A novel hyperspectral image classification approach based on multiresolution segmentation with a few labeled samples. (24th June 2017)

• Record Type: Journal Article
• Authors: Cui, Binge; Ma, Xiudan; Zhao, Faxi; Wu, Yanan
• Is Part Of: International journal of advanced robotic systems. Volume 14:Number 3(2017:May/Jun.)
• Publication Date: 2017-06-24
• Subjects: Principal component analysis -- segmentation -- edge-preserving filtering -- semi-supervised image classification -- support vector machine; Robotics -- Periodicals; Robotics; Periodicals; 629.892
• DOI: 10.1177/1729881417710219
• Languages: English
• ISSNs: 1729-8806

Abstract

Hyperspectral remote sensing technology becomes more and more popular in recent years which can be applied to satellite, plane, and flying robots. An important application of hyperspectral remote sensing is the classification of ground objects. However, when the number of labeled samples is very small, the classification accuracy of pixelwise classifiers will decline dramatically. In this article, a novel hyperspectral image classification approach is proposed based on multiresolution segmentation with a few labeled samples. The proposed method is motivated by the fact that pixels within a homogenous region are very likely to have the same class label, which can be utilized to increase the number of labeled samples. The proposed method consists of four steps. First, the hyperspectral image was segmented using the multiresolution image segmentation method. Second, the unlabeled neighbor pixels in the same region as the labeled pixels were selected randomly to assign the class labels. Next, one pixelwise classifier, that is, support vector machine, is used to classify the hyperspectral image with the new labeled sample set. Finally, edge-preserving filtering is performed on the classification result to remove the salt-and-pepper noise and preserve edges of ground objects. Experimental results on three real hyperspectral images demonstrate that the proposed method can improve the classification accuracy significantly when the number of labeled samples is relatively small.