基于多源数据的毛竹林信息自动提取  

Information Extraction of Phyllostachys edulis Forest based on Multi Source Data

作　　者：王广科 吴达胜[1,2,3] 方陆明 WANG Guang-ke;WU Da-sheng;FANG Lu-ming(School of Mathematics and Computer Science,Zhejiang A&F University,Hangzhou 311300,China;Key Laboratory of Forestry Intelligent Monitoring and Information Technology of Zhejiang Province,Hangzhou 311300,China;Key Laboratory of State Forestry and Grassland Administration on Forestry Sensing Technology and Intelligent Equipment,Hangzhou 311300,China)

机构地区：[1]浙江农林大学数学与计算机科学学院,浙江杭州311300 [2]浙江省林业智能监测与信息技术研究重点实验室,浙江杭州311300 [3]林业感知技术与智能装备国家林业和草原局重点实验室,浙江杭州311300

出　　处：《浙江林业科技》2023年第3期79-86,共8页Journal of Zhejiang Forestry Science and Technology

基　　金：浙江省科技重点研发计划资助项目(2018C02013)。

摘　　要：毛竹Phyllostachys edulis是重要的经济林种,快速准确地获取毛竹林的面积及郁闭度等信息可对毛竹林的高效经营管理提供巨大帮助。基于人工样地的森林资源调查耗时费力且效率低下,故利用遥感图像等较低成本的数据源估测大范围毛竹林的面积及其郁闭度等信息具有重要意义。本文以浙江省安吉县为研究区域,基于Sentinel-2中等分辨率遥感图像、数字高程模型(Digital Elevation Model,DEM)数据及森林资源二类调查数据,利用CatBoost、随机森林(Random Forest,RF)、支持向量机(Support Vector Machine,SVM)三种算法,通过二分类方法估算毛竹林面积和多分类方法估算毛竹林郁闭度。结果表明:在估算毛竹林面积时,CatBoost、RF、SVM的总体分类精度分别为97.75%、96.54%、95.05%,Kappa系数分别为0.96、0.93、0.90;在估算毛竹林郁闭度时,以上三种算法的总体分类精度分别为73.38%、73.49%、69.9%,Kappa系数分别为0.52、0.52、0.45。在本文研究数据中,0.7、0.8郁闭度样本(大类别样本)占样本总数的89.42%,剩余类别样本(小类别样本)只占样本总数的10.58%。样本失衡是导致毛竹林郁闭度估算精度严重偏低的重要原因。将样本分成两种类别后(占比较大的样本和占比较小的样本)分别进行建模,模型估测效果有了较大改善,各模型郁闭度估算精度均在78%以上,其中以CatBoost最优,总体精度达到83.45%。基于Sentinel-2遥感图像、DEM及森林资源二类调查数据的多源数据估算毛竹林面积和郁闭度,CatBoost算法具有最佳的性能指标,其估算结果可为竹林资源的监测提供重要借鉴。Based on images from Sentinel-2 medium resolution remote sensing imagery,digital elevation model(DEM)and bamboo(Phyllostachys edulis)data from forest management survey of Anji county,Zhejiang province in 2018,estimation was made on area and crown density of bamboo forest by CatBoost,random forest(RF)and support vector machine(SVM)model,dichotomy idea and multi-classification.The result demonstrated that classification accuracy of bamboo forest area was 97.75%by CatBoost,96.54%by RF and 95.05%by SVM,the kappa coefficients were 0.96,0.93 and 0.9 respectively.The result of Moso bamboo forest canopy density estimated were as follows:the overall classification accuracy of the above three algorithms were 73.38%,73.49%and 69.9%respectively,and the kappa coefficients were 0.52,0.52 and 0.45 respectively.The estimation accuracy of models was greatly improved after the samples divided into two categories(large sample and small sample)and modeled separately.The estimation accuracy of crown density of each model was greater than 78%,especially by CatBoost,83.45%.

关 键 词：毛竹林 遥感 机器学习 信息提取 

分 类 号：S711[农业科学—林学]