most-
-111
| were
ridden
rences
utputs
twork
til the
ual to
rk are
NETS
chines
ovides
ty of
arning
ration.
rksta-
were
image
based
L We
1596,
s with
'ained
aining
lamed
study
cover
infor-
aerial
vidual
were
ering)
| con-
| mar-
ie the
971),
onfu-
of the
id the
ency
tional
' and
ndard
) was
ibles.
itting
'ency
ounts
2.5 Evaluation of Classifications
Multiple comparisons were made to evaluate the four
classifiers, including NN-5%, NN-10%, NN-15%, and
NN-20%. By extracting the correct percentages of each
classification category from Tables 1 through 4, we pro-
duced a performance summary table of classifiers (Table
5). The Tukey multiple comparison method was used
for the evaluation of these four classifiers. Any two
population means of classifiers will be judged to be dif-
ferent from each other if the difference of the
corresponding sample means is greater than the Tukey
distance (Mendenhall and Sincich, 1989). The Tukey
multiple comparison method is supported by SAS
software (SAS Institute, 1988b).
3. RESULTS
The results of the Tukey multiple comparisons (Table 6)
provided the overall classification accuracies for the
classifiers of NN-5%, NN-10%, NN-15%, and NN-20%.
At a risk level of 5%, the results showed that (a)
classifiers NN-5%, NN-10%, and NN-15% did not differ
from one another, (b) classifiers NN-15% and NN-20%
did not differ from each other, but (c) classifiers NN-596
and NN-10% differed from classifier NN-20%. The
training rates of the neural network classifiers are illus-
trated in Figure 1.
4. DISCUSSION
As shown in table 8, we could train the neural network
with either 5%, 10%, or 15% TM data because the sta-
tistical evaluation showed no significant differences
among the three corresponding classifiers at a 5% risk
level. The evaluation was done based on individual
category accuracies highlighted in Tables 1 through 4.
However, interpretations of classified images (Figure 2)
show that the two classification results of NN-10% and
NN-15% were more uniform than the result of NN-5%.
We could not interpret that the classification result of
NN-20% differed from the results of NN-10% and NN-
15%.
As we increased the percentage of the TM data for train-
ing, the corresponding training rate also increased (Fig-
ure 1). When we increased 5% to 10% for training, the
training rate was decreased by 2 seconds/cycle. When
we increased 10% to 15% or 20%, the training rate was
reduced by 11 or 16 seconds/cycle, respectively. In this
project, the training periods ranged from 100 to 200
cycles.
CONCLUSIONS
Considering the evaluation results and the training rates,
we recommand using around 10% TM data to train a
neural network, and the performance of a neural net-
work classifier is satisfactory for this level of training.
531
5. ACKNOWLEDGEMENTS
This research was supported by NASA Research Grant
NAGW-1472 and the Laboratory for Applications of
Remote Sensing at Purdue University.
6. REFERENCE
Baffes, P.T., 1989. Nets User's Manual. Version 2.0.
AIS at NASA/JSC, Athens, Geogia, U.S.A., 76 p.
Benediktsson J.A., P.H. Swain, and O.K. Ersoy, 1990.
Neural network approaches versus statistical
methods in classification of multisource remote
sensing data. IEEE Trans. Geoscience and Remote
Sensing, GE-28(4):540-552.
Fienberg, S.E., 1971. A statistical technique for histori-
ans: standardizing tables of counts. Journal of
Interdisciplinary History, Vol. 1, pp. 305-315.
Fienberg, S.E., and P.W. Holland, 1970. Methods for
Eliminating Zero Counts in the Contingency
Tables. Random Counts in Scientific Work (G.P.
Patil, editor). Pennsylvania State University Press,
University Park, Pennsylvania, U.S.A., Vol. 1, pp.
233-260.
Hepner, G.F., T. Logan, N. Ritter, and N. Bryant, 1990.
Artificial neural network classification using a
minimal training set: comparison to conventional
supervised classification. Photogrammetric
Engineering and Remote Sensing, 56(4):496-473.
Mendenhall, W., and T. Sincich, 1989. A Second Course
in Business Statistics: Regression Analysis. Dellen
Publishing Company, San Francisco, California,
U.S.A., 864 p.
SAS Institute, 1988a. SAS/IML User's Guide, Release
6.03 Edition. SAS Institute Inc., Cary, North Caro-
lina, U.S.A., 357 p.
SAS Institute, 1988b. SAS/STAT user's Guide, Release
6.03 Edition. SAS Institute Inc., Cary, North Caro-
lina, U.S.A., 549 p.
Zhuang, X., B.A. Engel, M.F. Baumgardner, and P.H.
Swain, 1991. Improving Classification of Crop
Residues Using Digital Land Ownership Data and
LANDSAT TM Imagery. Photogrammetry
Engineering and Remote Sensing, 57(11):1487-
1492.
Zhuang, X., 1990. Determining Crop Residue Type and
Class Using Satellite Acquired Data, M.S.E.
Thesis. Department of Agricultural Engineering,
Purdue University, West Lafayette, Indiana,
U.S.A., 129p.