38 
  
THE USE OF HYPERSPECTRAL DATA FOR PRECISION FARMING 
K. Staenz, J.-C. Deguise, J.M. Chen, H. McNairn, and R.J. Brown 
Canada Centre for Remote Sensing, Ottawa Ontario, Canada 
T. Szeredi 
MacDonald Dettwiler and Associates, Richmond, British Columbia, Canada 
M. McGovern 
Agriculture Canada, Ottawa, Ontario, Canada 
Commission VII, Working Group I 
KEY WORDS: Leaf area index, precision farming, 
hyperspectral casi data cube, surface reflectance, spectral 
unmixing, crop fraction. 
ABSTRACT: A technique is proposed to estimate leaf 
area index (LAI) using the crop endmember fraction derived 
with spectral unmixing. 96-band hyperspectral data acquired 
with the Compact Airborne Spectrographic Imager (casi) over 
two agricultural sites were used to test the approach against 
measured LAI data and LAI computed with a NDVI-based 
semi-empirical model. For this purpose, the radiance data 
were converted to surface reflectance prior to the extraction of 
the crop fractions which were retrieved with constrained 
linear unmixing. The preliminary validation tests indicate that 
the proposed technique has potential to estimate the effective 
LAI from the crop fraction. This technique has the advantage, 
compared to other approaches, of separating the crop from 
unwanted portions of vegetation such as weeds. This should 
lead to a more accurate estimation of LAI. 
1.0 INTRODUCTION 
Image-based remote sensing can play a significant role in 
precision farming (Moran et al., 1997), providing information 
for crop management on a within-field basis. With the advent 
of the imaging spectrometer, a fundamental new remotely 
sensed data set became available for this purpose (Staenz 
1992). The high spectral dimensionality of such data enables 
the extraction of quantitative information never before 
possible with broad-band imaging sensors. In particular, 
hyperspectral data can be used to improve the detection of 
within-field variability with respect to crop production, to 
determine the cause of within-field spatial variability, and to 
pararmeterize and validate crop models (Moran et al., 1995; 
Carter, 1994; Maas, 1993). 
In this paper, the potential of hyperspectral Compact Airborne 
Spectrographic Imager (casi) data has been evaluated for the 
detection of spatial variability on a field basis for cash crops 
near Altona, Manitoba (Canada). Special emphasis was 
devoted to the extraction of the green leaf area index (LAI). 
LAI is a fundamental crop parameter that provides a valuable 
source for crop growth modeling (Moran et al, 1995; 
Bouman, 1992; Bauer, 1985; Wiegand et al, 1986). Since 
LAI is functionally related to spectral reflectance, a variety of 
techniques have been developed using remotely sensed data. 
Vegetation indices are involved especially for the estimation 
of LAI (Baret and Guyot, 1991; Wiegand and Richardson, 
1990; Clevers, 1989). One of the disadvantages of the use of 
vegetation indices is the fact that these indices are sensitive to 
the total amount of vegetation cover within a pixel without 
distinguishing between crop, weeds, and other vegetation 
components. À new technique that only takes the crop portion 
of the vegetation into account for estimating the LAI is 
presented in this paper. 
2.0 DATA ACQUISITION 
casi data were acquired over agricultural test sites near Altona 
and Birtle, Manitoba on July 25, 1996 during maximum 
vegetation growth. The flat test sites had a variety of cash 
crops, such as cereal grains, canola, sugar beets, and beans. A 
typical field size is 400 m by 800 m. 
The casi data sets were collected in the wavelength range 
from 458 nm to 1000 nm in 96 contiguous, 6.8 nm wide 
spectral bands, sampled at 5.8 nm intervals (Anger et al., 
1996). In this data acquisition configuration, the swath 
consists of 304 pixels with a ground resolution of 4 m across 
and 4 m along track at a flight altitude of 2745 m above sea 
level. 
Ground reference information relevant for this study includes 
crop type, LAI, and biomass. Biomass samples were 
collected from 8 to 12, 0.5 m by 0.5 m plots along a diagonal 
transect of selected fields with sample plots approximately 50 
m to 60 m apart. LAI was estimated from the leaf material 
collected from the biomass sample by dividing the total leaf 
area by the sample plot (unit) area. 
3.0 DATA PREPROCESSING 
The preprocessing of casi data included the removal of noise, 
most significant aircraft motion effects, surface reflectance 
retrieval, and post-processing of the retrieved surface 
reflectance spectra (Figure 1). 
Noise (non-periodic horizontal striping), which especially 
affects bands at the two extremes of the casi wavelength 
coverage, was removed in the principal component (PC) 
domain using the last 89 of 96 PC images. The average of 
each line per PC image was calculated and subsequently 
plotted against the line number. A Gaussian smoothing with a 
window of 15 lines was then applied to the data. This enabled 
the computation of a correction factor (gain) for each line to 
adjust the original mean values to the smoothed ones. The 
gain was calculated by dividing the smoothed values by the 
original line means. In a final step, the entire 96-PC image set 
was inversely transformed back to the spectral band domain. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998