Comparative analysis of 2D-PCA based dimensionality reduction and feature extraction

Abstract

The complex spectral and spatial characteristics of hyperspectral remote sensing images (HSI) lead to higher time-consuming in classification task. To address this question, we introduced the 2D-PCA dimensionality reduction method of linear mapping in the two-dimensional spatial domain on the basis of linear dimensionality reduction in the spectral domain, thereby compressing the complex spatial structure information of HSI into a limited low-dimensional space, and realizing space-spectrum dimensionality reduction and information fusion. The experimental results on three classic data sets of Salinas, Tea Farm, and Indian Pines show that 2D-PCA has a strong ability to condense and compress spatial structure characteristics. Compared with popular deep learning frameworks such as CNN and Mixer-MLP, conventional machine learning models based on 2D-PCA have significant advantages in terms of computing time under the premise of controllable accuracy loss, which makes 2D-PCA a promising method for dimensionality reduction and feature expression in hyperspectral pixel-wise classification.