A Novel Deep Hybrid Spectral Network for Hyperspectral Image Classification

Abstract

Image classification is the process of allocating land cover classes to picture elements. Hyperspectral image classification of land cover is not easy due to the problematic variability among the samples for training count of band spectra. Deep learning has network which is capable of learning from unstructured or unlabeled data without supervision. The convolutional neural network (CNN) is one of the most widely employed approaches for visual data processing based on deep learning. To classify the hyperspectral data, a hybrid network with 2D and 3D CNN is developed. The spectral information and spatial information are used together for hyperspectral image analysis that enhances the experiment result considerably. With pixel as the basic analysis unit, classification technique of convolutional neural network has been developed. Principal component analysis (PCA) is implemented which lowers the dimensionality of hyperspectral data. PCA implementation reduces feature size to increase computational efficiency. The first principal component has a superior function, as it has the highest variance compared to the other components. Three hyperspectral datasets are used for the analysis such as Pavia university, Indian pines, and Salinas scene. Hyperspectral data of Indian Pines is obtained from the Northwest Indiana on June 1992 by Airborne Visible Infrared Imaging Spectrometer (AVIRIS). Indian pine data has the size (145 × 145) pixels. The Pavia university is obtained from Reflective Optics System Imaging Spectrometer (ROSIS) Pavia, Northern Italy, in 2001. The data has 103 spectral bands with the size of 610 × 340 pixels. The Salinas scene is obtained from the AVIRIS Salinas Valley, CA, USA, in 1998, with 512 × 217 spatial dimension. The hybrid CNN is computationally efficient compared to the 3D CNN and for minimum training data it provides enhanced performance.