• 遥感大数据协同计算方法 •

### 综合多层优选尺度的高分辨率影像分割

1. 1. 浙江工业大学计算机学院,杭州 310023
2. 浙江省海洋大数据挖掘与应用重点实验室,舟山 316022
3. 中国地质大学（北京）信息工程学院,北京 100083
• 收稿日期:2016-02-15 修回日期:2016-04-18 出版日期:2016-05-10 发布日期:2016-05-10
• 作者简介:

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

• 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

YANG Haiping1,2,*(), MING Dongping3

1. 1. College of Computer Science & Technology, Zhejiang University of Technology, Hangzhou 310023, China
2. Key Laboratory of Oceanographic Big Data Mining & Application of Zhejiang Province, Zhoushan 316022, China
3. School of Information Engineering, China University of Geosciences (Beijing), Beijing 100083, China;
• Received:2016-02-15 Revised:2016-04-18 Online:2016-05-10 Published:2016-05-10
• Contact: YANG Haiping E-mail:yanghp@zjut.edu.cn

Abstract:

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.