Spectral-Spatial Feature Extraction Using PCA and Multi-Scale Deep Convolutional Neural Network for Hyperspectral Image Classification

Abstract

Hyperspectral imaging (HSI) is one of the emerging research fields in remote sensing technology as it contains huge information about a scene which can be further analyzed to detect various kind of objects. HSI contains several narrow and contiguous spectral bands which provide different research challenges for dealing with high dimensional data for better classification. Most of the traditional method takes only the spectral feature into account for classification. Principal Component Analysis (PCA) is widely used for this task. Recently with the introduction of the convolutional neural network, spatial features are being fused with the spectral features for better classification. However, traditional convolutional neural network (CNN) based methods take only a single scale spatial kernel in order to explore the spatial information. In this paper, we have proposed a hybrid approach of a dimensionality reduction technique based on PCA and a novel multi-scale convolutional neural network named PCA-MS-CNN in order to extract spectral and spatial feature for hyperspectral image classification. We have achieved an average accuracy of 99.10% on Indian Pines dataset using this method. This result shows that it outperforms other state-of-the-art deep learning based methods used for HSI classification.