欢迎访问林业科学,今天是

林业科学 ›› 2018, Vol. 54 ›› Issue (2): 68-80.doi: 10.11707/j.1001-7488.20180208

• 论文与研究报告 • 上一篇    下一篇

综合应用多源遥感数据的面向对象土地覆盖分类方法

李晓红, 陈尔学, 李增元, 李世明   

  1. 中国林业科学研究院资源信息研究所 北京 100091
  • 收稿日期:2016-02-22 修回日期:2016-06-28 出版日期:2018-02-25 发布日期:2018-03-30
  • 基金资助:
    高分辨率对地观测系统重大专项课题"高分森林资源调查应用示范子系统(一期)"(21-Y30B05-9001-13/15-1)。

Object Based Land Cover Classification Method Integrating Multi-Source Remote Sensing Data

Li Xiaohong, Chen Erxue, Li Zengyuan, Li Shiming   

  1. Research Institute of Forest Resource Information Techniques, CAF Beijing 100091
  • Received:2016-02-22 Revised:2016-06-28 Online:2018-02-25 Published:2018-03-30

摘要: [目的]针对国家森林资源宏观监测业务对区域森林资源空间分布信息的迫切需求,发展一种基于国家森林资源连续清查固定样地数据,可充分发挥GF-1宽幅多光谱数据、MODIS遥感数据相应空间和时间分辨率优势的面向对象土地覆盖分类方法,以提高林地和森林资源的监测精度和自动化程度。[方法]以黑龙江省小兴安岭某林区为研究区,主要数据源包括GF-1宽幅多光谱数据、MODIS NDVI(250 m,8天合成)时间序列遥感数据、国家森林资源连续清查固定样地数据以及少量外业实地调查数据等。首先,基于GF-1宽幅多光谱数据进行多尺度影像分割,将研究区划分为许多区域性的分割对象;然后,以分割对象为分析单元,分别提取GF-1宽幅多光谱遥感影像的光谱特征、纹理特征、形状特征等以及MODIS NDVI时间序列遥感数据的时序特征,并采用随机森林算法进行特征选择;最后,利用训练样本建立基于分类回归树分类器完成面向对象的土地覆盖分类方法研究,分别比较单一GF-1 16 m宽幅多光谱数据、单一MODIS NDVI时间序列遥感数据以及综合多源数据的分类结果,并基于混淆矩阵对分类结果进行分析。[结果]精度检验和分析结果表明,面向对象的综合多源遥感数据分类方法总体分类精度达89.46%,Kappa系数为0.874,明显优于仅基于GF-1宽幅多光谱数据、MODIS NDVI时间序列遥感数据的分类方法。[结论]综合应用多源遥感数据的面向对象土地覆盖分类方法适用于综合GF-1与GF-4数据的土地覆盖类型分别制图,可有效提高主要土地覆盖类型的分类精度。针对国家森林资源连续清查的业务需求和特点,本文所发展的方法在分类对象生成、特征提取、特征选择、分类器训练和精度检验等关键环节均进行了优化设计,有利于提高森林资源连续清查业务中主要林地类型遥感分类制图的自动化、标准化程度。

关键词: 多源数据, GF-1宽幅多光谱数据, MODIS NDVI遥感数据, 随机森林, 面向对象, 土地覆盖分类

Abstract: [Objective] In order to meet the urgent needs of national forest inventory (NFI) for monitoring national forest resources on a macro-scale, an object-oriented regional land cover classification method was developed in this paper by using permanent forest plot data of NFI and integrating the corresponding spatial and temporal resolution advantages of GF-1 WFV multi-spectral data and MODIS remote sensing data.[Method] The test site is located in the central of the Xiaoxing'an mountain region in Heilongjiang Province. The GF-1 WFV multispectral data, time series MODIS NDVI product of 8 days synthetic (250 m spatial resolution), the permanent forest plot data collected by the NFI and some field survey data are employed as the key data sources. After image segmentation, spectral, texture and shape features from GF-1 WFV multi-spectral data and NDVI features from times series MODIS NDVI data are extracted for each object. Based on these features, the random forests algorithm is adopt to select the best features automatically and then the classification and regression tree is used to administer the supervised classification. The permanent forest plot data are used to build and validate the decision tree classifier. This method has been investigated with the data collected for the study site.[Result] Experimental results show that the overall accuracy and Kappa coefficient of the developed method combing multi-sources data can reach 89.46% and 0.874 respectively, with significant improvement compared with that using either GF-1 WFV multi-spectral data or MODIS NDVI time series data alone.[Conclusion] The land cover classification method developed in this study is appropriate for integrating GF-1 data and GF-4 data for land cover classification mapping and can improve accuracy of land cover classification effectively.Compared with existed research work,this paper has developed one land cover classification method of much more practical application value with some special features such as focusing on the operational application needs and the characteristics of NFI,optimizing the key procedures such as classification objects generation,feature extraction,feature selection,classifier training,accuracy validation,and so on.The land cover classification method developed is useful for improving the automation and standardization of the technique flow to produce forest land cover maps implemented by NFI.

Key words: multi-sources data, GF-1 WFV multi-spectral data, MODIS NDVI remote sensing data, random forest, objected-oriented, land cover classification

中图分类号: