12 April 2016 Adaptive semisupervised feature selection without graph construction for very-high-resolution remote sensing images
Author Affiliations +
Abstract
Semisupervised feature selection methods can improve classification performance and enhance model comprehensibility with few labeled objects. However, most of the existing methods require graph construction beforehand, and the resulting heavy computational cost may bring about the failure to accurately capture the local geometry of data. To overcome the problem, adaptive semisupervised feature selection (ASFS) is proposed. In ASFS, the goodness of each feature is measured by linear objective functions based on loss functions and probability distribution matrices. By alternatively optimizing model parameters and automatically adjusting the probabilities of boundary objects, ASFS can measure the genuine characteristics of the data and then rank and select features. The experimental results attest to the effectiveness and practicality of the method in comparison with the latest and state-of-the-art methods on a Worldview II image and a Quickbird II image.
© 2016 Society of Photo-Optical Instrumentation Engineers (SPIE)
Xi Chen, Xi Chen, Jinzi Qi, Jinzi Qi, Yushi Chen, Yushi Chen, Lizhong Hua, Lizhong Hua, Guofan Shao, Guofan Shao, } "Adaptive semisupervised feature selection without graph construction for very-high-resolution remote sensing images," Journal of Applied Remote Sensing 10(2), 025002 (12 April 2016). https://doi.org/10.1117/1.JRS.10.025002 . Submission:
JOURNAL ARTICLE
12 PAGES


SHARE
Back to Top