偏最小二乘回归在Hyperion影像叶面积指数繁衍中的应用

叶面积指数(Leaf Area Index，LAI)是一个重要的森林结构参数指标，遥感技术被认为是区域LAI反演的有效手段。现有遥感反演模型多以单变量的曲线估计及线性回归模型为主，模型的通用性、建模精度以及植被指数的选择上需要更进一步的探讨。论文以攸县黄丰桥林场为研究区，Hyperion影像为数据源，提取归一化植被指数(NDVI)、比值植被指数(RVI)等13个因子，利用LAI-2000冠层分析仪开展130个样地(60 m×60 m)的叶面积指数测量，选用变量投影重要性(VIP)指标、变量解释能力及变量权重作为变量筛选的依据，采用偏最小二乘回归分析方法建立植被指数与实测样地的回归模型，开展叶面积指数反演并制图。研究结果表明：偏最小二乘回归分析在LAI反演中取得了较好的预测效果，其中以6个植被因子建立的回归模型预测精度最高，预测值与实测值的决定系数R2为0.91；LAI与植被指数之间具有良好的线性关系，其中RVI与LAI的相关性最大；残差分析表明，反演模型的自变量个数选取以4~6个为宜。

Partial Least Squares Regression Application in LAI Inversion Using Hyperion Data

Leaf area index （LAI） is an important indicator of forest structural parameters. Remote sensing was considered to be an effective means of regional LAI inversion. Single variable curve estimate and linear regression was the dominant model in LAI inversion using remote sensing, but the model generality, modeling accuracy and vegetation index on the choice of the need to further discussed. Hyperion data as a data source, through LAI-2000 instrument to obtain 130 samples （60 m×60 m） leaf area index of ground measurements at Huangfengqiao forest farm in Youxian County. Extract NDVI, RVI and other 13 factors from Hyperion image, Variable Importance in the Projection （VIP）, Proportion of Variance Explained and variable weights was used in variable selection, using partial least squares regression analysis （PLS） to establish vegetation index and measured plots of regression models, which to inversion the leaf area index and mapping in study area. The results as followed： PLS regression analysis had good prediction effect in LAI inversion, among all the fitting models, the effect of 6 vegetation factors atrial least-square regression was best with R 2 coefficient of 0.91; LAI and vegetation index had a good linear relationship. The results showed that the sensitivity of ratio vegetation index （RVI） was highest among all the modeling factors. Residual analysis showed that it was reliable to build model using 4 to 6 independent variables, prediction accuracy of partial least-square regression was highest.