三种回归分析方法在Landsat-5影像森林碳密度反演中的比较

随着遥感技术的快速发展,基于遥感影像和地面样地的方法成为目前森林碳密度精确估算的主要手段,然而没有找到具有普适性的建模因子和最佳的森林碳密度估算模型。鉴于此,本文通过分析研究区地面固定样地碳密度与Landsat-5影像及其衍生波段的相关性,筛选出估算森林碳密度的敏感因子。采用三种回归分析方法(逐步回归、偏最小二乘回归及非线性回归)分别建立森林碳密度的最优遥感估算模型。结果表明:1参与建模的遥感因子中,1/TM3与森林碳密度的相关性最大,敏感性最高;2三种回归分析方法建立的预测模型中,以4个遥感因子建立的非线性回归模型预测精度最高,预测值与实测值得决定系数R2为0.74;3通过测算,研究区平均森林碳密度为14.36 t/hm2,变化范围介于0.00~38.28 t/hm2之间。研究表明非线性回归在区域森林碳密度反演方面具有一定的潜力。

Comparison of Application of Three Regression Analysis Methods on Estimation of Forest Carbon Density using Landsat-5 Data

With the fast development of remote sensing technology,the method based on remote sensing image and sam-ple plots has become the major means of accurate estimation of forest carbon density.However,there were still no uni-versal factors and optimal models for the estimation of forest carbon density.The objective of this paper was to study the estimation of forest carbon density by combining plots data and remote sensing images of Landsat-5 using the methods of stepwise regression,partial least-squares regression and nonlinear regression respectively.First,various remote sensing factors derived from Landsat-5 images were generated using different transformations such as band ratios,vegetation indi-ces calculation,principal component analysis and texture transformation.Then,effective remote sensing factors were se-lected to conduct the estimation of forest carbon density,according to the correlation analysis between the fixed sample plots and factors derived from Landsat-5 images.Finally,the accuracy of Landsat-5 derived maps was assessed using R2 , Root Mean Square Error and Relative Error.The results showed that ①the correlation coefficients of 1/TM3 with plot values was the highest.② Among the built models,the effect of nonlinear model built by four remote sensing factors was the best with R2 of 0.74.③The mean value of forest carbon density of research area was 14.36 t/hm2 ,ranging from 0. 00 to 38.38 t/hm2 .This implied that nonlinear regression showed a certain potential in the aspect of region estimation of forest carbon density.