Scale-Span Classification of Multispectral Images Based on Feature Construction and Decision Trees

Ning Shu 1, 2+, Liqun Lin 1, Yan Gong 1, Jun Xiao 2 and Fangfang Jin 3 
1 The School of Remote Sensing Information Engineering, Wuhan University, Wuhan, China
2 National Lab for Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan, China
3 North Star Power Science and Technology CO., LTD, Hangzhou, China

Abstract. Most of the existing classification methods, based on the homogeneous-region, involve the best segmentation criterion choice. Using the so-called best scale to classify the multi-scale objects defined by human subjectivity, the paper doesn’t think it is the best way to correspond with the demand of the scale of human being. So the paper proposes a new scale-span classification method, based on multi-scale homogeneous-region model. The method uses the feature construction to fulfill the construction of scale-span features, and the best scale choice is implicit in the new constructive features, rather than directly carrying on the best scale choice. The experimental result proves the new constructive scale-span features can reduce the dimension of the feature space, and can fully use the longitude information of different scales, thus improve the classification accuracy.

Keywords: feature construction, scale-span, genetic programming, multispectral images, classification, homogeneous-region

In: Wan, Y. et al. (eds) Proceeding of the 8th international symposium on spatial accuracy assessment in natural resources and environmental sciences, World Academic Union (Press).

NingShu2008accuracy.pdf393.25 KB