基于Landsat 8 OLI和资源3号立体像对数据的桉树森林蓄积量估测  

Eucalyptus forest stock estimation study based on Landsat 8 OLI and Resource 3 stereo image

作　　者：张方圆 吴胜义 乔海亮 许舟[3] ZHANG Fangyuan;WU Shengyi;QIAO Hailiang;XU Zhou(Northwest Inventory and Planning Institute,National Forestry and Grassland Administration,Xi’an 710048,Shaanxi,China;Shaanxi Provincial Forestry Research and Planning Institute,Xi’an 710041,Shaanxi,China;Central South University of Forestry&Technology,Changsha 410004,Hunan,China)

机构地区：[1]国家林业和草原局西北调查规划院,陕西西安710048 [2]陕西省林业调查规划院,陕西西安710041 [3]中南林业科技大学,湖南长沙410004

出　　处：《中南林业科技大学学报》2024年第5期93-101,共9页Journal of Central South University of Forestry & Technology

基　　金：国家自然科学基金项目(31901311)。

摘　　要：【目的】探索Landsat8 OLI数据和立体数据在估算桉树森林蓄积量(forest stock volume,FSV)中的潜力,并且准确地估计桉树的FSV。【方法】以3幅Landsat8 OLI图像和资源3号立体数据为遥感数据源,并且结合少量地面调查数据实现了桉树FSV的遥感估计。研究中提取了三类遥感特征用于估计桉树FSV:第一类是包括植被指数和单波段反射率在内的光谱特征;第二类是基于Landsat 8 OLI图像的单波段提取的8种纹理特征;第三类是基于资源3号立体像对数据和开源的数字高程模型(digital elevation model,DEM)提取的冠层高度模型(canopy height model,CHM)。利用Boruta算法对三类遥感特征进行提取,之后建立了随机森林(random forest,RF)、K-最近邻(K-nearest neighbor,KNN)和支持向量机(support vector machine,SVM)3种机器学习模型以及传统的多源线性回归模型(multiple linear regression,MLR),并以决定系数(R2)、均方根误差(root mean square error,RMSE)和相对均方根误差(relative root mean square error,rRMSE)作为评价指标对模型结果进行评估。【结果】基于ZY-3立体像对数据和开源的DEM数据提取的CHM与桉树的FSV具有很强的相关性,Pearson相关系数达到了0.71。仅仅利用基于Landsat 8 OLI图像提取的光谱和纹理特征难以准确2地估计桉树的FSV,估测模型的R为0.29~0.38,rRMSE为35.65%~43.30%,存在严重的数据饱和问题。2当变量集中加入CHM后,模型的估测精度明显提高,R达到了0.64~0.66,rRMSE为25.74%~26.41%。【结论】使用Landsat 8 OLI数据估算桉树FSV时存在严重的数据饱和问题,并且使用空间分辨率为30 m的纹理特征难以有效地改善森林蓄积量的估计精度。利用资源3号立体像对数据和开源的DEM数据可以提取较为准确的CHM,并且所提取的CHM可以解决改善光学数据的饱和问题,从而提高桉树FSV的估计精度。【Objective】To explore the potential of Landsat8 OLI data and stereo data in estimating Eucalyptus forest stock volume(FSV)and to accurately estimate eucalyptus FSV.【Method】This study used three Landsat8 OLI images and ZY-3 stereo data as remote sensing data sources and combined with a small amount of ground survey.The remote sensing estimation of Eucalyptus FSV was achieved by combining a small amount of ground survey data.There were three main types of remote sensing features used to estimate the FSV of eucalyptus:the first was the spectral features including vegetation index and single-band reflectance;the second was the eight texture features extracted based on the single-band of Landsat 8 OLI images;the third was the canopy height extracted based on the Resource 3 stereo image pair data and the open-source digital elevation model(DEM).DEM extracted from the canopy height model(CHM).The Boruta algorithm was used to characterize the three types of remote sensing features,and then three machine learning models,Random Forest(RF),K-Nearest neighbor(KNN)and Support vector machine(SVM),as well as the traditional multiple linear regression model(MLR),were established.And evaluated the model results with coefficient of determination(R2),root mean square error(RMSE)and relative root mean square error(rRMSE)as evaluation metrics.The model results were evaluated.【Result】The CHM extracted from ZY-3 stereo relative data and open-source DEM data had a strong correlation with the FSV of Eucalypts,with a Pearson correlation coefficient of 0.71.It was difficult to accurately estimate the FSV of Eucalypts using only spectral and texture features extracted from 2 Landsat 8 OLI images,and the R of the estimated model was 0.29-0.38,rRMSE of 35.65%-43.30%,and a serious data saturation problem.2 When CHM was added to the variable set,the estimation accuracy of the model improved significantly,with R reaching 0.64-0.66 and rRMSE of 25.74%-26.41%.【Conclusion】There were serious data saturation problems when estimating E

关 键 词：森林蓄积量 Landsat 8 OLI 立体像对 冠层高度模型 遥感建模 

分 类 号：S771.8[农业科学—森林工程]