The back propagation-training algorithm is an iterative gradient
algorithm designed to minimize the mean square error between the
actual output of a multi layer feed-forward perception and the
desired output. The method involves modifying the weight values
between nodes including the weights for the hidden layers. The
algorithm is based on the Delta rule, which propagates the error
between the actual output and the desired output backward
through the neural network. The Delta rule attempts to minimize
the error using the steepest descent, similar to search technique
known as hill climbing (Penn et al., 1993). Back propagation
requires continuous differentiable non-linearities.
7. RESULTS
The results obtained from the sequential and composite data
classification using the two algorithms were evaluated on the basis
of two parameters namely, (1) classified paddy area expressed as
% of reference area and (2) accuracy of classification. The results
of sequential classification are shown in Table 4, which indicates
that the total estimated paddy area using ANN is 53% of reference
estimate, where as that of MLC is only 38%. However, the
accuracy of the estimate is 89% for ANN and 92% for MLC.
Thus, the performance of ANN is relatively better than MLC,
though in absolute sense, both the classifiers have failed to
achieve complete classification.
The results of the composite data classification (Table 5) indicate
that the estimated paddy area is significantly higher than that of
sequential classification. The paddy area classified by the two
algorithms in composite data classification, indicate that ANN has
classified more area at 74% of the reference estimate, where as
MLC has classified the paddy area that forms only 57% of
reference estimate. The accuracy of the estimate is 89.14% for
MLC and 89.84% with ANN. Thus, ANN classification has
delineated more paddy area, compared to MLC without
compromising with the accuracy of classification.
Table 4 Sequential classification - comparison of estimates
Artificial Neural
Networks Maximum Likelihood
% of % of
Est area [Reference| Acc. | Est.area |Reference Acc..
(ha) area % (ha) area %
February
paddy | 36792 33 92 | 28998 26 94
March
paddy | 23034 20 86 | 14032 12 88
Total
paddy | 59827 53 89 | 43030 38 92
Table 5 Composite data classification - comparison of estimates
Classifier Estimated % of Accuracy
paddy area (ha) Reference (%)
estimate
ANN 82875 74 89.14
MLC 63623 57 89.84
8. CONCLUSIONS :
The results of the study have shown conclusive evidence that with
a given set of training areas, the ANN has resulted in superior
classification in terms of area classified and accuracy of
IAPRS & SIS, Vol.34, Part 7, "Resource and Environmental Monitoring", Hyderabad, India,2002
classification compared to Maximum Likelihood. Therefore,
Artificial Neural Networks has the potential for improving crop
classifications using multi-temporal remote sensing images. The
study also indicates that ANN classification process can be
implemented with no extra manual efforts compared to traditional
MLC. However, the extra computational efforts involved in ANN
would not be a major limitation, since powerful computational
facilities are now available.
9. ACKNOWLEDGMENTS
Grateful thanks are due to Dr. R.R. Navalgund, Director and Dr.
A. Bhattacharya, Deputy Director (Remote Sensing & GIS),
National Remote Sensing Agency, Hyderabad, India, for
according permission and for providing required facilities for
successful completion of this study. Without the help and
cooperation from colleagues in the Water Resources Group, in
NRSA, this study could not have been successfully completed.
10. REFERENCES
Badhwar, G. D., 1984, Classification of corn and soybean using
multi temporal thematic mapper data, Remote Sensing of
Environment, 16, pp.175-182.
Benediktsson, A. Jon, Philip, H. Swain and Okam, K. Ersoy,
1990, Neural Network approaches versus statistical methods in
classification of multi source remote sensing data,
LE.E.E.Transactions on Geoscience and Remote Sensing 28,
pp.540-551.
Conese Claudio, and Maselli Fabio, 1991, Use of multi temporal
information to improve classification performance of TM scenes
in complex terrain, ISPRS Journal of Photogrammetry and
Remote Sensing, 46, pp.187-197.
Hlavaka, C.A., Haralick, R.M., Carlyle, S.M. and Yokoma, R.,
1980, The discrimination of winter wheat using a-growth state
signature, Remote Sensing of Environment, 9, pp.277-294.
King, T., 1989, Using neural networks for pattern recognition, Dr.
Dobb's Journal, January, pp.14-18.
Lambin, F. Eric. and Strahler, H. Alan., 1994, Change vector
analysis in multi temporal space; A tool to detect and categorize
land cover change processes using high temporal resolution
satellite data, Remote Sensing of Environment, 48, pp.231-244.
Lippman, R.P. 1987. An introduction to computing with neural
nets, IEEE ASSP May, 4, pp.4-22.
Liu, Xuehua, Skidmore, K. Andrew, and Oosten van Henk, 2000,
Discrimination ability of Neural Network and Maximum
Likelihood Classifiers, In XIX, ISPRS: IAPRS, vol.XXXIII,
Amsterdam, The Netherlands, 16-23 July 2000.
Lo H. C. Thomas, 1986, Use of multi temporal spectral profiles in
Agricultural land cover classification, Photogrammetric
Engineering and Remote Sensing, 52, pp.535-544.
Pax-Lenney and Wood Cock, E. Curtis, 1997, Monitoring
Agricultural land in Egypt with multi temporal Landsat TM
imagery; How many images are needed?, Remote Sensing of
Environment, 59, pp.522-529.
un
N= N A