欢迎访问林业科学,今天是

林业科学 ›› 2021, Vol. 57 ›› Issue (7): 80-91.doi: 10.11707/j.1001-7488.20210709

• 论文与研究报告 • 上一篇    下一篇

结合光学与雷达影像多特征的热带典型人工林树种精细识别

黄翀1,张晨晨1,2,刘庆生1,李贺1,杨晓梅1,刘高焕1   

  1. 1. 中国科学院地理科学与资源研究所 资源与环境信息系统国家重点实验室 北京 100101
    2. 中国科学院大学 北京 100049
  • 收稿日期:2020-01-19 出版日期:2021-07-25 发布日期:2021-09-02
  • 基金资助:
    国家自然科学基金项目(41890854);国家自然科学基金项目(41901309);国家自然科学基金项目(41801353)

Multi-Feature Classification of Optical and SAR Remote Sensing Images for Typical Tropical Plantation Species

Chong Huang1,Chenchen Zhang1,2,Qingsheng Liu1,He Li1,Xiaomei Yang1,Gaohuan Liu1   

  1. 1. State Key Laboratory of Resources and Environmental Information System Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences Beijing 100101
    2. University of Chinese Academy of Sciences Beijing 100049
  • Received:2020-01-19 Online:2021-07-25 Published:2021-09-02

摘要:

目的: 结合Sentinel-2光谱特征、纹理特征和Sentinel-1雷达后向散射特征,开展光学与雷达影像协同人工林树种分类研究,评价不同数据源和不同特征在树种分类中的作用,获取最优分类策略,为热带典型人工林树种精细识别提供新的技术途径。方法: 获取Sentinel-2光学影像43个光谱特征,将优选的Sentinel-2光谱特征与Sentinel-2纹理特征、Sentinel-1雷达后向散射特征组合,采用随机森林分类算法对研究区橡胶林、油棕林和桉树林进行精细提取,并评价不同特征对人工林树种识别的贡献。结果: 1)选择重要性排名前17的Sentinel-2光谱特征参与分类时,OOB score达最高值0.947 2,其中Sentinel-2的蓝波段、红边和近红外波段及其相应的植被指数在树种识别中重要性较高;2)Sentinel-2纹理特征和Sentinel-1雷达后向散射特征对于不同树种均具有一定差异性,可作为树种分类的有效特征;3)仅利用Sentinel-2光谱特征对橡胶林、油棕林和桉树林的区分度有限,生产者精度和用户精度的调和平均值(F1)分别为0.70、0.74和0.61;结合光谱特征、纹理特征和后向散射特征,3种人工林的分类精度达到最高,F1均大于0.80,同时其他地物的分类精度也有较大提高,总体分类精度为0.85,Kappa系数为0.83。结论: Sentinel-2光学影像的光谱特征是树种识别的重要基础,Sentinel-2光学影像派生的纹理特征及由Sentinel-1雷达影像提取的后向散射特征可为不同人工林树种识别提供有益补充。考虑到Sentinel-2和Sentinel-1较高的时间分辨率和空间分辨率,同时结合其光谱特征、纹理特征和后向散射特征对热带典型人工林树种进行精细识别具有极大潜力。

关键词: 人工林, Sentinel-2, 合成孔径雷达, Sentinel-1, 树种识别, 红边指数, 多特征

Abstract:

Objective: Accurate plantation map plays an important role in sustainable forest resource management. In this paper,a multi-feature classification method based on Sentinel-2 optical and Sentinel-1 SAR data combination was proposed for discriminating and mapping typical plantation species in the tropical region. Method: 43 spectral features were extracted from a scene of Sentinel-2 image. Their importance was sorted through the out-of-bag(OOB) error estimation,to select out the optimal spectral features for plantation species classification. In addition,4 textural features were also calculated from the Sentinel-2 image to represent the spatial information of each plantation species,and 4 backscattering coefficient features were extracted from Sentinel-1 SAR image to represent the structural characteristics. Then,different combinations using the selected spectral features,together with textural and backscattering coefficient features were input to the random forest classifier for plantation species classification. The performance and contribution of different combinations were assessed based on their classification accuracies. Result: 1) The OOB score reaches the highest value of 0.947 2 when the top 17 important Sentinel-2 spectral features are employed. Among them,blue,red-edge,NIR spectral bands,and vegetation indices based on NIR and red-edge bands have a high importance for plantation species identification. 2) Eucalyptus,oil palm and rubber plantations show some differences in different textural features. Among the three plantations,oil palm has the largest CON(contrast) and ENT(entropy),and the smallest ASM(angular second moment) value,while eucalyptus has the largest ASM and COR(correlation),and the smallest ENT value. For rubber,the values of CON and COR are the smallest,while the values of ASM and ENT are between those of eucalyptus and oil palm. The backscattering coefficients of the three plantations also exhibit much separability,especially for the ratio of VV to VH. 3) When only the Sentinel-2 spectral features were used for plantation species classification,the F1 values of eucalyptus,oil palm and rubber plantations were 0.61,0.74 and 0.70,respectively. However,the combination of spectral features,textural features and backscattering coefficient features have been seen the highest classification accuracy,with the overall accuracy and Kappa coefficient of 85% and 0.83,respectively. The F1 values of eucalyptus,oil palm and rubber plantation increase by 0.19,0.15 and 0.10,respectively. Conclusion: The spectral bands of Sentinel-2 optical image could provide an important basis for plantation species recognition,while the derived textural features and the backscattering coefficient features extracted from Sentinel-1 SAR image may also provide supplementary information which contributes a lot to the identification of different plantation species. Considering the high temporal and spatial resolution of Sentinel-2 and Sentinel-1,the integration of spectral,textural and structural information might show great potentials to improve the accuracy of plantation species discriminating at a fine scale.

Key words: plantation, Sentinel-2, synthetic aperture radar(SAR), Sentinel-1, plantation species classification, red-edge index, multi-features

中图分类号: