A Novel Deep Learning Framework by Combination of Subspace-Based Feature Extraction and Convolutional Neural Networks for Hyperspectral Images Classification

Approaches based on deep learning have gained an increased attention in the recent years in particular Remote Sensing. Convolutional Neural Networks (CNNs) as one of these deep learning techniques has demonstrated remarkable performance in visual recognition applications. However, using well-known pre-train models such as GoogleNet and VGGNet in the area of hyperspectral image classification due to the high dimensionality and the insufficient training samples is intractable. The current study proposed a new and fixes CNN architecture for two real hyperspectral data sets. To overcome curse of dimensionality we perform a subspace-based feature extraction method by calculating the orthonormal basis of correlation matrix for each class to reduce the dimensionality of hyperspectral images and increasing signal to noise ratio. This framework combines the proposed CNN architecture and subspace reduction method to prepare informative features (from subspace method) and designing optimized CNN by considering limitation of training samples. Also, feature generated by subspace reduction method is compatible by the nature of class based CNNs and a logistic regression as a classifier in the last layer of proposed architecture. Experimental results from two real and well-known hyperspectral images, the Indiana Pines and the Pavia University scenes show that the proposed strategy leads to a performance improvement, as opposed to using the original data and conventional feature extraction strategies which have been employed during the recent approaches. The classification overall accuracy of ca. 98.1% and 98.3% were obtained in Indian Pine and Pavia University respectively.