基于星载光子计数雷达数据的森林郁闭度估测模型优化

Optimization of forest canopy closure estimation model based on spaceborne photon counting LiDAR data

【目的】旨在评估星载光子计数雷达数据估测森林郁闭度(Forest Canopy Closure,FCC)的潜力,为优化森林管理规划提出一种高效率、低成本估测区域尺度FCC的新技术方法。【方法】研究以星载激光雷达ICESat-2/ATLAS光子点云数据为信息源,以滇西北生态脆弱区香格里拉为研究区,结合54块地面实测样地数据,在前期对点云数据进行去噪、分类等预处理的基础上,对研究区74808个有林地光斑冠层参数进行提取(共计59个),使用支持向量机的递归特征消除算法(SVM-RFE)优选特征变量,采用普通克里格(OK)插值出区域尺度特征变量的空间分布,基于贝叶斯优化(BO)算法改进后的随机森林(RF)、K-最近邻值法(KNN)、梯度回归(GBRT)模型建模,以决定系数(R^(2))、均方根误差(RMSE)、总体预测精度(P)、残差平方和(R_(SS))和相对均方根误差(R_(RMSE))作为模型评价指标,以此构建研究区FCC估测模型。【结果】1)由ICESat-2/ATLAS提取的光斑冠层参数经SVM-RFE优选后,6个(asr、n_toc_photons、n_ca_photons、h_min_canopy、toc_roughness、photon_rate_can)冠层参数的平均交叉验证精度高为0.60,可作为OK插值变量;2)以优选的冠层参数作为OK插值变量拟合最佳半方差函数,所有变量的块金效应(SR<25%)较弱,具有强烈的空间自相关性,除asr变量的最佳拟合模型为球状模型(R^(2)=0.689,R_(SS)=2.05×10^(-6),R_(RMSE)=0.1602)外,其余5个变量的最佳拟合模型均为指数模型(R^(2),0.71~0.93;R_(SS),2.34×10^(-9)~1.54×10^(-4);R_(RMSE),0.0239~0.1886);3)在BO-RF、BO-GBRT、BO-KNN郁闭度估测模型中,以BO-RF模型综合建模精度最高(R^(2)=0.73,RMSE=0.09、P=80.13%),可作为研究区FCC遥感估测模型;4)使用BO-RF模型估测的研究区FCC进行空间制图,均值为0.53,主要分布在0.3~0.6之间,占比77.44%。FCC高值区域总体呈现出由东南向北延伸分布的趋势,与研究区森林资源分布情况基本一致。【结论】该方法可为优化森林资源管理提供一种技术与方法上的参考。

【Objective】This paper aims to evaluate the potential of spaceborne photon counting LiDAR data to estimate forest canopy closure(FCC),also in order to propose a new technical method for optimizing forest management planning with high efficiency and low cost to estimate regional-scale FCC.【Method】The study took the spaceborne lidar ICESat-2/ATLAS photon point cloud data as the information source,and took the ecologically fragile area of Shangri-La in northwest Yunnan as the study area,combined with the data of 54 ground measured sample plots.In the early,based on the pre-processing of point cloud data such as denoising and classification,74808 footprints canopy parameters of forested land in the study area were extracted(59 in total).The recursive feature elimination algorithm of support vector machine(SVM-RFE)was used to optimize the feature variables,and the spatial distribution of regional-scale feature variables was obtained by ordinary Kriging(OK)interpolation.Modeling of improved random forest(RF),k-nearest neighbor(KNN),and gradient regression tree(GBRT)models by Bayesian Optimization(BO)algorithms,the determination coefficient(R^(2)),root mean square error(RMSE),overall prediction accuracy(P),residual sum of squares(R_(SS))and relative root mean square error(R_(RMSE))were used as model evaluation indexes,which construct the FCC estimation model in the study area.【Result】1)The average crossvalidation accuracy of the six(asr,n_toc_photons,n_ca_photons,h_min_canopy,toc_roughness,photon_rate_can)footprint canopy parameters extracted by ICESat-2/ATLAS after SVM-RFE optimization was 0.60,which could be used as OK interpolation variables.2)The optimal canopy parameter was used as the OK interpolation variable to fit the best variance function,the nugget effect(SR<25%)of all variables was weaker and had strong spatial autocorrelation.Excepting that the best fitting model of asr variable was spherical model(R^(2)=0.689,R_(SS)=2.05×10^(-6),R_(RMSE)=0.1602),and the best fitting models of the other 5 variables were exponential models(R^(2),0.71-0.93;R_(SS),2.34×10^(-9)-1.54×10^(-4);R_(RMSE),0.0239-0.1886).3)Among BO-RF,BO-GBRT and BO-KNN of forest canopy closure estimation models,the BO-RF model had the highest comprehensive modeling accuracy(R^(2)=0.73,RMSE=0.09,P=80.13%),which could be used as FCC remote sensing estimation model in the study area.4)The spatial mapping of the study area FCC estimated by the BO-RF model had a mean value of 0.53,mainly distributed between 0.3 and 0.6,accounting for 77.44%.The high value area of FCC generally showed a trend of extending the distribution from southeast to north,which was basically consistent with the distribution of forest resources in the study area.【Conclusion】This method can provide a technical and methodological reference for optimizing forest resource management.