Imitate geometric manifold coverage method for one-class classification of remote sensing data
- PDF / 850,790 Bytes
- 8 Pages / 595.276 x 790.866 pts Page_size
- 92 Downloads / 195 Views
ORIGINAL PAPER
Imitate geometric manifold coverage method for one-class classification of remote sensing data Yiliang Zeng & Jinhui Lan
Received: 28 August 2013 / Accepted: 3 January 2014 # Saudi Society for Geosciences 2014
Abstract One-class classification (OCC) of remote sensing image only pays attention to the class of interest, without regarding to other classes. Traditional classifiers are inefficient for OCC because it requires all the classes in an image labelled. In this paper, a simple and reliable Imitate Geometric Manifold Coverage (IGMC) method is proposed for solving OCC problems in remote sensing image. First of all, the IGMC method establishes the initial geometric covering space by spectral angle, and the coverage condition using geometry space relationship is presented to determine whether the current testing region is covered in the current geometry. Then, in order to use unlabeled data to help build classifier, the constraint condition using relative correlation dimensions and “Shift and Shrinking” is proposed. Finally, the regions of specific class are labelled as positive label in the whole image. Experiments are conducted using QuickBird multispectral image and MERIS subset. And the results of the proposed framework are compared to Support Vector Data Description, one-class Support Vector Machine (OC-SVM) and objectbased SVM methods. The advantages of the new method are that it requires only a small set of training data and provides more accurate and compact classification results.
Keywords One-class classification . IGMC . Relative correlation dimensions . “Shift and Shrinking”
Y. Zeng : J. Lan (*) Department of Instrument Science and Technology, School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing 100083, People’s Republic of China e-mail: [email protected] Y. Zeng e-mail: [email protected]
Introduction With the development of remote sensing data, multispectral remote sensing images have shown their usefulness in numerous applications. Multispectral images provide huge amounts of data for terrain analysis and help to provide more detailed information for remote sensing object classification (Tuia et al. 2009). Traditionally, all the classes in an image should be completely labelled for classification (Li and Guo 2011). In most classification problems, training data is available for all classes of instances. In this case, the learning algorithm can use the training data for the different classes to determine decision boundaries that discriminate between these classes (Hempstalk et al. 2008). However, it is difficult and timeconsuming to select the set of training samples from the complex ground surface, which causes the classification accuracy decreasing (Tuia and Munoz-Mari 2012). For some applications, we may only refer to the specific land-cover class of interest or only the target class which is known (Foody et al. 2006). For example, if the goal for existing disaster monitoring system is to retrieve mountain landslide f
Data Loading...
