The effect of the principal component analysis in convolutional neural network for hyperspectral image classification

Hyperspectral imagery is widely used in image classification due to their continuity and the abundance of spectral information. Recently, manifold deep learning algorithms were proposed for hyperspectral image classification (HSIC) demonstrating a competitive edge in performance over existing methods due to their ability to automatically extract high-level features. However, one significant drawback of hyperspectral images is their high dimensionality, which increases the learning time and processing complexity. To address this problem, several studies have exploited principal component analysis (PCA) as a pre-processing step in the deep learning framework to compress the entire image through a simple statistical processing of the spectral dimension while preserving the spatial information. However, since PCA can result in a loss in the spectral information, other studies have applied the original hyperspectral imagery as input to deep learning networks. At the same time, the impact of dimensionality reduction using PCA in deep learning networks on achieving efficient HSIC is understudied. Hence, the purpose of this study is to analyze the effect of PCA in deep learning for HSIC. In this paper, we verified the efficiency of deep learning networks through various conditions of PCA. We employed a convolutional neural network (CNN), which can extract spatial-spectral features of hyperspectral imagery. To analyze the sensitivity of PCA depending on CNN architectures, a two-dimensional CNN (2D-CNN) and a three-dimensional CNN (3D-CNN) were applied. We quantitatively analyzed the experimental results, which revealed that PCA can effectively reduce an image to its optimal spectral dimension according to CNN models for efficient CNN-based HSIC.