地球信息科学学报 ›› 2021, Vol. 23 ›› Issue (6): 1050-1062.doi: 10.12082/dqxxkx.2021.200372

• 遥感科学与应用技术 • 上一篇    下一篇

面向对象高分遥感影像典型自然地物半自动提取

张春森1(), 贾欣1, 吴蓉蓉1, 崔卫红2, 史书1, 郭丙轩3   

  1. 1.西安科技大学测绘科学与技术学院,西安 710054
    2.武汉大学遥感信息工程学院,武汉 430079
    3.武汉大学测绘遥感信息工程国家重点实验室,武汉 430079
  • 收稿日期:2020-07-15 修回日期:2020-10-29 出版日期:2021-06-25 发布日期:2021-08-25
  • 作者简介:张春森(1963— ),男,陕西西安人,博士,教授,主要从事摄影测量与高分遥感应用研究。E-mail: zhchunsen@aliyun.com
  • 基金资助:
    国家自然科学基金重大研究计划项目(92038301);自然资源部城市国土资源监测与仿真重点实验室开放基金资助课题(KF-2018-03-052)

Object Oriented Semi-automatic Extraction of Typical Natural Areas from High-Resolution Remote Sensing Images

ZHANG Chunsen1(), JIA Xin1, WU Rongrong1, CUI Weihong2, SHI Shu1, GUO Bingxuan3   

  1. 1. College of Geomatics, Xi'an University of Science and Technology, Xi'an 710054, China
    2. School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079, China
    3. State Key Laboratory of Information Engineering in Surveying Mapping and Remote Sensing,Wuhan University,Wuhan 430079, China
  • Received:2020-07-15 Revised:2020-10-29 Online:2021-06-25 Published:2021-08-25
  • Supported by:
    Major Research Plan of the National Natural Science Foundation of China(92038301);Laboratory of Urban Land Resources Monitoring and Simulation,Ministry of Natural Resources(KF-2018-03-052)

摘要:

针对目前高空间分辨率遥感影像(简称高分遥感影像)地物全自动提取无法完全实现的现实,本文结合自然地物的光谱和纹理特征,提出一种面向对象的高分遥感影像典型自然地物半自动提取方法。首先构建最小生成树(Minimum Spanning Tree, MST)进行影像初始分割,根据影像灰度平均归一化值和标准差统计对象的光谱、纹理等特征。用户通过“种子点”交互选取提供前景样本,并基于区域邻接图(Region Adjacency Graph, RAG)寻找合并代价最小的区域扩充前景样本。在自动构建的环形缓冲区内选择背景样本,利用特征空间高斯滤波实现全连接条件随机场中均值场更新。依据全连接条件随机场描述全局信息,结合不同地物的提取准则最终得到自然地物的提取结果。以航空和高分二号(GF-2)遥感影像为实验数据,分别对林地、草地、耕地、裸地和水体等典型自然地物进行提取。结果显示,基于本文方法的航空影像典型自然地物提取总精度和Kappa值为0.959和0.948,相较于SVM方法分别提升了20.757%和0.268。高分二号(GF-2)遥感影像的提取总精度和Kappa值为0.959和0.941,相比SVM方法分别提高了1.698%和0.133。证明所给方法能够通过较少的用户交互,实现高分遥感影像典型自然地物高精度智能提取。

关键词: 高分遥感影像, 半自动提取, 面向对象, 全连接条件随机场, 最小生成树, 区域邻接图, 高斯混合模型, 支持向量机

Abstract:

Given that the automatic extraction of features from high resolution remote sensing images (HR remote sensing images) cannot be fully realized, this paper proposes an object oriented semi-automatic method to extract typical natural objects from HR remote sensing images based on spectral and textural features. We introduce the concept of "seed points" that provides limited seed points from the ground objects to be extracted to realize the auto-extraction of large-scale typical natural ground objects under the premise of ensuring the extraction accuracy. We firstly segment the initial images based on the Minimum Spanning Tree algorithm, and then the spectral and textural features of each object are calculated according to the average normalized value and standard deviation of the image gray. Secondly, users provide foreground samples through interactive selection of "seed points", and then find the regional expanded foreground samples with the least merging cost based on Region Adjacency Graph (RAG). Thirdly, the Gaussian Mixture Models (GMMs) can better adapt to the color of the image, effectively capture the subtle differences between the foreground and the background, and approximate any probability distribution with arbitrary precision. They have been widely used in image processing and model building. This paper uses GMMs to estimate whether the current object belongs to the foreground or background models. We then update the mean field in the fully connected conditional random field through the feature space Gaussian filter. The fully connected condition random field can describe the relationship between each node and all other nodes. Finally, the contour extraction results are obtained according to the extraction criteria of different natural objects and the global information description with the fully connected conditional random fields. In this study, we utilized aerial images and GF-2 images to verify our method, and extracted five typical natural objects including woodland, grassland, cropland, bare land, and waters. The results show that the total accuracy and Kappa value of typical natural features extracted from aerial images were 0.959 and 0.948, respectively. They increased by 20.757% and 0.268, respectively, compared with the SVM method. The total accuracy and Kappa value using Gaofen-2 (GF-2) remote sensing images were 0.959 and 0.941, respectively. They increased by 1.698% and 0.133, respectively, compared with SVM. Our results indicate that the proposed method can effectively extract the natural objects in HR remote sensing images with less interaction and time.

Key words: high spatial resolution remote sensing image, semi-automatic extraction, object oriented, fully connected conditional random fields, minimum spanning tree, region adjacency graph, Gaussian Mixture Models, Support Vector Machines