基于PCA的森林生物量遥感信息模型研究

森林生物量和遥感多光谱数据、植被指数及地学因子存在相关关系,但这些因子间可能存在着多重相关性,如利用这些因子直接建模估测森林生物量,则可能出现病态模型。因此,文章采用主成分分析方法,提取遥感及地学因子的主成分,再建立主成分与生物量多元线性回归模型,估测森林生物量,达到既可保留多个遥感及地学因子的主要信息,又可避免因子间共线性的问题,以及降维,简化模型的作用。文章以高黎贡山自然保护区常绿阔叶林为研究对象,利用地面样地胸径每木调查数据,结合生物量相对生长式,得样地生物量。利用2006年印度卫星(IRS)数据,包括B2、B3、B4、B5四个波段,提取DVI、NDVI、PVI、RVI、VI3、SLAVI六种植被指数,利用DEM提取海拔、坡度、坡向值共13个遥感及地学因子。在此基础上,提取13个因子的主成分,第一主成分至第五主成分的累计贡献率达98.7%。以前5个主成分值作自变量,建立主成分与地面生物量的回归模型,模型经方差分析及相关性检验,达到显著相关水平,相关系数R=0.7129,可用于森林生物量估测。

Remote-sensing information model of forest biomass based on principal components analysis

Forest biomass is correlated to the factors such as remotely sensed data, vegetation indexes and topographic features. However, these factors are usually strongly correlative. The forest biomass model becomes an ill-posed one if the model is built directly with those factors. The principal components (PCs) for those factors are obtained by principal components analysis (PCA), and then the forest biomass model is built by linear regression with PCs as the input variables. In that case, not only are the most information of these factors reserved in the model, but also the multi-colliearity problem of these factors is avoided. Moreover, the number of the variables decreases and the model is optimized. The procedure for the model building is as follows: firstly the forest biomass of the plots in the evergreen broadleaved forest in Mt. Gaoligong National Nature Reserve is calculated based on the relative growth equations and the diameter of all the trees in the plot. Secondly the PCs for the 13 factors, which include 4 multi-spectral bands of B2, B3, B4, B5 of IRS in 2006, 6 kinds of vegetation index of DVI, NDVI, PVI, RVI, VI3, SLAVI , Elevation, slope and aspect generated from DEM, are analyzed by PCA. The results show that the accumulative ratio of contribution of the first 5 PCs is 98.7%. Finally, a forest biomass model is set up by regression based on these first 5 PCs. F test examination shows that forest biomass is correlated significantly to these first 5 PCs with coefficient R being 0.712 9. The remotely sensed information model based on PCA is effective for the estimation of forest biomass.