Summary 
Learning Algorithm 
(1) Initialize weight vectors (Wi, 
i-1,2..— kl) for the k output neurons to 
either small random values or small 
uniform values because sample- 
dependent initialization avoids many 
pathologies that can distort nearest- 
neighbor learning. 
Competitive 
(2) Present random input vector, 
Y. Find the closest or "winning" weight 
vector (Wj) by using excitation function 
E. 
Ej =E (| Y - Wj) = min | Wi- Y| 
I 
where || Y ||2 = Yı2+Y22 + Y32 + 
Yn?, which defines the squared 
Euclidean vector norm of Y. 
(3) Update the winning weight 
vector (Wj*) by the following learning 
rule: 
Wj* 2 Wj*n(Y-Wj) 
where n is an learning rate, 
and 0 «mn«1. 
(4) Present next random input 
vector. 
Typical a set of random input 
vectors will cycle through the network a 
number of times until a stable clustering 
gas evolved. Besides, in the majority of 
cases the distance Ej are calculated using 
the Euclidean distance function although 
frequently the Manhattan ("city block") 
or some kind of statistical distances 
(variance) are used as well. 
3. Data Source and Experimental 
777 
PRE D EE SI EEE PE ASIE EI 
In this research, Landsat-7 
Thematic Mapper (TM) image is used as 
input to a competitive learning neural 
network. The TM 7-band imagery of the 
region was obtained in Columbus, Ohio, 
1988 (Dr. Merry, personal 
communication, 1995). The network is 
unsupervised training to associate the 
spectral data of each pixel with one of 
10, 15, and 30 possible land cover 
categories. There are a total of 270,000 
pixels (with band 3, 4, 5) and 630,000 
pixels (from band 1 to 7) of testing 
datasets. These pixels are contained 
within a 300 x 300 pixel region within 
the area which is about 5 miles north of 
Columbus, Ohio. The ground truths of 
land cover data (road and water) for the 
region were obtained by topographic 
map (from Engineering Department of 
Franklin County, Ohio) at scale about 
1:4748 that was cover the area on 1987. 
Compared using TM bands 3, 4, 
and 5 with TM bands 1-7, the visual 
viewing results are not great difference. 
However, the final performance of the 
CL is measured by the proportion of the 
total pixels assigned to the correct land 
cover category, and the overall 
percentage ^ correctly characterized 
(PCC). The PCC for each category type 
is calculated. A minimum classification 
accuracy of 8596 has been suggested for 
remote sensing data. This level of 
performance is not expected at this 
research stage. The results are far below 
than suggested accuracy, and one of the 
output results is in the Appendix. 
4. Discussion and Conclusion e 
AS" noted" by Grossberg, 
Rumelhart and Zipser, and among 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996 
EEE ENTE EEE a ———— NM a