REPUBLIC OF SERBIA

MINISTRY OF DEFENCE

www.mod.gov.rs

MINISTRY OF DEFENCE

Material Resources Sector

Defensive Technologies Department

Military Technical Institute

www.vti.mod.gov.rs

object-BASED remote sensing (RS) images classification

yumin tan

Beihang University, Beijing, China, tanym@buaa.edu.cn

jianzhu huai

Beihang University, Beijing, China,

Abstract: A framework aimed at classifying remote sensing (RS) images based on object attributes is proposed in this paper. The framework consists of two modules, the first module segmenting images into objects, the second classifying those objects considering their geometrical features. Our previous achievement, two-stage segmentation method composed of initial graph-based partition and subsequent hierarchical clustering is used in the first module to produce image objects. In the second module, these objects are classified by the supervised maximum likelihood classifier given both geometrical and spectral features of objects. Experimental results were compared with conventional pixel-based maximum likelihood classifier and those provided in commercial software. It was found randomly chosen features might not benefit classification and this framework achieved comparable accuracy to pioneering methods and had a potential to improve speed.

Key words: object-based classification, hierarchical clustering, maximum likelihood.