欢迎访问林业科学,今天是

林业科学 ›› 2019, Vol. 55 ›› Issue (2): 87-96.doi: 10.11707/j.1001-7488.20190209

• 论文与研究报告 • 上一篇    下一篇

基于航空正射影像的面向对象林隙识别

毛学刚, 邢秀丽, 李佳蕊, 谭良全, 范文义   

  1. 东北林业大学林学院 哈尔滨 150040
  • 收稿日期:2016-12-19 修回日期:2017-02-05 出版日期:2019-02-25 发布日期:2019-03-20
  • 基金资助:
    国家重点研发计划(2017YFD0600902);中央高校基本科研业务费专项资金项目(2572018BA02);国家级大学生创新训练项目(201810225116)。

Object-Oriented Recognition of Forest Gap Based on Aerial Orthophoto

Mao Xuegang, Xing Xiuli, Li Jiarui, Tan Liangquan, Fan Wenyi   

  1. School of Forestry, Northeast Forestry University Harbin 150040
  • Received:2016-12-19 Revised:2017-02-05 Online:2019-02-25 Published:2019-03-20

摘要: [目的]研究分割尺度及对象特征对航空正射影像面向对象林隙识别的影响,评价基于航空正射影像林隙识别适宜性。[方法]以真彩色航空正射影像为基础数据,采用面向对象分类方法,以东北典型天然次生林帽儿山实验林场东林施业区为试验区进行林隙识别。在面向对象分类过程中,对航空正射影像的3个分量(Blue、Green和Red)采用10种尺度(10~100,步长为10)进行分割,应用拓扑组合准确度(RA(or)和RA(os))和位置准确度(Dsr)评价分割结果。对每种尺度分割结果,应用航空正射影像的光谱及光谱+纹理结合特征,采用带有线性核的支持向量机(SVM)分类器进行林隙、非林隙和树冠分类,共获得20种分类结果。利用混淆矩阵计算的生产者精度、用户精度、分类总精度和Kappa系数4个评价指标对分类结果进行精度评价。[结果]对象特征(大小和形状)受尺度参数影响,小尺度分割产生小面积对象,大尺度分割产生大面积对象,但大尺度不能有效将明显的林隙从冠层中分离出来,存在明显分割不足的现象。拓扑组合准确度(RA(or)和RA(os))和位置准确度(Dsr)说明与参考对象最相似的分割对象结果是在尺度20时获得的分割结果(RA(or)和RA(os)较大且接近,Dsr较小)。光谱及光谱+纹理结合特征分类方案分类总精度具有相似的变化特征,即小尺度分类总精度较低,随着尺度增大,分类总精度也再提高并在某个尺度达到最大值,之后分类总精度随尺度增大而降低,并趋于平稳。光谱+纹理结合特征的分类总精度低于仅使用光谱特征的分类总精度,在中小尺度上尤其明显。在尺度参数为20时使用光谱+纹理结合特征分类总精度低19个百分点,在尺度参数为30时低13个百分点。基于航空正射影像分割最优尺度参数为20。[结论]基于航空正射影像进行林隙识别,最高精度仅为74%(Kappa=61%),林隙生产者和用户精度在60%以上,非林隙生产者和用户精度在90%左右。基于航空正射影像林隙识别纹理特征的加入还将继续降低识别精度。

关键词: 林隙, 影像分割, 对象特征, 航空正射影像, 支持向量机, 面向对象

Abstract: [Objective] The effects of different segmentation scales and characteristics of objects on object-oriented forest gap based on aerial orthophoto were studied,and the suitability of forest gap classification based on aerial orthophoto was also evaluated in this research.[Method] In this study,based on true color aerial orthophoto data and object-oriented classification method,northeastern typical natural secondary forest-Mao'ershan Experimental Forest Farm Donglin industry zone was selected as the study area for classification of forest gap.In the process of object-oriented classification,for three components of aerial orthophoto (Blue,Green,Red),10 scales (10-100,step size 10) was divided to carry out segmentation,and topological combination accuracy (RA(or) and RA(os)) and position accuracy (Dsr) were used to evaluate the classification results.For each segmentation result of different scales,spectral features and combination features of spectrum and texture derived from aerial orthophoto were used,and support vector machine (SVM) with linear kernel classifier of object-oriented was used to classify the study area into forest gap,non-forest gap and canopy,totally 20 classification result were obtained.Later the accuracy of each classification result was evaluated by the 4 different evaluation indexes,namely producer accuracy calculated by Confusion matrix,user accuracy,classification accuracy and Kappa coefficient.[Result] Characteristics of objects (size and shape) were affected by scale parameter.Objects in small area were created by small scale segmentation,and objects in big area were created on large scale segmentation,however,segmentation of large scale could not efficiently distinguish obvious forest gaps from canopies,it indicates an obvious insufficient segmentation phenomenon.The topological combination accuracy (RA(or) and RA(os)) and position accuracy (Dsr) illustrated that the result of segmentation objects on 20 scale (RA(or) and RA(os) were higher and close,Dsr was lower) was most similar to the reference objects comparing to the result on other segmentation scales.The segmentation schemes of spectral features and combination features of spectrum and texture showed a similar variable trend:classification accuracies were lower on small scales,and then increased with scale increasing,after reached the maximum on a certain scale,they decreased to stable state.The classification accuracy using combination features of spectrum and texture was lower than that using spectral features only,especially on small scales.On the segmentation scale of 20,the overall classification accuracy using combination features of spectrum and texture was 19% lower than that using only spectral features,and accordingly 30% lower on the segmentation scale of 30.The optimal scale parameter of segmentation based on aerial orthophoto was 20.[Conclusion] The highest recognition accuracy was just 74%(Kappa=0.61) when using aerial orthophoto to recognize forest gaps,the producer and user accuracy of forest gap were all above 60%,and which of non-forest gap were around 90%.The classification accuracy based on aerial orthophoto would continually decrease when adding texture features.

Key words: forest gap, image segmentation, object features, aerial orthophoto, SVM, object-based

中图分类号: