Estimating foliage nitrogen concentration from HYMAP data using continuum removal analysis

Abstract

The concentrations of various foliar chemicals can be estimated by analyzing the spectral reflectance of dried ground leaves. The continuum-removal analysis of Kokaly and Clark [ Remote Sens. Environ. 67 (1999) 267] has been an improvement on the standard derivative analysis in such applications. Continuum-removal analysis enables the isolation of absorption features of interest, thus increasing the coefficients of determination and facilitating the identification of more sensible absorption features. The purpose of this study was to test Kokaly and Clark's methodology with aircraft-acquired hyperspectral data of eucalypt tree canopies, which are more complex than are spectra from many coniferous canopies and much more complex than the spectra from dried ground leaves. The results of the continuum-removal analysis were most encouraging. It identified, in one experiment or another, almost all of the known nitrogen absorption features. The coefficient of determination in one case increased from 0.65, using the standard derivative analysis, to 0.85 with the continuum-removal analysis. It is recommended that continuum-removal analysis become at least a supplement to standard derivative analysis in estimating foliar biochemical concentrations from remote sensing data. This study also reports several other findings: (1) the neural network method generally achieved higher coefficients of determination and lower errors of estimation [root mean square error (RMSE) and standard error of cross validation (SECV)] than did the modified partial least squares (PLS) or stepwise regression methods, probably indicating nonlinear relationships between biochemical concentrations and canopy reflectance; (2) modified partial least squares (MPLS) proved a better statistical method than conventional stepwise regression analysis in many cases in terms of both coefficient of determination and RMSE; and (3) the maximum spectrum of a cluster of tree pixels represents canopy reflectance at least as well as the mean spectrum of the cluster, especially when used in conjunction with the modified partial least squares method.