characteristic analysis and data processing like imputation and augmentation. In addition, compared with other alternative machine learning algorithms, such as deep learning (DL), decision tree (DT), and support vector regression (SVR), the proposed POSM-GPR also shows superior precision and respectable efficiency in RCS modeling.