对象置信度指引下的高分辨率遥感影像分割

Object confidence index guided high-resolution remote sensing image segmentation

如何减小分割结果与实际地理对象间的差异,是目前高分辨遥感影像分割中面临的一个难点问题。为此,构建了一种新的对象置信度(OC)指标来衡量任意区域与地理对象间的匹配程度,进而提出了一种面向地理对象的多尺度分割算法。该算法主要包括两个步骤:首先,通过对影像进行过分割来构建初始种子区域集合,并确定尺度参数集合;而后,通过跟踪对象置信度指标OC的尺度间变化来指引多尺度区域合并过程,使区域合并结果逐步逼近实际的地理对象。多组实验结果表明,所提出的算法能够显著改善过分割及欠分割问题,准确识别建筑物、道路等地理对象的完整轮廓,在定性分析及定量精度评价中均显著优于商业软件e Congnition及传统多尺度分割算法。

How to reduce the difference between segmentation result and practical geographical object is a difficult problem faced by high- resolution remote sensing image segmentation currently. Aiming at this issue, in this paper a new OC （Object Confidence） index is constructed to measure the matching degree between any region and geographical object, and a multi-scale segmentation algorithm facing to geographical objects is proposed. This algorithm mainly contains two steps： firstly, this algorithm establishes an initial seed regional set through conducting over segmentation to the image and determines the scale parameter set; secondly, this algorithm guides the process of multi-scale region merging through tracking the inter-scale change of OC index, and makes the region merging result gradually approach to practical geographical object. The multi-group experiments indicate that the proposed algorithm can obviously improve the over-segmentation and insufficient-segmentation problems, and identify the complete outlines of buildings, roads as well as other geographical objects accurately. The proposed algorithm is obviously superior to the commercial software eCongnition and traditional muhi-scale segmentation algorithm in both qualitative analysis and quantitative precision evaluation.