A Methodology for Unsupervised Feature Learning in Hyperspectral Imagery Using Deep Belief Network

Deep learning approaches have received major interest in the field of remote sensing. Hyperspectral imaging has rich data that are distributed in multi-dimensions. It is challenging to apply deep learning algorithms due to the limited amount of labelled data. So, unsupervised feature extraction approaches are used to overcome this limitation. In this study, we propose an unsupervised feature learning approach using deep belief network (DBN). In the proposed framework, the input hyperspectral image is segmented using entropy rate superpixel segmentation and filtered by domain transform recursive filter which extracts spatial and spectral information effectively. Then the features are learned by improved DBN. In the traditional methods, DBN is stacked with restricted Boltzmann machine (RBM) which is suitable for only binary value data. In the proposed framework, we used Gaussian–Bernoulli RBM which was constructed for real value data such as images. The experiments were carried out using Pavia University dataset. The results show that the proposed network has good performance in terms of classification accuracy and computation time.