تصویربرداری فراطیفی توسط یک روش رده بندی نیمه نظارتی موثر

ABSTRACT An efficient semi-supervised classification approach for hyperspectral imagery

Hyperspectral sensors provide abundant spectral information with hundreds of spectrally continuous bands, and the collected hyperspectral imagery can be used to conduct finer classification that could not be achieved by traditional multispectral imagery (Bioucas-Dias et al., 2013; Du et al., 2012; Mountrakis et al., 2011; Tan and Du, 2008, 2011; Tan et al., 2013; van der Meer et al., 2012). Over the past several decades, novel pattern recognition methods have been widely used to remote sensing image processing tasks, such as supervised classification (Li et al., 2013b; Zhong and Zhang, 2012), unsupervised classification (Villa et al., 2013; Yang et al., 2010), feature extraction (Du et al., 2008; Du and Yang, 2008; Jia et al., 2013), target identification (Du and Zhang, 2014; Manolakis et al., 2001), and change detection (Bioucas-Dias and Nascimento, 2008; Camps-Valls et al., 2008).

support vector machines (SVM), as one of the research focuses of machine learning, has attracted more attention in remote sensing (Camps-Valls et al., 2006; Fauvel et al., 2008; Melgani and Bruzzone, 2004; Plaza et al., 2009). SVM has advantages such as less rigid requirements for prior knowledge and training samples, fitness to high-dimensional data, and more robustness to noise (Du et al., 2010; Jain et al., 2000). However, it is often difficult for a traditional SVM classifier to offer satisfactory performance in hyperspectral image classification.

Specifically, for a high-dimensional hyperspectral image with a limited number of training samples, classification accuracy usually is significantly decreased due to the Hughes phenomenon (i.e., for a limited number of training samples, classification accuracy is decreased with the feature dimension being increased) (Hughes, 1968). Moreover, hyperspectral imagery is usually short of training sets, because sample collection generally involves extensive and time-consuming fieldwork (Dopido et al., 2013; Li et al., 2013a).

A relevant advanced solution to this classification challenge is the introduction of semi-supervised learning (SSL) techniques (Camps-Valls et al., 2007; Chapelle et al., 2008). SSL utilizes a large number of unlabeled data together with the available labeled data, to build classifiers that may be stronger than usual. SSL also requires less human efforts in sample collection, so it is of great importance in practical applications. The research on SLL has experienced a quick development in the past few years, and some popular methods have been proposed, such as generative mixture models (Krithara et al., 2011), self-learning models (Dopido et al., 2013), multiview learning models (Culp et al., 2009; Sun and Shawe-Taylor, 2010), transductive support vector machines SVMs (TSVMs) (Bruzzone et al., 2006; Ratle et al., 2010), and graph-based methods (Bai et al., 2013; Camps-Valls et al., 2007).

In particular, semi-supervised SVM (S3VM) has been successfully applied to remote sensing image classification. Marconcini proposed a novel composite S3VM for the spectral–spatial classification of hyperspectral images, which can significantly improve classification accuracy compared to both supervised SVMs and progressive S3VM (Marconcini et al., 2009). Munoz-Mari proposed an algorithm to modify the OC-SVM kernel by modeling the data marginal distribution with a data graph Laplacian (Munoz-Mari et al., 2010). Gu successfully used Laplacian SVM for semi-supervised hyperspectral image classification (Gu and Feng, 2013).

A novel approach for semi-supervised learning that adapts available active learning methods to a self-learning framework was presented by Dopido et al. (2013). The aforementioned SSL methods improve classification performance by different strategies. However, as the number of training samples is increased, it may be unbalanced and unbearable for a classifier to correctly exploit all the proper training samples due to computational issues.

Furthermore, if a mislabeled sample is added for training, it may degrade classification accuracy. Thus, a semi-supervised algorithm should be designed in a computationally efficient fashion with the objective that classification performance can be truly improved with the use of additional samples. In this paper, we evaluate the feasibility of adapting spatial information into an SSL paradigm, in which the S2SVMSE itself selects accurate unlabeled samples for training.

It makes use of spatial information in selecting new unlabeled training samples and generating classifier ensemble. It is assumed that samples in a homogeneous spatial segment mostly likely belong to the same class, which helps generate a smooth classification map; meanwhile, samples in the same segment also exhibit within-class variations, which can be useful for classifier training. Up to now, few studies that employ spatial information to select unlabeled samples are reported in the literature.

Shi et al. proposed a semi supervised dimension reduction algorithm, where unlabeled samples were selected based on multilevel segmentation results and t spectrally nearest neighbors similarity measure, and it notes that the number of the neighbors t controls the accuracy of similarity measurement, so they recommended to set a larger t to relax the constraint (Shi et al., 2013). However, it could not perform perfectly with limited or few labeled samples, so a-one-nearest neighbor is proposed to take the action of similarity measurement when labeled samples are very limited. In this work, the mean shift algorithm is used for image segmentation and spatial feature extraction, and a strongly constrained algorithm called a-one-nearest neighbor is applied for integrating spatial information when selecting new training samples. As a result, our proposed approach combines the spectral–spatial information in the SSL strategy.

The paper is organized as follows. In Section 2, the mean shift based segmentation is briefly reviewed, and then the proposed approach is introduced. Section 3 demonstrates the experimental results. Finally, Section 4 concludes with some remarks.