Journal of Geo-information Science ›› 2020, Vol. 22 ›› Issue (9): 1897-1909.doi: 10.12082/dqxxkx.2019.190598

Previous Articles     Next Articles

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

CHEN Ang1(), YANG Xiuchun1,2,*(), XU Bin1,2, JIN Yunxiang1, ZHANG Wenbo1, GUO Jian1, XING Xiaoyu1, YANG Dong1   

  1. 1. Key Laboratory of Agri-informatics, Ministry of Agriculture/Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081, China
    2. College of Grassland Science, Beijing Forestry University, Beijing 100083, China
  • Received:2019-10-12 Revised:2019-12-10 Online:2020-09-25 Published:2020-11-25
  • Contact: YANG Xiuchun;
  • Supported by:
    National Key Research and Development Program of China(2017YFC0506504);National Natural Science Foundation of China(41571105)


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, ortho-rectification, 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 was 90.14%, 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 higher spatial resolution, richer texture and shape information. Object-based method had high applicability for both kinds of images, and the accuracy of three classifiers were similar.The deep learning method was more suitable for UAV image in this paper, it can effectively reduce the misclassification phenomenon in UAV image.In the future, a higher quantity and quality sample database should be constructed to further improve the accuracy of deep learning method and provide support for the management and research of elm sparse forest.

Key words: Elm sparse forest, UAV, object-based method, machine learning, deep learning, Hunshandake sandy land