nbul 2004 
ssification 
is slightly 
he object- 
inal pixel- 
Gaussian 
d with the 
ction the 
ed using 
which we 
uilt upon 
re, ES 
' by using 
re (notice 
ication of 
d 2 show, 
eat field 
ed for its 
thesis and 
nal based 
cted pixel 
by a pixel 
1agery we 
'scription, 
eloped an 
im (called 
y relation 
mong the 
adjacency 
stics in an 
quentially 
pixel and 
pace) are 
ppropriate 
presented 
ed to real 
e object- 
inal pixel- 
(overall 
ance and 
mparative 
e object- 
volume is 
ure-space 
e than 25 
uracy of 
asured by 
using the 
nage data 
ral-spatial 
pixels is 
In object 
onsidered 
ected that 
ct-feature 
ent of the 
oration of 
1 decision 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
rule; however, in addition to that, it could be also a 
consequence of complexity reduction by data compaction 
(Hughes, 1968). 
Since the classification process is performed in the feature- 
space (rather than in the observation-space) the algorithm is 
much faster than conventional ones. The object appearance in 
the feature-map can be incorporated (by visual assessment) into 
the feature selection strategy for extraction of more complex 
objects in the scene. In summary, it appears that the proposed 
object-feature extraction process has several advantages over 
most of the conventional techniques. 
REFERENCES 
Ghassemian, H. and D. Landgrebe. 1987. An Unsupervised 
Feature Extraction Method for High Dimensional Image Data. 
IEEE Proc. on System, Man and Cybernetics, vol.2, pp.540- 
544, Oct. 1987. 
Ghassemian, H. and D. Landgrebe, 1988. On-Line Object 
Feature Extraction for Multispectral Scene Representation. 
NASA_TR_EE 88-34, Aug. 1988. 
Ghassemian, H. and D. Landgrebe, 2001. Multispectral Image 
Compression by an On-Board Scene Segmentation. Proc. Of 
IEEE Int. Geoscience and Remote Sensing 2001 Symposium. 
Scanning the Present and Resolving the Future, July 2001. 
Ghassemian, Hassan, 1990. Adaptive Feature Extraction for 
Multispectral Image Data Representation. /A4STED Control and 
Modelling Conf. pp.277-282, July 1990. 
Hapke, B., 1993. Theory of Reflectance and Emittsnce 
Spectroscopy. Cambridge, U.K.: Cambridge Univ. Press, 1993. 
Hughes, G., 1968. On the mean accuracy of statistical pattern 
recognizers. [EEE Trans. Information Theory, vol. IT-14, no. 
l, pp. 55-63, 1968. 
Kettig, L. and D. Landgrebe, 2001. Classification of Remotely 
Sensed Multispectral Image Data by Extraction and 
Classification of Homogeneous Objects. /EEE Transactions on 
Geoscience and Remote Sensing, Vol. GE-39, No.1, pp.4-16, 2001. 
Landgrebe, David, 2004. Hyperspectral Image Data Analysis. 
Dept. of EE, Purdue University. http://dynamo.ecn.pur-due.edu 
/biehl/Mulyispect. 
Tso, B. and P.M. Mather, 2001. Classification Methods for 
Remotly sensed Data, Taylor & Francis printed in London and 
New York, 2001. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Number of Features=369600 Bytes 
True Class Classifier results 
Corn Soybeans | Woods Wheat | Sudex Oats Pasture Hay Nonfarm | Totals | %Corret 
Corn 8942 102 145 149 ] 22 0 22 721 10104 88.5% 
Soybeans 6 11717 482 108 8 87 0 14 488 12910 90.8% 
Woods 4 10 328 3 0 ] 2 0 41 389 84.3% 
Wheat 0 8 8 732 0 24 0 9 163 944 77.5% 
Sudex 0 17 0 0 1175 21 0 2 4 1219 96.4% 
Oats | 12 0 8 3 508 0 28 43 603 84.2% 
Pasture 0 0 0 0 0 0 307 0 32 339 90.6% 
Hay 22 | M 21 3 S2 0 502 54 746 79.4% 
Nonfarm 17 69 14 68 ] 111 9 81 3176 3546 89.6% 
Totals 8992 11936 978 1089 1191 826 318 748 4722 30800 89.2% 
Overall Performance = 89.2% CPU Time = 51.52 Seconds 
  
  
  
Table 1. Pixel-Feature performance using Bayes MLC 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Number of Features=13,692 Bytes. Compresstion Coefficient = 27 
True Class Classifier results 
Corn Soybeans | Woods Wheat | Sudex | Oats Pasture Hay Nonfarm Totals | %Corrct 
Corn 9592 123 17 67 0 6 0 66 233 10104 94.9% 
Soybeans 24 12409 209 74 | 27 0 11 155 12910 96.1% 
Woods 0 4 385 0 0 0 0 0 0 389 99.0% 
Wheat 6 11 12 824 0 11 0 0 80 944 87.3% 
Sudex 0 9 0 0 1193 13 0 3 1 1219 97.9% 
Oats 4 I 0 2 0 588 0 0 8 603 97.5% 
Pasture 0 0 0 0 0 0 339 0 0 339 100.0% 
Hay -| 45 0 0 0 9 | 0 691 0 746 92.696 
Nonfarm 69 136 12 94 8 244 0 118: 2865 3546 80.8% 
Totals 9740 12693 635 1061 1202 890 339 889 3342 30800 93.8% 
Overall Performance = 93.8% CPU Time = 1.88 Seconds 
  
  
  
Table 2. Object-Feature performance using Bayes MLC 
825