Deep support vector machine for hyperspectral image classification

Highlights

• Deep Support Vector Machine (DSVM) introduced in hyperspectral image classification.

• DSVM outperformed five other classification algorithms, including Deep Neural Network.

• DSVM is a state-of-the-art algorithm to improve hyperspectral image classification.

Abstract

To improve on the robustness of traditional machine learning approaches, emphasis has recently shifted to the integration of such methods with Deep Learning techniques. However, the classification problems, complexity and inconsistency in several spectral classifiers developed for hyperspectral images are some reasons warranting further research. This study investigates the application of Deep Support Vector Machine (DSVM) for hyperspectral image classification. Two hyperspectral images, Indian Pines and University of Pavia are used as tentative test beds for the experiment. The DSVM is implemented with four kernel functions: Exponential Radial Basis Function (ERBF), Gaussian Radial Basis Function (GRBF), neural and polynomial. Stand-alone Support Vector Machines form the interconnecting weights of the entire network. The network is trained with one hundred input datasets, and the interconnecting weights of the network are initialised using the regularisation parameter of the model. Numerical results show that the classification accuracies of the DSVM for Indian Pines and University of Pavia based on each DSVM kernel functions are: ERBF (98.87%, 98.16%), GRBF (98.90%, 98.47%), neural (98.41%, 97.27%), and polynomial (99.24%, 98.79%). By comparing the DSVM algorithm against well-known classifiers, Support Vector Machine (SVM), Deep Neural Network (DNN), Gaussian Mixture Model (GMM), K Nearest Neighbour (KNN), and K Means (KM) classifiers, the mean classification accuracies for Indian Pines and University of Pavia are: DSVM (98.86%, 98.17%), SVM (76.03%, 73.52%), DNN (94.45%, 93.79%), GMM (76.82%, 78.35%), KNN (76.87%, 78.80%), and KM (21.65%, 18.18%). These results indicate that the DSVM outperformed the other classification algorithms. The high accuracy obtained with the DSVM validates its efficacy as state-of-the-art algorithm for hyperspectral image classification.

Introduction

Hyperspectral Image (HSI) has been found invaluable because of its numerous applications and ability to obtain remotely sensed information from the visible through the near infrared wavelength ranges thus providing multi-spectral channels from the same location (e.g. [10], [27]). HSIs are highly innovative remote sensing imageries that consist of hundreds of contiguous narrow spectral bands, which unlike the conventional panchromatic and multispectral imageries enable a better distinct discrimination of object classes [34]. However, the major challenge for scientists is how to efficiently classify HSIs (see, e.g., [17]). Some of these challenges as detailed by Ghamisi et al. [10] include increased presence of redundant spectral information and high dimensionality in observed data, among others. Several conventional unsupervised and supervised machine learning classifiers have been used for classifying HSIs and includes prominent unsupervised conventional classifiers such as Fuzzy C-Means (FCM) and K Means (KM). While notable conventional supervised classifiers (e.g., K Nearest Neighbour (KNN) and Gaussian Mixture Model (GMM)) have been used in the classification of HSIs, the use of contemporary classifiers such as Support Vector Machine (SVM) and Artificial Neural Network (ANN) are gradually emerging [20], [31].

But recently, emphasis has shifted from conventional methods to the integration of Deep Learning (DL) and ANN, which scientists have argued grossly enhanced the robustness of the traditional ANN. For example, Paoletti et al. [28] showed that the use of DL to train the conventional ANN significantly increased the efficiency of the ANN for HSI classification. Haut et al. [11] implemented an integrated Deep Convolutional Neural Network (DCNN) using a new Bayesian approach and found that the hybrid of DL and the traditional ANN classifier enhanced the efficiency of the traditional ANN classifier. Furthermore, a novel guided filter based Deep Recurrent Neural Network (DRNN) for HSI classification has proved to be more efficient than the traditional ANN classifier [8]. Zhao et al. [35] recently proposed an integrated Convolutional Neural Network (CNN) and Gray Level Co-occurrence Matrix (GLCM) textural features for HSI classification using limited training sample. More recently (e.g., [19]), other DL ANN algorithms such as the use of Deep Belief Network have been employed in HSI classification with major highlights on their merits as opposed to conventional methods. However, classification problems associated with small-size training dataset in traditional machine learning techniques have been reported (e.g., [4]).

