基于面向对象与深度学习的榆树疏林识别方法研究  

Research on Recognition Methods of Elm Sparse Forest based on Object-based Image Analysis and Deep Learning

作　　者：陈昂 杨秀春[1,2] 徐斌 金云翔[1] 张文博 郭剑[1] 邢晓语[1] 杨东[1] CHEN Ang;YANG Xiuchun;XU Bin;JIN Yunxiang;ZHANG Wenbo;GUO Jian;XING Xiaoyu;YANG Dong(Key Laboratory of Agri-informatics,Ministry of Agriculture/Institute of Agricultural Resources and Regional Planning,Chinese Academy of Agricultural Sciences,Beijing 100081,China;College of Grassland Science,Beijing Forestry University,Beijing 100083,China)

机构地区：[1]农业部农业信息技术重点实验室中国农业科学院农业资源与农业区划研究所,北京100081 [2]北京林业大学草业与草原学院,北京100083

出　　处：《地球信息科学学报》2020年第9期1897-1909,共13页Journal of Geo-information Science

基　　金：国家重点研发计划项目(2017YFC0506504);国家自然科学基金项目(41571105)。

摘　　要：榆树疏林是浑善达克沙地中一种特殊的植被类型,它对于维持区域生态系统稳定具有重要意义,在防风固沙、涵养水源、调节气候等方面发挥着重要的作用。本文利用无人机影像与GF-2影像,对高分辨率数据源中榆树疏林的两种自动识别方法进行了研究。在面向对象方法中,首先通过计算影像对象的局部方差变化率得到了最佳分割尺度;其次采用随机森林算法对初选特征的重要性进行排序,并删除无关特征;最后分别对支持向量机(SVM)、随机森林(RF)、深度神经网络(DNN)3种分类器进行参数寻优与榆树疏林提取。此外,在ENVI5.5中基于TensorFlow框架,利用U-Net构建深度学习模型对榆树疏林进行了提取,并与面向对象方法进行对比。结果显示:①通过面向对象方法过程的优化,最终的识别精度较以往研究有所提升,GF-2影像中SVM总体精度为90.14%,RF总体精度为90.57%,DNN总体精度为91.14%;无人机影像中SVM总体精度为97.70%,RF与DNN总体精度为97.42%。②深度学习方法中,GF-2影像的总体精度为91.00%,无人机影像的总体精度达到了98.43%。研究结果说明在榆树疏林提取中,无人机影像具有更高的空间分辨率,更丰富的纹理、形状等信息,能达到比GF-2影像更高的精度。面向对象方法对于2种影像都有较高的适用性;深度学习的方法在本文中更适用于无人机影像,它可以有效地减少无人机影像中的错分现象。Elm sparse forest is a special vegetation type in Hunsh and ake sandy land.It has important significance for maintaining the stability of regional ecosystem,and plays a key role in sand fixation,water conservation and climate regulation.Rapid and accurate access to the distribution of elm sparse forest is conducive to the protection of the fragile ecosystem in the area.In this paper,the automatic recognition methods of elm sparse forest in high spatial resolution data source was studied by using Unmanned Aerial Vehicle(UAV)image and GF-2 image.After processing the original images of UAV,the Digital Ortho photo Map and the Canopy Height Model were obtained.The preprocessing of GF-2 data included atmospheric correction,orthorectification,image fusion et al.In the object-based method,firstly,the optimal segmentation scale was obtained by calculating the change rate of local variance in the image objects;Secondly,the importance of the selected features was sorted by the random forest algorithm,and the irrelevant features were deleted;Finally,the parameters of three classifiers,namely,Support Vector Machine(SVM),Random Forest(RF)and Deep Neural Network(DNN),were optimized,and then they were used to identify the elm sparse forest.In addition,based on the Tensor Flow framework in ENVI 5.5,a deep learning model based on U-Net was constructed to identify elm sparse forest.The results showed that:(1)through the optimization of the object-based method process,the final recognition accuracy was improved than the privious study.In GF-2 image,the overall accuracy of SVM was90.14%,the overall accuracy of RF was 90.57%,and the overall accuracy of DNN was 91.14%.In UAV image,the overall accuracy of SVM was 97.70%,and the overall accuracy of RF and DNN were 97.42%.(2)In the deep learning method,the overall accuracy of the GF-2 image was 91%,and the overall accuracy of the UAV image reached 98.43%.The results illustrated that UAV image can achieve higher accuracy than GF-2 image in elm sparse forest recognition because of its highe

关 键 词：榆树疏林 无人机 面向对象 机器学习 深度学习 浑善达克沙地 

分 类 号：S718.5[农业科学—林学]