高空间分辨率遥感影像多尺度分割优化组合算法

A Multi-scale-segmentation Optimal Composition Algorithm for High Spatial Resolution Remote Sensing Imagery

高空间分辨率遥感影像能够充分地描述地表覆盖空间异质性,可用于提取地面目标物。然而高空间分辨率在像元尺度的目标提取时易产生"椒盐效应"问题,面向对象的小尺度影像分割也受此效应影响;而大尺度的影像分割造成较小目标的遗漏。本文提出了一种针对高空间分辨率遥感影像的多尺度分割优化组合算法MOCA(Multi-scale-segmentation Optimal Composition Algorithm),基于后验概率信息熵指标选择影像中每个地面目标的最优分割尺度并集成组合,获得高空间分辨率遥感影像的多尺度分割优化组合结果。本文使用F指标和BCI(Bidirectional Consistency Index)两种指标评估地面目标物提取精度,并将MOCA与同类多尺度分割方法进行比较。实验结果表明,本文提出的MOCA算法可实现多个分割尺度的最优组合,并获得较高的地面目标物提取精度。

High spatial resolution remote sensing imagery can fully delineate the heterogeneity of land covers and has been widely used for extracting surface objects. However, pixel-based object extraction from high spatial resolution images may bring in＇salt-and-pepper effect＇, and this effect also occurs when segmentation scale is small using object- based analysis. Choosing a big segmentation scale will omit small surface objects. In this research, a Multi-scale- segmentation Optimal Composition Algoritiim （MOCA） for object-based extraction with high spatial resolution imagery are proposed. MOCA selects optimal segmentation scale for each surface object based on entropy of posterior probabilities of land covers, and combines different scales to produce the suitable scales composition. F-measure and Bidirectional Consistency Index （BCI） are used to evaluate the accuracy of surface object extraction and compared with the existing multi-scale-segmentation method. Results show that MOCA can achieve the optimal composition of multiple scales and obtain a high accuracy of object extraction.