摘要
叶面积指数(leaf area index,LAI)是体现林分冠层结构的一项重要参数,其准确估测对于精准林业的实施具有重要意义.为了快速、无损地监测毛竹林LAI,采用ISI921VF-256野外地物光谱辐射计和LAI-2200冠层分析仪获取福建省西北部毛竹林分冠层光谱反射率和LAI值,通过敏感波段的选取,新建了8类植被指数,分析了LAI值与对应植被指数的相关性,进而利用随机森林回归、支持向量回归和反向传播神经网络法构建了毛竹林分冠层LAI高光谱估测模型,以决定系数(R2)、均方根误差(ERMS)、平均绝对误差(EMA)和估测值与实测值的回归线斜率为指标评价并比较了模型预测精度.结果表明:新建的NDVI674、NDVI687、GNDVI563、GRVI563、RVI674、RVI687、DVI674、DVI687八类植被指数与LAI均呈极显著相关(P<0.01).建立的RFR模型中,决定系数R2达到0.7323,分别比SVR模型和BP模型提高了0.1066和0.2470;其EMA为0.4062,分别比SVR模型和BP模型减少了0.0448和0.4811;其ERMS为0.6463,略高于SVR模型,但远小于BP模型;其实测值与估测值的回归线斜率接近1,优于SVR模型和BP模型的回归线斜率.RFR模型对毛竹林分冠层LAI的高光谱估测效果优于SVR模型和BP模型,可用于大区域范围毛竹林冠LAI的高光谱估测.
Abstract
Leaf area index ( LAI) is an important parameter to embody forest canopy structure and its accurate estimation is a great significance for implementation of precision forestry.For monitoring Phyllostachys e dulis LAI rapidly and non-destructively, ISI921VF-256 field spectral radiometer and LAI-2200 canopy analyzer were used to acquire P.edulis canopy and LAI value in the northwest of Fujian Province, respectively.Sensitive bands were selected to construct 8 new vegetation indexes and the correlation between LAI and its vegetation indexes was analyzed.And then random forest regression (RFR), support vector regression (SVR) and back propagation (BP) were used to construct hyperspectral estimation models of P.edulis forest canopy LAI that the coefficient of determination (R2 ), root mean square error (ERMS), the mean absolute error (EMA ), and the slope of the regression line between the estimated and actual value were as evaluation indexes and applied to compare model prediction accuracy.Results showed that 8 new built vegetation indexes which were NDVI674, NDVI687, GNDVI563, GRVI563, RVI674,RVI 687, DVI674, DVI687 were significantly correlated with LAI (P<0.01).The coefficient of determination (R2) was 0.7323 by using RFR model, which improved 01.066 and 0.2470 than SVR model and BP model, respectively.TheE MA was 0.4062 by using RFR model, which decreased 0.0448 and 0.4811 than SVR model and BP model, respectively.The ERMS was 0.4062 by using RFR model, which was slightly more than SVR model and much smaller than BP model, respectively.The slope of the regression line between the estimated and actual value was close to 1 by using RFR model, superior to the SVR and BP model.Hyperspectral estimation effect of RFR model for hyperspectral estimation of P. edulis forest canopy LAI had an advantage to SVR model and BP model, suggesting RFR model can be applied to region-wide hyperspectral estimation of P.edlu is forest canopy LAI.
国家科技期刊平台
NSTL
维普
万方数据