Abstract

Improving Classification Accuracy by Enriching Feature space in High Resolution Multi-spectral Images

Hamed Ashoori
Msc. Student of Remote Sensing
K.N. Toosi University of Technology, Iran
h_ashouri@sina.kntu.ac.ir

M. Javad Valadan Zoej
Assistant Professor
K.N. Toosi University of Technology, Iran
valadanzouj@kntu.ac.ir

Abbas Alimohamadi
Assistant Professor
K.N. Toosi University of Technology, Iran
alimoh_abb@yahoo.com

Barat Mojaradi
Ph.D. Student
K.N. Toosi University of Technology, Iran
mojaradi@albors.kntu.ac.ir


Abstract :
Although conventional remotely sensed image classification methods which use only spectral information can provide accurate results when implement on law and medium resolution images they could not achieve so accurate results especially for high resolution images. Increasing spatial resolution conduces to decrease spectral resolution, whereas spatial information such as image context will be increased. As a result considering contextual information obtaining from these images, could improve the classification accuracy. Using texture quantization could help producing valuable features to enrich feature space. In this paper several statistical features such as mean, variance and entropy as well as other features originated from run-length matrix are generated. Autocorrelation and geostatistics methods are also employed to produce new features. Adding best combination of these features in classification process improves the accuracy of results.