A Study of Neural Network Classification of Jers-1/Ops Images
4 Network and parameter selection
We select three-layer MLP model using the back-propagation (BP) algorithm in this study. The number of input nodes is specified by the dimension of the input patterns. For the JERS-1/OPS image data, the input pattern is from one pixel consists of three bands at a precision of 8-b/band to give 24-b per input pattern.
Similarly, the output nodes are determined by the number of categories to be classified or the desired output mapping. In this paper, we classify four categories. Therefore, we have four bits per output pattern. Each bit of the desired output Tpi in eq. (4) presents state 1 or state 0, state 1 for "belongs to" and state 0 for "not belongs to" a category (eg., 1 0 0 0 is the desired output pattern for "belongs to" category 1).
The number of hidden nodes usually define at least as number of nodes in the input layer [5]. Based on Kolmogorov theory [6], 2N+1 hidden nodes should be used for one hidden layer (where N is number of input nodes). For 24 input nodes, we will have 49 hidden nodes. The learning factor (h) and the momentum factor (a) in eq. (5) are set to 0.01 and 0.9, respectively. The summed squared error (SSE) in eq. (4) is set to 0.003. In tble 1, the number of pixels in each category which we selected to train the neural network is shown.
5 Result
We designed MATLAB program to train this network on Pentium pro 200 workstation (64 MG RAM), and used 2 " days for training time. After training process, 2,334 pixels in table 1 was selected to test the correct classification percent of two methods (the neural network method and the maximum likelihood method (The results are shown in Table 3 and Table 4, respectively.
As is possible to see by reading the table, the neural network method seem to work a little better, The overall correct classification for the neural network method is about 92.8 percent, and the 85.6 percent value for the maximum likelihood method. This situation is explained by a better attribution in the neural network case of water, urban and vegetation categories, whereas the maximum likelihood method make less errors than the neural network method in one case which is bare soil category.
Table 2. Result of the neural network classification.
| True Category |
Clasified as |
Correct (%) |
| Water |
Urban |
Vegetation |
Bare Soil |
| Water |
953 |
4 |
8 |
58 |
95.5 |
| Urban |
22 |
465 |
0 |
0 |
98.9 |
| Vegetation |
5 |
0 |
563 |
1 |
72.2 |
| Bare Soil |
63 |
0 |
8 |
184 |
72.8 |
| Overall accuracy |
92.8 |
Table 3. Result of the maximum likelihood classification.
| True Category |
Clasified as |
Correct (%) |
| Water |
Urban |
Vegetation |
Bare Soil |
| WAter |
908 |
0 |
21 |
94 |
88.8 |
| Urban |
0 |
395 |
0 |
92 |
81.1 |
| Vegetation |
0 |
0 |
443 |
126 |
77.9 |
| Bare Soil |
0 |
0 |
3 |
252 |
98.8 |
| Overall accuracy |
85.6 |
6 Conclusion
The MLP neural network model using the BP algorithm for classification of JERS-1/OPS image data was simulated and processed. We found that, it is easily modified to accommodate more channels or to include spatial and temporal information. The input layer of the network can simply be expanded to accept the additional data. Although it is slow to train, but it is fast in the classification state.
Acknowledgement
The authors with to thank the National Research Council of Thailand (NRCT) for providing the satellite image data.
Reference
-
J.L. McCelland and D. E. Rumelhar, Eds., Parallel Distributed Processing, Vol. 1.
Cambridge, MA : MIT Press, 1986.
- Y.H. Pao, Adaptive pattern Recognition and Neural Network, Addition-Wesley Publishing Company, Inc., 1989.
- J.A. Bendiktsson, P.H. Swain and O.K. Ersoy, "Neural Network Approaches Versus Statistical Methods on Classification of Multisource Remote Sensing Data", IEEE Trans. Geosci. Remote Sensing, vol. 28, pp. 540-552, July 1990.
- H. Bichof, W. Schneider and A.J. Pinz, "Multishpectral Classification of Landsat Image using Neural Networks", IEEE Trans. Geosci. Remote Sensing. Vol. 30, pp. 482-480, May 1992.
- P.D. Heermann and N.Khozenie, "Classification of Multispectral Remote Sensing Data Using a Back-propagation Neural Network", IEEE Trans. Geosci, Remote Sensing, vol. 30, pp. 81-88; Jan. 1992.
- A.J. Annema, Feed-Forward Neural Network : Vector Decomposition Analysis Modeling and Analog Implementation, Kluwer Academic Publishers , 1995.
