资源科学 ›› 2019, Vol. 41 ›› Issue (5): 992-1001.doi: 10.18402/resci.2019.05.15

• 资源生态 • 上一篇    下一篇

基于Sentinel-2A影像特征优选的随机森林土地覆盖分类

何云1,2(), 黄翀1(), 李贺1, 刘庆生1, 刘高焕1, 周振超3, 张晨晨1,2   

  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
    3. 吉林大学 地球探测科学与技术学院,长春 130026
  • 收稿日期:2018-08-16 修回日期:2019-02-26 出版日期:2019-05-25 发布日期:2019-05-25
  • 作者简介:

    作者简介:何云,女,湖南益阳人,硕士生,主要从事生态遥感研究。E-mail: cug_heyun@163.com

  • 基金资助:
    国家自然科学基金国际(地区)合作与交流项目(41661144030);国家自然科学基金项目(41471335)

Land-cover classification of random forest based on Sentinel-2A image feature optimization

Yun HE1,2(), Chong HUANG1(), He LI1, Qingsheng LIU1, Gaohuan LIU1, Zhenchao ZHOU3, Chenchen ZHANG1,2   

  1. 1.State Key Laboratory of Resources and Environmental Information System, Institute of Geographic Sciences and Natural Resources Research, CAS , Beijing 100101, China
    2. University of Chinese Academy of Sciences, Beijing 100049, China
    3. College of Geo-Exploration Science and Technology, Jilin University, Changchun 130026, China
  • Received:2018-08-16 Revised:2019-02-26 Online:2019-05-25 Published:2019-05-25

摘要:

中南半岛地处热带、亚热带地区,由于水热条件适宜,植被生长旺盛,土地利用强度高,地表覆盖类型的光谱特征时空变异复杂,使用传统的基于光谱特征的遥感分类精度难以保证。Sentinel-2A卫星遥感数据具有较丰富的光谱波段和较高的空间分辨率,为土地覆盖遥感分类提供了多维特征空间。但多维特征参与分类容易造成信息冗余,从而导致分类速度和精度降低。因此,如何充分利用Sentinel-2A数据丰富的光谱和空间信息,并通过高维特征空间降维进行特征优选对于提高分类精度具有重要意义。本文以中南半岛典型地区土地覆盖分类为例,利用Sentinel-2A多波段光谱特征,归一化植被指数(NDVI)、比值植被指数(RVI)、差值植被指数(DVI)、归一化水体指数(NDWI)等指数特征以及对比度、相关性、能量、均值、熵等纹理特征,在随机森林模型框架下,采用平均不纯度减少方法对不同特征在土地覆盖分类中的重要程度进行识别;利用袋外(OOB)误差方法,对重要特征组合进行了优选;利用优选特征进行随机森林土地覆盖分类,并与原始随机森林分类结果进行对比。结果表明:Sentinel-2A影像的光谱特征和纹理特征在土地覆盖分类中具有较为重要的作用,光谱特征中短波红外、可见光、植被红边波段重要性较大,纹理特征中均值、能量法重要性较高。选择重要性列前9位的特征参与分类时,OOB精度达到最高;继续增加特征会使模型复杂度过高,容易发生过拟合而使得分类精度不增反降。通过特征优选高效利用了Sentinel-2A丰富的光谱和纹理信息,其总体分类精度达87.53%,Kappa系数达0.8461,优于原始随机森林方法,一定程度上提高了热带亚热带地区复杂土地覆盖分类精度。

关键词: Sentinel-2A, 特征优选, 随机森林, 土地覆盖分类, 袋外(OOB)误差方法, 中南半岛, 泰国穆河流域

Abstract:

Due to the suitable hydrothermal conditions, vigorous vegetation growth, high land use intensity and complex spatiotemporal variation of spectral characteristics of surface cover types, it is difficult to guarantee the accuracy of remote sensing classification using traditional spectral characteristics in tropical and subtropical regions. Multi-spectral, high spatial resolution Sentinel-2A imageries provide a new source of data for land-cover classification. In order to improve the speed and accuracy of land-cover classification using Sentinel-2A images, we propose a classification method with feature-optimized random forests. In this study, we took the Mun River Basin of Indo-China Peninsula as our research area and made full use of the rich spectral characteristics, normalized vegetation index (NDVI), ratio vegetation index (RVI), difference vegetation index (DVI), normalized water body index (NDWI), and texture features including contrast, correlation, energy, mean, and entropy, of Sentinel-2A images for the analyses. We used the average impurity reduction method in random forests to evaluate the importance of different spectral features, indices, and texture features. Combining the out-of-bag (OOB) error to select features, the results of land-cover classification with feature-optimized random forests were obtained. They show that the spectral features and texture features of Sentinel-2A images play an important role in our classification compared with the original random forest land-cover classification results. The short-wave infrared, visible, and vegetation red-edge bands are of greater importance in spectral features, and the mean and energy are of high importance in texture features.The accuracy of OOB is the highest when the top 9 important features are selected. Sentinel-2A images have good adaptability in tropical and subtropical region land-cover classification. It can effectively improve the accuracy of land-cover classification in tropical and subtropical regions. The accuracy of our classification method reaches 87.53%, and the Kappa coefficient reaches 0.8461, better than the original random forest method. The random forest method based on feature optimization not only has a fast classification speed, but also can guarantee high classification accuracy under the condition that the sample is representative, especially suitable for the land-cover classification of medium and high spatial resolution images of Sentinel-2A.

Key words: Sentinel-2A, feature optimization, random forest, land-cover classification, out-of-bag (OOB) method, Indo-China Peninsula, Mun River Basin of Thailand