Recent advances in convolutional neural networks, including activation function, loss function, regularization, optimization and fast computation have been documented. Gu et al. [7] who provided details on these advances also highlighted the weakness of CNN, indicating computational efficiency and the choice of a suitable hyper-parameters (e.g., learning rate, kernel sizes of convolutional filters) are still challenging issues, especially for large-scale data. The use of a new CNN architecture for the classification of hyperspectral images was therefore predicated on the computational constraints of CNN algorithms to high-dimensional data contained in multidimensional data cubes [28]. Although the application of deep learning techniques, especially CNN in image-based cancer detection, diagnosis and other disciplines have shown significant strength [13], [18], improvements are required to handle large-scale multi-resolution data cubes. It is against this background, that assessing the skills of other non-parametric deep learning algorithms such as the deep SVM has become necessary.

In addition to other classification methods such as random forests, neural networks, and logistic regression-based techniques, the SVM [5] is another robust classifier that has been used in hyperspectral data classification (e.g., [2], [10]). Since the introduction of the SVM it has proven to be very efficient in remote sensing (RS) image classification, tide analysis, and prediction of urban land use change (e.g., [24]). Due to SVM's ability to model complex real-world data, they were found to be relatively better predictive models as opposed to agent-based models whose inability to evaluate model constraints and results have been highlighted in previous studies (e.g., [23], [25], [29], [30], [36]). Even though new algorithms such as the Supervised Fuzzy Partitioning are now competitive with the SVM, the latter is still largely characterized as a state-of-the-art algorithm [1]. SVM is intrinsically a binary classifier, it can be modified however, for multi-class problems by using mainly the One Against All (OAA) or One Against One (OAO) technique [14]. The OAA and OAO techniques have proven to be considerably effective in the classification of remote sensing images [26].

While classification problems still exist with traditional machine learning techniques, the complexity (e.g., availability of training samples) surrounding the implementation of different classification algorithms are some reasons warranting further research on ideal techniques for HSI. This was echoed in Ghamisi et al. [10] who noted the inconsistency in several spectral classifiers developed for HSI based on selected metrics. Furthermore, the use of deep belief network in improving classification outputs of hyperspectral images and multi-temporal images has recently been demonstrated [15], [19]. Whereas several parametric and prominent non-parametric algorithms have been widely used in image classification (see, e.g., [10], [20], [31]), the assessment and accuracy of HSI classification based on Deep Support Vector Machine (DSVM) however, is largely undocumented. One of the key challenges with HSI classification is limited training samples. It is for this reason that the development of new optimization algorithms such as deep learning is increasingly becoming popular and effective in the fields of image recognition and classification, especially for the classification of large multi-spectral and hyperspectral datasets [28]. Moreover, the success of these new algorithms in automatic feature extraction, computer vision, language processing and speech recognition have recently been re-echoed [17]. There are still some constraints nonetheless, on the application of these methods to multispectral and hyperspectral images. A regularized ensemble framework of deep learning that incorporates SVM is therefore crucial to further explore these challenges.

Arguably, deep learning algorithms have attracted the attention of the remote sensing community and several other experts in the fields of speech recognition, computer vision, and natural language processing among others. To explore the application of deep learning methods in hyperspectral image classification, a multi-grained network that appears to be an ensemble deep learning method has been proposed [27]. Another hybrid model that integrates unsupervised deep belief network with a one-class SVM were found to be scalable and computationally efficient in an earlier study [6]. While theoretical foundations and optimization techniques for learning deep CNN architectures are required [7], ensemble deep learning methods have shown considerable potentials in the classification of HSI classification. For example, the coupling of deep belief network with a SVM algorithm addressed complexity and scalability issues of SVM in large datasets [6]. The use of hybrid models is therefore emerging as efficient, accurate and scalable techniques that can improve the classification accuracy of large-scale and high-dimensional data.

The aim of this research therefore is to integrate DL and SVM to formulate a hybrid DSVM. The architecture of the DSVM proposed for this experiment imitates the Deep Neural Network (DNN) that consists of multiple hidden layers. Normally the hidden layer neurons are connected by series of weights; but instead the weights are initialised by several SVM functions modified with the SVM regularisation parameter. The optimal DSVM output for each input is found by updating all the connecting SVM functions in the hidden layer. Two HSIs, Indian Pines and University of Pavia are used as experimental and tentative test beds for the study. The OAA multi-class technique are used to modify the SVM for multi-class separation. To assess the robustness of our hybrid DSVM model, the results of the DSVM are compared to those of the SVM, DNN, GMM, KNN, and KM.