地球信息科学学报 ›› 2018, Vol. 20 ›› Issue (12): 1777-1786.doi: 10.12082/dqxxkx.2018.180310

• 遥感科学与应用技术 • 上一篇    下一篇

基于随机森林特征选择的城市绿化乔木树种分类

温小乐1,2,3(), 钟奥1,2, 胡秀娟1,2,3,*()   

  1. 1. 福州大学环境与资源学院,福州 350116
    2. 福州大学遥感信息工程研究所,福州 350116
    3. 福建省水土流失遥感监测评估与灾害防治重点实验室,福州 350116
  • 收稿日期:2018-07-03 出版日期:2018-12-25 发布日期:2018-12-20
  • 通讯作者: 胡秀娟 E-mail:wenxiaole@sina.com;huxiujuan@fzu.edu.cn
  • 作者简介:

    作者简介:温小乐(1976-),女,博士,副教授,研究方向为环境资源遥感与环境评价。E-mail: wenxiaole@sina.com

The Classification of Urban Greening Tree Species Based on Feature Selection of Random Forest

WEN Xiaole1,2,3(), ZHONG Ao1,2, HU Xiujuan1,2,3,*()   

  1. 1. College of Environment and Resources,Fuzhou University, Fuzhou 350116, China
    2. Institute of remote Sensing Information Engineering, Fuzhou University, Fuzhou 350116, China
    3. Fujian Provincial Key Laboratory of Remote Sensing of Soil Erosion and Disaster Protection, Fuzhou University, Fuzhou 350116, China
  • Received:2018-07-03 Online:2018-12-25 Published:2018-12-20
  • Contact: HU Xiujuan E-mail:wenxiaole@sina.com;huxiujuan@fzu.edu.cn

摘要:

城市绿化在改善空气、水和土壤质量,吸收和减少二氧化碳及各种污染物,缓解城市热岛和减少雨水径流等方面发挥着重要作用。及时准确地获取树种信息是城市规划与绿化管理的先决条件,对进一步改善城市生态环境也具有重要意义。基于遥感技术,使用高空间分辨率的WorldView-2卫星影像,采用光谱、纹理、指数以及几何等多种特征相结合的面向对象方法,并通过随机森林进行特征选择,对福州大学旗山校区北部的榕树、杧果、香樟、重阳木、羊蹄甲、垂叶榕以及木棉7种主要绿化乔木进行树种分类。实地验证结果表明:通过特征选择可以减少或规避数据冗余以及休斯效应的产生,该方法可以提高现有同类型树种分类的精度,当淘汰全部特征的20%,利用34个特征(包括15个光谱特征、6个纹理特征、8个指数特征和5个几何特征)进行分类时,总精度最高,可达74.95%,Kappa系数为0.67。其中,光谱平均值的特征重要性最高,而各波段的标准差的重要性较低。WorldView-2卫星影像的4个新增波段,特别是黄光和红边波段及其构建的指数特征重要性较高,也说明这些波段在植被遥感,特别是树种分类中极具应用前景。

关键词: WorldView-2, 面向对象, 随机森林, 特征选择, 树种分类

Abstract:

Since Urban forests played important roles in improving air, water and land quality, absorbing and mitigating carbon dioxide and many pollutants, mitigating urban heat island and reducing storm water runoff, its monitoring is a major issue for urban planners. It is of great significance to obtain the tree species timely and precisely in urban planning and green space management. At present, urban forest tree species mapping has benefitted from advances in remote sensing techniques. Using an object-oriented method combing spectral, textural, indicial and geometric features from high-resolution WorldView-2 imagery, this paper aimed to carry out the classification of seven main tree species in Fuzhou university, including Banyan (Ficus microcarpa), Mango(Mangifera indica L.), Camphor tree (Cinnamomum camphora), Bishop wood (Bischofia polycarpa), Chinese orchid tree(Bauhinia purpurea L.), Weeping fig (Ficus benjamina L.), and Kapok tree (Bombax malabaricum DC.). A random forest method was employed to determine the feature selection in this study. When eliminating 20 percent of the total features, the in situ validation results showed that the overall accuracy reached a highest value of 74.95% with Kappa coefficient of 0.67 when using 34 features for classification, which including 15 spectral features, 6 textural features, 8 indicial features and 5 geometric features, and the feature of mean spectral was the most significant, however, the standard deviation of each band is less important. The results also revealed that the feature selection of random forest could reduce or avoid the data redundancy and Hughes phenomenon, and thus could improve the classification accuracy of same type tree species. Moreover, the four additional bands of WorldView-2 imagery, especially the yellow and red edge band, and their composite indexes showed a higher importance in classification, which also indicates that these bands have great application prospects in vegetation remote sensing, especially in tree species classification.

Key words: WorldView-2, object-oriented, random forest, feature selection, tree species classification