Collaborative Superpixelwised PCA for Hyperspectral Image Classification

Extracting spectral-spatial features from Hyperspectral imagery (HSI) has been proven to be efficient for classification tasks. A recently developed superpixelwised principal component analysis (PCA) (SuperPCA), which has shown its promising performance, is a prominent technique in spectral-spatial feature extraction. However, we have discovered that SuperPCA may lead to an intraclass dispersion problem, which can result in a decrease in classification accuracy. In this article, a novel method called collaborative superpixelwised PCA (CSPCA) is proposed to address this issue. The main idea behind CSPCA is to collaboratively learn the projections for each superpixel. Specifically, CSPCA first employs a superpixel segmentation technique to generate superpixels. Next, the mean vectors of samples within each superpixel are utilized to model the manifold structure of the data. Then, a novel objective function is formulated, which aims to simultaneously preserve the obtained manifold structure between superpixels and the structure within each superpixel. To optimize the objective function, the Manopt toolbox is employed in the proposed method. The effectiveness of the proposed approach is validated through experimental evaluations conducted on five HSI datasets.