the area 
th about 
rea data 
nent was 
for both 
The 
with the 
on of the 
he spray 
rm rate 
yortion is 
ventional 
ninus the 
/eed map 
ol zones 
different 
analysis 
ffect the 
  
>10 
100 
from the 
tion map, 
. Table 1 
ind based 
orrelation 
ta. The 
generated 
he spatial 
n remote 
IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring", Hyderabad, India, 2002 
  
sensing data equivalents to the weed map data resolution (0.76 
m/pixel), the correlation between image indices and weed density 
map was low. The correlation increased with the decrease in 
resolution up to 3.8-m/pixel, stayed stable from 3.8 to 4.5-m/pixel 
and then decreased (Figure 6).  Theoretically, the correlation 
should be highest at a resolution the same as that of the image and 
weed data. The correlation should decrease at lower resolutions 
due to the error of averaging. It is believed that the geo- 
referencing error along with the low horizontal positioning 
accuracy of the GPS systems compared to the high resolution of 
weed and the aerial image data had contributed to overlay errors 
at higher resolutions. However, the comparatively high 
correlation of 0.9 (for NDVI) at lower resolutions of 4-m/pixel to 
4.5-m/pixel proved that the error in geographic location of data 
points was averaged out at these lower resolutions. It would be 
possible to develop a fairly accurate weed map from an aerial 
image of a field under similar conditions at this resolution using 
GPS with sub-meter accuracy. 
Table 3. Herbicide savings over the uniform application method. 
Very high weed Normal weed density 
  
density plot plot (STD=0.05) 
(STD*=0.18) 
Single 6% 52% 
threshold 
method 
Variable 18% 71% 
rate 
method 
  
_* Standard deviation of weed density 
  
Figure 3a. Example original color near-infrared image 
of a test plot at a resolution of 1meter per pixel (Soybean 
field, 7-23-1998). 
The variation of correlation between weed density and image 
indices at different resolutions indicated that there existed a best 
resolution for mapping spatial variation in weed density. This 
best resolution could be slightly different for different fields 
depending mainly on the various errors involved in the data and 
other factors such as soil characteristics that influence the 
variability in weed distribution. 
A CIR image obtained through remote sensing of the field was 
classified using unsupervised classification and overlaid with a 
ground truth map of weed density for visual comparison (Figure 
7). The classified image enhanced the spatial patterns in the 
original CIR image by painting them with distinct false colors. 
The patterns observed in the image corresponded to both soil and 
vegetation characteristics. The spatial variation in green 
vegetation was mainly due to the various herbicide treatments 
within the field. On visual analysis of the classified image 
overlaid with the weed map (Figure 4), it was found that the 
variation in weed density corresponded to the variation in the 
total vegetation density within the field. The vegetation density 
varied due to different herbicide applications and planting dates, 
and under different soil types. Image calibration eliminated the 
effect of soil background and planting date, and hence the visible 
spatial pattern in the calibrated image was assumed to be due to 
the difference in vegetation. 
Based on the weed detection from the images in this randomly 
sampled data set, the weed distributions were best approximated 
by the negative binomial, which is coincident with some other 
weed distribution research (Cardina, et al., 1997). Figure 8 
depicts the weed density frequency in the experimental field. In 
the experimental field, more than 80 percent of the control zones 
had less than 20 percent weed coverage. 
   
Figure 3b. Image taken on the same day with higher 
spatial resolution aerial imaging system (300 mm/pixel).