Applied Mathematics & Information Sciences

Author Country (or Countries)



In this paper, a novel Multi-Dimensional Recurrent Deep Neural Network is proposed for classifying hyperspectral images. Deep Learning Networks have developed rapidly with applications in several fields including computer vision, healthcare, bioinformatics and machine learning. Multi-Dimensional Recurrent Deep Neural Networks are a special case of directed acyclic graph networks in which standard Recurrent Neural Networks are realized by giving recurrent connections along all spatio-temporal dimensions of the data and the recurrent connection size is equal to the dimension of the data. In this work, two Recurrent Neural Networks are replaced by one Multi-Dimensional Recurrent Deep Neural Network to learn middle-level visual patterns and spatial dependencies between them. In the last stage, fully connected layers are used to learn a global image representation. Due to the recurrent connections, this method is robust to local distortions such as image rotation and shear. Without suffering from scaling problems, it brings additional advantages over Recurrent Neural Networks to multi-dimensional data. This paper investigates hyperspectral image classification with the proposed network and the results have been validated with hyperspectral datasets namely Pavia University and Salinas images. There is an improvement in the classification accuracy of this newly proposed method in comparison to classical methods like Support Vector Machine, Convolutional Neural Network (CNN) and Recurrent Convolutional Neural Network (RCNN).

Digital Object Identifier (DOI)