欢迎访问《应用生态学报》官方网站,今天是 分享到:

应用生态学报 ›› 2018, Vol. 29 ›› Issue (12): 3995-4003.doi: 10.13287/j.1001-9332.201812.015

• 研究论文 • 上一篇    下一篇

基于不同决策树的面向对象林区遥感影像分类比较

陈丽萍,孙玉军*   

  1. 北京林业大学森林资源和环境管理国家林业局重点开放性实验室, 北京 100083
  • 收稿日期:2018-06-04 修回日期:2018-10-01 出版日期:2018-12-20 发布日期:2018-12-20
  • 作者简介:陈丽萍, 女, 博士研究生, 1990年生. 主要从事林业遥感与信息技术. E-mail: chenlp_1990@foxmail.com
  • 基金资助:
    本文由国家自然科学基金项目(31870620)和国家林业局‘948’项目(2015-4-31)资助

Comparison of object-oriented remote sensing image classification based on different decision trees in forest area

CHEN Li-ping, SUN Yu-jun*   

  1. State Forestry Administration Key Laboratory of Forest Resources & Environmental Management, Beijing Forestry University, Beijing 100083, China
  • Received:2018-06-04 Revised:2018-10-01 Online:2018-12-20 Published:2018-12-20
  • Supported by:
    This work was supported by the National Natural Science Foundation of China (31870620) and the 948 Project of State Forestry Administration of China (2015-4-31).

摘要: 面向地理对象影像分析技术(GEOBIA)是影像分辨率越来越高的背景下的产物.如何提高高分辨率影像分类精度和分类效率是影像处理的重要议题之一.本研究对QuickBird影像多尺度分割后的对象进行分类,分析了C5.0、C4.5、CART决策树算法在林区面向对象分类中的效率,并与kNN算法的分类精度进行比较.利用eCognition软件对遥感影像进行多尺度分割,分析得到最佳尺度为90和40.在90尺度下分离出植被和非植被后,在40尺度下提取不同类别植被的光谱、纹理、形状等共21个特征,并利用C5.0、C4.5、CART决策树算法分别对其进行知识挖掘,自动建立分类规则.最后利用建立的分类规则分别对植被区域进行分类,并比较分析其精度.结果表明: 基于决策树的分类精度均高于传统的kNN法.其中,C5.0方法的精度最高,其总体分类精度为90.0%,Kappa系数0.87.决策树算法能有效提高林区树种分类精度,且C5.0决策树的Boosting算法对该分类效果具有最明显的提升.

Abstract: Geographic Object-Based Image Analysis (GEOBIA) was a product under the background of increasing high-resolution remote sensing data. How to improve the accuracy and efficiency of classification of high-resolution images is one of the important topics in image processing. After objects segmented multiscale by QuickBird image was classified, the efficiency of C5.0, C4.5, and CART decision trees in object-oriented classification of forest areas was analyzed. The accuracy of those three methods were compared with kNN method. The eCognition software was used to multiscale segmentation of remote sensing images, with the result showing that 90 and 40 were the optimal scales. After separating vegetation and non-vegetation at 90 scale, 21 features such as spectrum, texture and shape of different vegetation types were extracted at 40 scale, knowledge mining was carried out by using C5.0, C4.5 and CART decision tree algorithms respectively, and classification rules were automatically established. The vegetation area was classified based on the classification rules and the classification accuracy of different methods was compared. The results showed that the classification accuracy based on decision-tree was higher than that of the traditional kNN method. The accuracy of C5.0 method was the best, with the overall accuracy and Kappa coefficient reaching 90.0% and 0.87, respectively. The decision tree algorithm could effectively improve the accuracy in classification of forest species. The Boosting algorithm of the C5.0 decision tree had the most significant improvement on the classification.