关键词: 高分辨率遥感影像, 面向对象, 几何结构, 分割/分类评价

Abstract:

Traditional classification accuracy assessments in terms of overall accuracy or Kappa coefficient based on isolating pixels statistic, cannot capture the geometrical properties of the segmented image objects, and therefore do not provide an accurate evaluation in object based segmentation or classification. Considering the importance of bordering pixels and geometric shape in object segmentation, it is thus of our interesting to design some measuring indexes introduced the border pixels and geometric structure information to evaluate the object-based resultant image segmentation or classification. This paper improved five indexes originally proposed by Persello and Bruzzone to evaluate the resultant segmentation of high spatial resolution image. Realizing that the original indices cannot well capture the gap among different categories, resulting in difficulties in discriminating the assessment results of different categories, we use normalized techniques based on geometric shape to overcome it. The indices depend on the geometry features of each object of the thematic map including over segmentation, under segmentation, edge location, fragmentation error and shape error. Moreover, we realized a prototype system contained the aforementioned evaluated indexes based on IDL platform to support the object-based image processing and analysis. To validate these indexes, a subset QuickBird image located on Fuzhou was implemented and the results of the Meanshift segmented algorithm demonstrate that the proposed indexes can provide better accuracy evaluation of each land cover class, and can make users more effectively choose the best classification map. Moreover, our experiments also demonstrate that, compared with OA and Kappa coefficient, the proposed indexes have advantages on characterizing the detailed ground materials and are helpful in aiding the optimal parameters selection for the Meanshift segmented algorithm.

Key words: segmentation/classification evaluation, object-based, high-resolution remote sensing image, geometrical structure