For comparison, a second eight-band data set was 
generated from the spectral data using band 
averaging to match the PMI spatial band set #3 
(GEOBOTANY) given in Table 2. The GEOBOTANY band 
set #3 is well positioned for crop discrimination. 
The same feature selection criteria were applied 
to this new data set for the same classes. The 
best four-band subset was (1,4,5,6); that is, 
491.42 nm, 675.75 nm, 708.26 nm, and 734.22 nm. 
The maximum likelihood classification results for 
each band combination selected by feature 
selection are given in Table 3. 
The classification results are also summarized in 
Figure 6 for the two different data sets. It can 
be seen that the weighted mean classification 
accuracy using the GEOBOTANY data set varies 
between 57.03 ± 10.59 (s.e.m.)% for the best one- 
band subset (band 4, 675.75 nm) and 98.59 ± 0.45% 
for the eight-band combination. For the band 
moments the classification accuracy varied between 
61.38 ± 12.49% for band 6 (skewness) and 99.41 ± 
0.27% for all bands. With band moments, an 
excellent result of 94.86 ± 1.74% was already 
achieved with the three-band subset (1,7,8) where 
(1) is the mean, (7) the kurtosis, and (8) the 
band-concentrated moment. A similar accuracy 
level was reached with the GEOBOTANY four-band 
subset (1,4,5,6) leading to a result of 94.54 ± 
1.50%. The band-moments method produces slightly 
higher accuracies than the GEOBOTANY band set by 
a margin up to 4% depending upon the different 
band combinations. The standard error of the mean 
classification accuracy is generally lower for the 
band moments indicating slightly less confusion 
between classes, especially between barley, wheat, 
and fallow. Confusion between these three targets 
was resolved for the four-band subset using the 
band moments while for the corresponding GEOBOTANY 
subset a 5% confusion with wheat and fallow 
remains. These classification results are 
promising for further application of the band- 
moments analysis concept to other imaging spectro 
meter data sets. 
CONCLUSIONS 
This paper describes the methodologies necessary 
to bring the PMI data to an acceptable data 
quality level for classification purposes, as well 
as preliminary classification results. 
Preprocessing steps include assessment and 
correction techniques for bad data in the spectral 
and spatial domains and radiometric alignment of 
the five different PMI cameras. The camera 
adjustment procedure involves a polynomial 
regression approach which removes part of the 
viewing angle effect. This effect is quite 
substantial due to the large PMI scan angle range 
of 75.5 degrees. Data variations within a 
specific camera (a field of view of 14.5 degrees) 
caused by the viewing angle and other sensor 
related effects were encountered, especially in 
camera 4. Future software development will take 
this phenomenon into account. Further data 
assessments with respect to noise and effective 
wavelength position of the sensor indicate satis 
factory results. The random noise shows, for 
example, a signal-to-noise ratio of 600:1 for an 
average signal level of 803.4 x 10 B W cm -2 nm 1 
sr' 1 in band 247 (751.21 nm) for the different 
cameras. The radiometric calibration data for 
normalization purposes was found to be suspect 
because of noise and the (erroneous) magnitude of 
the calibration values. 
Despite the problems encountered during the data 
evaluation, preliminary classification results 
involving data reduction techniques followed by 
maximum likelihood classification are encouraging. 
Band-moment analysis in the spectral domain and 
band averaging to match the PMI spatial band set 
GEOBOTANY were used to create two eight-band data 
sets. Feature selection, with a branch and bound 
algorithm, was applied to these data in order to 
establish the best band subsets for the eight 
classes considered. Similar classification 
accuracies were achieved with both methods for the 
agricultural data set considered. The best four- 
band subset, consisting of the moments mean (1), 
skewness (6), kurtosis (7), and band-concentrated 
moment (8), resulted in an overall accuracy of 97% 
compared to 95% for the best four-band GEOBOTANY 
subset (1,4,5, and 6). Band-moment analysis is an 
automated procedure for data reduction in order to 
achieve similar classification results as with a 
carefully selected band set derived from a 
detailed spectroscopic analysis. Further inves 
tigations will be concentrated on the application 
of the band-moment analysis concept to other 
imaging spectrometer data sets (e.g. AVIRIS), 
multispectral imagers, and ground-based 
spectrometer data involving not only agricultural 
objects but also different forest types. 
ACKNOWLEDGEMENTS 
The authors would like to thank Dr. R. Buxton, L. 
H. Gray, and R. Gordon from Moniteq Ltd. for their 
technical advice with respect to PMI, as well as 
Dr. K.I. Itten and P. Meyer from the University of 
Zurich-Irchel for their support during the data 
acquisition. The valuable discussions with Dr. P. 
M. Teillet and the assistance of B. Thibault and 
A. Boisvert in analyzing the data are gratefully 
acknowledged. Our thanks go to A. Kalil for the 
excellent typing of this manuscript. 
REFERENCES 
Borstad, G.A., Edel, H.R., Gower, J.F.R. and 
Hollinger, A.B., 1985. Analysis of Test and 
Flight Data from the Fluorescence Line Imager. 
Canadian Special Publication of Fisheries and 
Aquatic Sciences 83, Department of Fisheries and 
Oceans, Ottawa, Ontario, Canada, 38 pages. 
Brown, R.J., Ahern, F.J., Ryerson, R.A., Thomson, 
K.P.B., Goodenough, D.G., McCormick, J.A. and 
Teillet, P.M., 1980. Rapeseed: Guidelines for 
Operational Monitoring. Proceedings of the 6th 
Canadian Symposium on Remote Sensing, Canadian 
Aeronautics and Space Institute, Ottawa, Ontario, 
Canada, pp.321-333. 
Buechel, S.W., Philipson, W.R. and Philpot, W.D., 
1989. The Effects of a Complex Environment on 
Crop Separability with Landsat TM. Remote Sensing 
of Environment, Vol.27, pp.261-272. 
Curran, P.J. and Dungan, J.L., 1988. Estimating 
the Signal-to-Noise Ratio of AVIRIS Data. NASA 
Technical Memorandum 101035, NASA/AMES Research 
Center, Moffet Field, California, 21 pages. 
Goodenough, D.G., Narendra, P.M. and O'Neill, K., 
1978. Feature Subset Selection in Remote Sensing. 
Canadian Journal of Remote Sensing, Vol.4, No.2, 
pp.143-148. 
Hlavka, C., 1986. Destriping AIS Data Using 
Fourier Filtering Techniques. Proceedings of the 
2nd Airborne Imaging Spectrometer Data Analysis 
Workshop, JPL 86-35, Pasadena, California, pp.74-