Hyperspectral Image Classification Using Modified Convolutional Neural Network

Abstract

Classification of hyperspectral images (HSI) signifies giving each and every pixel in the image a class on the basis of some information obtained. Convolutional neural networks (CNNs), in recent years, have been widely explored and applied for HSI classification as they have shown better performance than other existing techniques in terms of accuracy of classification. Most of the existing techniques such as traditional techniques using manually crafted features, feature reduction techniques like independent component analysis (ICA), decision trees (DT) and support vector machines (SVMs) consider only the spectral information for classification. They are unable to utilize the spatial information adequately which hinders classification performance. CNN has shown great promise as it is able to conceive the spatial information along with the spectral information thus enhancing classification accuracy. However, as HSI has very high dimensionality and insufficient samples for training, effectively and efficiently applying CNNs for classification of HSI remains a demanding task. Therefore, in this paper, we have proposed a modified CNN architecture to boost its discriminative capability for extraction of both spectral and spatial information from original HSI and using those information predict classification results using a softmax regression (SR) classifier. The proposed architecture is different from conventional CNN models as it makes use of 1×1 filters along with FxF filters during convolution operations in order to properly cope with the HSI information. A global average pooling layer (GAP) replaces the traditional fully connected (FC) layers in the proposed architecture which helps to alleviate the problem of overfitting to some extent. Comparing our work with considered techniques on two benchmark datasets, it is observed that our proposed approach gives better results in terms of overall accuracy (OA), average accuracy (AA) and kappa coefficient (K).