Validation of PCA and LDA for SAR ATR

Both principal component analysis (PCA) and linear discriminant analysis (LDA) have long been recognized as tools for feature extraction and data analysis. There has been reports in the open literature regarding the performance of both LDA and PCA as feature extractors in various types of classification and recognition problems. Many of the reports claim a better performance with LDA than with PCA. However, the grounds of comparison have mostly been quite narrow. In the current paper PCA and LDA based classifiers are evaluated for the problem of synthetic aperture radar based automatic target recognition problem. The results show that in terms of absolute performance, PCA outperforms LDA. Results of PCA based classifier are also found to be of higher confidence than those from LDA based classifiers, as observed from the error-bar analysis of the classifiers. With decreased amount of training dataset, the degradation in the performance of the classifiers are almost similar in nature. The current work concludes that LDA is not suitable for radar image based target recognition task. This is in line with reports from some works in the open literature which claim that the success of LDA will depend on the type of data and whether there is exhaustive data available during the training phase or not.