综合多层优选尺度的高分辨率影像分割
作者简介:杨海平(1987-),女,博士,研究方向为高分辨率遥感影像信息提取。E-mail:yanghp@zjut.edu.cn
收稿日期: 2016-02-15
要求修回日期: 2016-04-18
网络出版日期: 2016-05-10
基金资助
国家自然科学基金项目(41271367、41501453、41371347)
浙江省海洋大数据挖掘与应用重点实验室开放课题项目(OBDMA201512)
国家高分辨率对地观测系统重大专项(03-Y30B06-9001-13/15-01)
Optimal Scales Based Segmentation of High Spatial Resolution Remote Sensing Data
Received date: 2016-02-15
Request revised date: 2016-04-18
Online published: 2016-05-10
Copyright
采用面向对象方法处理高空间分辨率遥感影像时,影像分割质量对后续影像的信息提取结果影响很大。本文主要针对高分辨率影像分割中地物多尺度的问题,提出了一种基于多层优选尺度的高分辨率影像分割算法。该算法首先采用一系列规律变化的尺度对高分辨率影像进行多尺度分割,然后通过单分割层全局标准差的变化与尺度的关系确定一组最优分割尺度。在此基础上,通过各优选分割层之间的包含关系,局部建立多层次对象树,从整体上形成影像森林;通过局部同质性异质性综合评价指数的比较及父层光谱特征的限制来选取多层次对象树中的优势对象,从而获得最终的高分辨率影像分割结果。最后,本文分别采用了Geoeye和ZY3多光谱影像进行了2组分割实验,结果表明本文算法能有效地提高正常分割影像对象的比例。
杨海平 , 明冬萍 . 综合多层优选尺度的高分辨率影像分割[J]. 地球信息科学学报, 2016 , 18(5) : 632 -638 . DOI: 10.3724/SP.J.1047.2016.00632
The quality of image segmentation has a great impact on the results of information extraction from high spatial resolution remote sensing imagery when the object-based method is employed. During the segmentation of high spatial resolution remote sensing images, the scale parameter directly affects the construction of segmented image objects. A small scale is likely to produce broken image objects, while a large scale probably results in the mixed image objects. To solve this problem, an image segmentation framework based on a set of optimal scales is proposed in this paper. First of all, the high spatial resolution remote sensing image is processed using multi-scale segmentation methods with respect to a group of regularly distributed scales. Then the relationship between the global standard deviation of a single segmented layer and its corresponding scale is determined, from which a group of optimal scales are selected. Since the object in a layer that is segmented by a big scale parameter contains the corresponding object in a layer that is segmented by a small scale parameter, a hierarchical tree with nodes of multi-scale image objects can be created. Within this hierarchical tree, the image object of the layer that is segmented by the maximum scale is set as the root. In this manner, each image object of the layer that is segmented by the maximum scale can generate a hierarchical tree, which all together forms the image forest. Two types of features are considered when the optimal image object is selected from each hierarchical tree, which are the comprehensive evaluation index and the spectral features. The comprehensive evaluation index keeps a balance between the homogeneity and heterogeneity of the image objects. And the spectral features of the children nodes should be consistent with the parent nodes in order to dismiss the mixed image objects. Finally, the segmented result is created after the optimal image objects from all hierarchical trees are selected. In the experiment presented in this paper, the Geoeye and ZY3 images are adopted. Results show that the proposed method can effectively improve the percentage of properly segmented image objects.
Fig.1 Process of image segmentation based on a series of optimal scales图1 基于多层优选尺度的高分影像分割方法 |
Fig.2 Diagram of a hierarchical tree with nodes of multiscale image objects图2 多层次对象树示意图 |
Fig.3 Geoeye multispectral image图3 Geoeye多光谱影像 |
Fig.4 Comparison of the segmentation results using different methods for the Geoeye multispectral image图4 Geoeye多光谱影像分割实验中不同尺度分割结果和本文方法分割结果的精度比较 |
Fig.5 Details of the segmentation results for the Geoeye multispectral image图5 Geoeye多光谱影像分割实验的局部结果图 |
Fig.6 ZY3 multispectral image图6 ZY3影像 |
Fig.7 Comparison of the segmentation results using different methods for the ZY3 multispectral image图7 ZY3多光谱影像分割实验中不同尺度分割结果和本文方法分割结果的精度比较 |
Fig.8 Details of the segmentation results of factory buildings for the ZY3 multispectral image图8 ZY3多光谱影像分割实验的局部结果图 |
The authors have declared that no competing interests exist.
[1] |
|
[2] |
|
[3] |
|
[4] |
|
[5] |
|
[6] |
|
[7] |
|
[8] |
|
[9] |
[
|
[10] |
[
|
[11] |
|
[12] |
|
[13] |
|
[14] |
[
|
[15] |
eCognition Developer. eCognition developer 8.7: reference book[M]. Munich, Germany: Trimble Germany GmbH, 2011.
|
[16] |
[
|
[17] |
|
[18] |
[
|
[19] |
[
|
[20] |
[
|
[21] |
|
/
〈 |
|
〉 |