433
MM
EXTRACTION OF FOLIAR BIOCHEMISTRY FROM HYPERSPECTRAL DATA USING
WAVELET DECOMPOSITION.
G. A. Blackburn 3 ’ *, J. G. Ferwerda b
a Department of Geography, Lancaster University, Lancaster, LAI 4YB, UK - alan.blackbum@lancaster.ac.uk
b Department of Zoology, Oxford University, Oxford, 0X1 3PS, UK -jelle@bio-vision.nl
KEY WORDS: Leaf, Biochemicals, Water content, Reflectance, Hyperspectral, Wavelet transform.
ABSTRACT:
This study explores the potential of wavelet decomposition of leaf reflectance spectra for quantifying foliar biochemicals and water.
A leaf-scale radiative transfer model was used to generate a very large spectral data set with which to develop and rigorously test the
technique. The size of the data set enabled a thorough statistical analysis of the performance a range of alternative methods for
constructing predictive models including the selection of specific wavelet functions, continuous or discrete transforms, reflectance or
derivative input spectra and number of wavelet coefficients used as predictors. The results demonstrated that wavelet decomposition
techniques can generate accurate predictions of protein, lignin/cellulose and water content, despite wide variations in of all of the
biochemical and biophysical factors that influence leaf reflectance. Wavelet analysis outperformed predictive models based on
untransformed spectra and enabled the greatest improvements in performance for protein followed by lignin/cellulose then water
content. Hence, the study highlights the capabilities of wavelet decomposition for extracting information concerning leaf
components that have narrow, weak absorption features, which are otherwise difficult to characterise in untransformed reflectance
spectra.
1. INTRODUCTION
The spatial and temporal dynamics of foliar biochemicals and
water content are intimately linked to the interactions between
plants and the environment. Quantifying such dynamics can
provide important evidence concerning the functioning of
ecosystems and the physiological status of crops. It is with
variable success that hyperspectral remote sensing has been
applied to quantifying foliar biochemistry and water content
and there is considerable scope for developing more accurate
and robust analytical techniques. One particular approach,
developed by the authors, that has proven successful in
quantifying chlorophyll concentrations is based on the wavelet
decomposition of reflectance spectra (Blackburn and Ferwerda,
2008).
Wavelet decomposition is used to separate complex signals into
their basic component signals. Comprehensive descriptions of
wavelet analysis are provided by Graps (1995) and Strang and
Nguyen (1996). In the context of leaf reflectance spectra, the
component signals result from the various biochemical and
biophysical properties that control the optical properties of
leaves. Some of these properties such as leaf structure, affect
broad regions of the reflectance spectrum, while other
properties such as particular biochemicals can affect specific
narrow wavebands. Hence a technique which is able to
discriminate between effects at different spectral scales and
identify wavelength positions that contain the greatest
information would be valuable in a quantitative remote sensing
context. Wavelet analysis is able to decompose signals
according to both frequency and time domains, meaning that
the wavelet coefficients that result contain information
concerning the scale and position of component signals. This
capability has proven useful in extracting information
concerning chlorophyll concentrations from leaf reflectance
spectra despite large variations in the range of other controlling
factors. The aim of the present study is to determine the
suitability of wavelet-based methods for quantifying other foliar
biochemicals and water.
2. METHODS
The experimental strategy adopted was to generate a data set
that incorporated leaves with a wide range in biochemical and
water content and encompassed a large variability in all of the
other factors that influence leaf reflectance spectra. Our
previous experience in collecting leaf reflectance spectra and
measuring leaf biophysical and biochemical properties
indicated that the time and resources required to physically
collect the large data set needed for this project would be
prohibitive. Therefore, leaf radiative transfer (RT) modelling
was used to generate spectral data sets based on a very wide
range and combinations of leaf properties that influence
reflectance.
The LIBERTY model adapts RT theory for determining the
optical properties of powders for describing light interaction
within cells which are roughly spherical and separated by air
spaces (Dawson et al., 1998). LIBERTY simulates conifer
needle reflectance over the wavelengths 400 to 2500nm using
the input parameters: chlorophyll content, water content,
lignin/cellulose content, protein content, average cell diameter,
intercellular air space, leaf thickness, baseline absorption and
albino leaf absorption. The parameter described in this paper as
protein content is termed nitrogen content in the original model
description but this is actually based on the absorption spectrum
of protein and does not incorporate absorption of other
significant nitrogen containing components such as chlorophyll.
LIBERTY was used to simulate 1100 spectra and for each
model run the value of each input parameter was selected
Corresponding author.