Improving Principal Component Analysis Performance for Reducing Spectral Dimension in Hyperspectral Image Classification

High dimensional data will cause problems in classification tasks and storage. Hyperspectral image is a 3D image with hundreds of spectrals. A dimensional reduction method is required to solve problem. PCA is a well known and common method for reducing data dimension. However, the standard PCA needs to be extended in order to improve its ability in hyperspectral image classification. From the previous study, PCA have limited capability on extracting information from the spectral in hyperspectral image. In this study, we propose a new method to improve PCA performance: as a feature extraction method and a dimensional reduction method. We have inserted a hidden layer to transform the data into a new dimension and used PCA to extract the information. The experiment was conducted using Indian Pines hyperspectral image which contains 200 spectrals. We also use datasets from UCI repository such as Iris and Seed datasets. The results showed that the proposed method is able to increase the standard PCA performance and it is comparable to Nonlinear PCA.