热带地理 ›› 2019, Vol. 39 ›› Issue (4): 512-520.doi: 10.13284/j.cnki.rddl.003147

• 专刊:无人机在生态学和地理学中的应用 • 上一篇    下一篇

基于低空无人机的草原灌丛遥感辨识方法

张 顺1,2,赵玉金2,白永飞2,杨 龙3,孙中宇3   

  1. (1. 山东科技大学,山东 青岛 266590;2. 中国科学院植物研究所,北京 100093;3. 广州地理研究所,广州 510070)
  • 出版日期:2019-07-10 发布日期:2019-07-10
  • 通讯作者: 赵玉金(1988—),男,山东莱芜人,助理研究员,研究方向为生态遥感,(E-mail)zhaoyj@ibcas.ac.cn。
  • 作者简介:张顺(1994—),男,山东青岛人,硕士研究生,研究方向为生态遥感,(E-mail)1403092529@qq.com;
  • 基金资助:

    中国科学院STS项目(KFJ-STS-ZDTP-036);国家重点研发计划:京津冀风沙源区沙化土地治理关键技术研究与示范(2016YFC0500804);基于生理性状的草地物种多样性遥感监测方法研究(41801230)

Remote Sensing Identification of Grassland Shrubs Using Low-Altitude Unmanned Aerial Vehicles

Zhang Shun1,2, Zhao Yujin2, Bai Yongfei2, Yang Long 3 and Sun Zhongyu3   

  1. (1. Shandong University of Science and Technology, Qingdao 266590, China; 2. Institute of Botany, the Chinese Academy of Sciences, Beijing 100093, China; 3. Guangzhou Institute of Geography, Guangzhou 510070, China)
  • Online:2019-07-10 Published:2019-07-10

摘要:

以位于中国科学院内蒙古草原生态研究定位站灌丛化样地实验平台为研究区,基于低空无人机遥感影像,结合实地调查,开展草原灌丛遥感辨识方法研究。通过对灌丛、草地和裸地归一化植被指数(NDVI)的方差统计分析,确定了裸地与植被的分割阈值为-0.08,并使用该阈值提取植被覆盖区,然后分别利用面向对象的决策树(DT)、贝叶斯(Bayes)、K最邻近(KNN)、支持向量机(SVM)机器学习分类器进行灌丛辨识。研究表明:借助Estimation of Scale Parameter(ESP)最优分割尺度评价工具可以快速确定分割参数,获取灌丛、草地影像对象;利用特征空间优化工具选取了18个的对象特征,可以有效避免盲目选择而导致的计算量增大;通过对不同分类器分类结果的对比和样本数量敏感性实验得出:Bayes分类器精度稳定、无需设置参数,灌丛分类精度最高,总体精度和Kappa系数分别达到92%和0.83,结果与影像地物嵌合最好,能够精确识别单株灌丛;根据Bayes分类器分类结果统计得研究区灌丛盖度为14.74%,平均冠幅为0.6 m2,与样方调查结果基本一致。由于4种分类器的算法特征以及对训练样本数量的敏感性各不相同,因此选择合适的分类器还需根据具体影像的地物特征、空间分辨率和研究区范围来确定。

关键词: 灌丛化, 低空无人机, 遥感辨识, 面向对象, 机器学习

Abstract:

Satellite remote sensing imagery, when used in shrub monitoring, cannot accurately identify individual plants or young shrubs due to limitations in spatial resolution. However, using low-altitude unmanned aerial vehicles (UAVs) for remote sensing has the potential to identify such plants while yielding ultrahigh-resolution images. Therefore, we carried out shrub identification based on remote sensing with a low-altitude UAV, combined with a field investigation, at a shrub encroachment in Xilinhot, Inner Mongolia. An analysis of variance of the normalized differential vegetation index for shrubland, grassland, and bare land determined that the segmentation threshold of the bare and vegetated lands was -0.08. The vegetation covered area was extracted using this threshold and four object-oriented machine learning classification classifiers, i.e., Decision Tree (DT), Bayes, k-nearest neighbor (KNN), and Support Vector Machine (SVM), were used to identify the shrubs. Results showed that the tool used to evaluate the optimal segmentation scale, i.e., estimate the scale parameter, could rapidly determine the segmentation parameters, obtain shrub and grass images, and provide the basis for studying the former. We effectively avoided blind selection and increased the number of calculations by using the feature selection and optimization tool to select 18 object features with the highest degree of discrimination. A comparison of results from the different classifiers revealed that the Bayes classifier had the highest classification accuracy, with an overall accuracy and Kappa coefficient of 92% and 0.83, respectively. Its classification results matched the image features well and identified individual shrubs precisely. According to Bayes classification, shrubs covered 14.74% of the study area, and the average crown was 0.6 m2. The results of the remote sensing experiments were almost the same as those from ground-based measurements in the field. A sensitivity experiment on the sample size showed that: 1) SVM obtained a high classification accuracy with relatively few counts in the training sample but was sensitive to parameter settings, which may lead to abnormal classification results. 2) the Bayes classifier had a high precision with stable number of samples, if sufficient training samples were provided. 3) the overall classification accuracy of the DT classifier varied with the number of samples, and 4) the overall accuracy of the KNN classifier was the lowest among the four classifiers. Due to their different algorithmic features and different sensitivities to the number of training samples, the selection of a suitable classifier depends on the surface object being observed and spatial resolution of the image, along with the scope of the study area.

Key words: shrub encroachment, low-altitude UAV, remote sensing identification, object-oriented, machine learning