Application of linear and nonlinear PCA to SAR ATR

Abstract

This paper explores the use of Principal Component Analysis (PCA) techniques for the development of classification systems for Synthetic Aperture Radar (SAR) Images. The concept of Principal Component Analysis is centered on feature extraction and dimensionality reduction. Through the exploitation of spatial differences and variances between data points of a specific data domain, application of PCA techniques allows the reduction of datasets to representations consisting of principal components only. The effect hence forth being the reduction of dataset sizes, which translates to a reduction in processing time on these datasets, for almost any application the mathematical technique is applied to. Open literature provides examples of software computation domains to which PCA has been applied, examples being face recognition and geo-environmental forecasting applications. Both linear and nonlinear PCA forms are covered in this paper. Application of linear PCA to SAR based automatic target recognition has been covered extensively in open literature. This investigation therefore aims to improve on the performance achieved by linear PCA application, using non linear PCA. Three systems were developed for the purpose of the investigation, which were a linear PCA system, a nonlinear PCA system using a polynomial kernel, and a nonlinear PCA system using a Gaussian kernel. The systems were tested for how well they responded to a reduction in training dataset, as this is a real-world problem experienced in ATR systems. The performance of the systems in terms of their running times were also evaluated. As anticipated, the nonlinear PCA approach outperformed the linear PCA approach, and the performance of the polynomial kernel PCA system was observed to be the best of all the three systems.