摘要
【目的】为有效估测毛竹(Phyllostachys edulis)林面积提供技术参考,该文尝试基于卷积神经网络(CNN)的深度学习和面向对象相结合的分类方法提取毛竹林信息。【方法】以研究区毛竹林为研究对象,以无人机影像为数据源,采用CNN的深度学习模型识别无人机影像中各典型地物,在此基础上对无人机影像进行多尺度分割,使用隶属度函数分类方法对毛竹林进行分类,并与采用随机森林(RF)的分类方法提取毛竹林的结果进行比较。【结果】(1)该文采用的分类方法既可以避免面向对象的分类方法难以充分利用影像的深层次特征,又能解决CNN在影像分类时易丢失地物目标边缘、形状等细节信息的难题。(2)分类方法提取的毛竹林总体精度为91.04%,Kappa系数为0.89;RF分类方法总体精度为85.07%,Kappa系数为0.82,总体精度提高5.97个百分点,Kappa系数提高0.07。(3)根据2023年度林草湿资源图成果数据统计,研究区毛竹林面积为150.31 hm^(2),该文采用的分类方法提取毛竹林面积为161.99 hm^(2),提取结果比实际多11.68 hm^(2),总体上吻合程度较高。【结论】结合CNN的深度学习和面向对象的分类方法可以提高毛竹林信息提取精度,可为毛竹林的精细化提取提供有效技术参考。
【Objective】In order to provide technical reference for effectively estimating the area of bamboo forests,the paper tried to make information extraction of Phyllostachys edulis forests with a joint method of convolutional neural networkbased(CNN)deep learning and objectoriented classification.【Method】Taking the Ph.edulis forests in research area as the research object and the UAV image as the data source,the deep learning model of CNN was used to identify the typical ground objects in the UAV image,and then the distribution information of Ph.edulis forests was extracted by using the objectoriented multiscale segmentation and the classification method of the membership function.The accuracy of the classification results was compared with another accuracy extracted by using the random forest(RF)classification method.【Result】(1)The classification method adopted in this paper can not only avoid the difficulty of objectoriented classification method to make full use of the deep features of images,but also solve the problems of detailed information loss such as edge and shape of ground objects during image classification of CNN.(2)The overall accuracy of Ph.edulis forests extracted by the classification method adopted in this paper was 91.04%,and the Kappa coefficient was 0.89;another overall accuracy of RF classification method was 85.07%,and the Kappa coefficient was 0.82.The overall accuracy was increased by 5.97 percentage points,Kappa coefficient was increased by 0.07.(3)According to the statistical data of forest,grassland and wet land resource results in 2023,the area of Ph.edulis forests in the study area was 150.31 hm^(2).However,the area of Ph.edulis forests extracted with the classification method adopted in this paper was 161.99 hm^(2),and the extracted result was 11.68 hm^(2)more than the actual result.In general,the degree of coincidence is high.【Conclusion】The joint method of deep learning based on CNN and objectoriented classification can improve the accuracy of information extraction,and it can also provide effective technical reference for the refinded extraction of Ph.edulis forests.
作者
王海龙
郜昌建
胡昱彦
林荫
应建平
徐达
WANG Hai-long;GAO Chang-jian;HU Yu-yan;LIN Yin;YING Jian-ping;XU Da(Zhejiang Forestry Survey Planning and Design Co.,Ltd.,Hangzhou 310020,Zhejiang,China;East China Academy of Inventory and Planning of NFGA,Hangzhou 310019,Zhejiang,China;Longyou Forestry and Water Resources Bureau,Longyou 324400,Zhejiang,China;Zhejiang Forest Resources Monitoring Center,Hangzhou 310020,Zhejiang,China)
出处
《竹子学报》
2024年第3期1-11,共11页
Journal of Bamboo Research
