A novel semisupervised SVM for pixel classification of remote sensing imagery
- PDF / 981,020 Bytes
- 12 Pages / 595.276 x 790.866 pts Page_size
- 32 Downloads / 203 Views
ORIGINAL ARTICLE
A novel semisupervised SVM for pixel classification of remote sensing imagery Ujjwal Maulik • Debasis Chakraborty
Received: 12 May 2011 / Accepted: 24 October 2011 / Published online: 15 November 2011 Ó Springer-Verlag 2011
Abstract This article introduces a semisupervised support vector machine classification technique that exploits both labeled and unlabeled points for addressing the problem of pixel classification of remote sensing images. The proposed technique is based on applying the margin maximization principle to both labeled and unlabeled patterns. Semisupervised SVM progressively searches a reliable discriminant hyperplane in the high dimensional space through iterative method exploiting both labeled and unlabeled samples. In particular, the dynamic thresholding and successive filtering of the unlabeled set are exploited to find a reliable separating hyperplane. The proposed technique is first demonstrated for six labeled datasets described in terms of feature vectors and then identifying different land cover regions in remote sensing imagery and compared with the standard SVM. Experimental results confirm that employing this learning scheme removes unnecessary points to a great extent from the unlabeled set and increases the accuracy level on the other hand. Comparison is made in terms of accuracy, ROC, AUC and F-measure for the labeled data and quantitative cluster validity indices as well as classified image quality for the image data.
U. Maulik Department of Computer Science and Engineering, Jadavpur University, Calcutta 700032, West Bengal, India e-mail: [email protected] D. Chakraborty (&) Department of Electronics and Communication, Murshidabad College of Engineering and Technology, Berhampore 742101, West Bengal, India e-mail: [email protected]
Keywords Support vector machines Remote sensing satellite images Quadratic programming Semisupervised classification
1 Introduction Remote-sensing research focusing on image classification has long attracted the attention of the remote sensing community because classification results are the basis for many environmental and socioeconomic applications. A suitable classification system and a sufficient number of training samples are prerequisites for a successful classification [15]. In this context, two major problems need to be addressed: (1) a problem related to the quantity of the available training samples and (2) a problem related to the quality of the available training points [9]. Regarding the quantity of the training samples, acquisition of sufficient number of reference data is often a critical problem to design a successful classifier in real life applications. In particular, small number of training points compared to the number of features [5] gives rise to the well-known problem of the curse of dimensionality [18]. As a result, there is a risk of overfitting the training data and may involve poor generalization capabilities in the classifier. As regards to the quality of the training data, two major issues ar
Data Loading...
