Machine Learning-based Modeling and Uncertainty Quantification for Radar Cross Section of a Cone-like Target

Radar cross section (RCS) plays an important role in the recognition of targets. RCS varies dramatically with the incident angle and the size of targets, and it is difficult to accurately predict the RCS values. In this paper, an efficient modeling and uncertainty quantification method based on the support vector machine and the k-nearest neighbor is proposed for the RCS prediction of cone-like targets. The proposed method is compared with two uncertainty quantification methods, an ensemble based on lower upper bound estimation and a neural network with dropout. The root mean square error, the prediction interval coverage probability, the mean prediction interval width and the computation time are used as the performance metrics, and the experimental results demonstrate that the proposed method is superior to the compared methods in accuracy and efficiency.

I. Introduction

The appearance, size and posture of a target and many other factors are all able to affect its radar cross section (RCS) values, and the RCS is a key element in the target recognition [1]. To advance electromagnetic scattering theory and applications, many computational electromagnetics approaches about RCS have been developed, and they can be divided into three main categories, namely, exact analytical methods, numerical calculation methods, and high-frequency asymptotic methods [2]. Despite the rapid development of the hardware and software, these approaches usually cost plenty of computational, physical and manpower resources, especially when uncertainties are taken into account [3]– [5].