Comparing prediction power and stability of broadband and hyperspectral vegetation indices for estimation of green leaf area index and canopy chlorophyll density

doi:10.1016/S0034-4257(00)00197-8

Remote Sensing of Environment

Volume 76, Issue 2, May 2001, Pages 156-172

https://doi.org/10.1016/S0034-4257(00)00197-8 Get rights and content

Abstract

Hyperspectral reflectance data representing a wide range of canopies were simulated using the combined PROSPECT+SAIL model. The simulations were used to study the stability of recently proposed vegetation indices (VIs) derived from adjacent narrowband spectral reflectance data across the visible (VIS) and near infrared (NIR) region of the electromagnetic spectrum. The prediction power of these indices with respect to green leaf area index (LAI) and canopy chlorophyll density (CCD) was compared, and their sensitivity to canopy architecture, illumination geometry, soil background reflectance, and atmospheric conditions were analyzed. The second soil-adjusted vegetation index (SAVI2) proved to be the best overall choice as a greenness measure. However, it is also shown that the dynamics of the VIs are very different in terms of their sensitivity to the different external factors that affects the spectral reflectance signatures of the various modeled canopies. It is concluded that hyperspectral indices are not necessarily better at predicting LAI and CCD, but that selection of a VI should depend upon (1) which parameter that needs to be estimated (LAI or CCD), (2) the expected range of this parameter, and (3) a priori knowledge of the variation of external parameters affecting the spectral reflectance of the canopy.

Introduction

Spectral reflectance of vegetation in the visible (VIS) region of the electromagnetic spectrum is primarily governed by chlorophyll pigments (Thomas & Gausman, 1977). Developments within the field of hyperspectral remote sensing imaging sensors have allowed for new ways of monitoring plant growth and estimating potential photosynthetic productivity.

Many studies have focused on the relationship between pigment concentration and optical properties of leaves Horler et al., 1983, Jacquemoud et al., 1996, Lichtenthaler et al., 1996. A number of investigators have studied the relationship between canopy spectral reflectance and canopy characteristics for major crops Baret et al., 1987, Gilabert et al., 1996, Jackson & Pinter, 1986. For example, spectral vegetation indices (VIs) calculated as linear combinations of near infrared (NIR) and VIS red reflectance have been found to be well correlated with canopy cover, leaf area index (LAI), and absorbed photosynthetically active radiation (APAR) Elvidge & Chen, 1995, Myneni & Williams, 1994. However, it has been shown that most traditional VIs are sensitive to soil background, especially at low LAIs Huete, 1989, Huete et al., 1985.

The wavelength region located in the VIS–NIR transition has been shown to have a high information content for vegetation spectra Collins, 1978, Horler et al., 1983. The spectral reflectance of vegetation in this region is characterized by very low reflectance in the red part of the spectrum followed by an abrupt increase in reflectance at 700–740 nanometer (nm) wavelengths. This spectral reflectance pattern of vegetation is generally referred to as the “red edge.” Several studies have shown that measures based on the red edge position or shape are likewise well correlated with biophysical parameters at the canopy level, but less sensitive to spectral noise caused by the soil background and by atmospheric effects Baret et al., 1992, Demetriades-Shah et al., 1990, Guyot et al., 1989, Mauser & Bach, 1995.

The objective of the present study is to compare different VIS–NIR spectral reflectance-based approaches for estimation of LAI and canopy chlorophyll density (CCD). As part of this assessment the effects of the soil background and the atmosphere are considered. The analyses are based on simulated canopy spectral reflectance data using acknowledged radiative transfer models in combination with real soil reflectance data.

The study is composed of three phases addressing (1) the effects of structural and biochemical variation in the canopy, (2) the effects of variations in soil background reflectance, and (3) the effects of varying atmospheric conditions. A canopy reflectance database was created for each phase of the study.

Section snippets

Canopy reflectance simulations

Canopy spectral reflectance was simulated using the PROSPECT leaf optical model Baret et al., 1992, Jacquemoud & Baret, 1990 coupled with the Scattering by Arbitrarily Inclined Leaves (SAIL) canopy reflectance model (Verhoef, 1984) modified to include the hot spot effect (Kuusk, 1991). The SAIL model is an analytical, physically based four-stream radiative transfer model that considers the canopy a homogeneous, infinitely extended vegetation layer made up of leaves distributed at random. The

Calculation of VIs

Most commonly used VIs are based on discrete Red and NIR bands, because vegetation exhibits unique reflectance properties in these bands. The early indices are generally divided into ratio indices and orthogonal indices depending upon their nature. Whereas ratio indices are calculated independently of soil reflectance properties, the orthogonal indices refer to a base line specific to the soil background. This soil line is normally defined by the coefficients a and b giving the slope and

Relating LAI and CCD to VIs

Because most VIs, including the REIP, approach a saturation level with increasing green biomass, they can be fitted to an exponential function. A modified version of Beer's law has been suggested Baret & Guyot, 1991, Wiegand et al., 1992 to describe the relationship between a VI and LAI or APAR. This model was adopted in this study to quantify the sensitivity of the calculated indices to solar zenith angle, mean leaf tip angle, and background reflectance. $VI = VI_{∞} +(VI_{g} − VI_{∞}) exp^{(−K_{VI}LAI)} .$ The model

Sensitivity analysis

The sensitivities of the VIs to external factors were analyzed using the relative equivalent noise approach (REN) proposed by (Baret & Guyot, 1991). These authors used the local slope of the exponential function fitted to the data to calculate the standard deviation of the greenness estimate according to the equation: $REN_{LAI} = σ_{LAI} LAI = σ_{VI} LAI d (VI) d (LAI)^{−1} .$ The local slope (Eq. (24)) can be found by differentiation of Eq. (23): $d (VI) d (LAI) =− K_{VI} (VI_{g} −VI_{∞}) exp^{(−K_{VI}LAI)} .$

The advantage of the REN measure is

The REIP

REIP was determined in two fundamentally different ways. The polynomial fitting procedure (Eq. (17)) and the inverted Gaussian model (Eq. (15)) proposed by Bonham-Carter (1988) and Miller et al. (1990) both approximate the spectral shape of the red edge by fitting a function to the spectral data. This method has a built-in smoothing routine because REIP is determined by differentiation of these functions. The Lagrangian technique proposed by Dawson and Curran (1998) forces the interpolation

Conclusion

Classic VIs based on broadband (TM sensor configuration) and narrow band (ideal 10 nm wide bands) reflectance data were compared. It was shown that the broadband indices were less affected by external factors when used as estimators of LAI or canopy chlorophyll content.

The performance of these indices was then compared with the performance of various hyperspectral indices, i.e. recently proposed indices that are based on narrow band reflectance data. The classic broadband VIs generally seem to

Acknowledgements

The authors are indebted to Dr. Craig S.T. Daughtry (ARS-USDA) for facilitating our use of the soil reflectance database, and to Dr. Stephane Jacquemoud (University of Paris) for supplying the code for the SAIL+PROSPECT model.