were identified through a field study in the time of satel 
lite snap shoting. This study and maps formed the ground 
truth data base for our field definition and class identifica 
tion. The digital computations were performed using inter 
active image analysis system developed by the author. The 
mean vector and covariance matrix of each training field 
were calculated to develop the spectral signature represen 
tative of land-cover classes. The objective of the analysis, 
was to discriminate among following 15 agricultural classes: 
water, red clover, white clover, wood, wheat, maize, mix 
ture, millet, grass, harvested rapeseed, rapessed, sugarbeet, 
wheat II, clover and residential area. The resubstitution 
estimations of the probability of correct classification are 
summarised in the table. 
performance. The overall performance of contextual clas 
sifier is better in tested example than the non-context per- 
point Bayesian one. Similar improvement is seen for all 
classes to be searched. This result was reached with the 
smallest possible context window of eight surrounding pix 
els. The optimal size of window depends on local condi 
tions ( average field area ) and on the used image resolu 
tion. Computational time of presented algorithm was in 
tested example approximately five times longer than for 
the standard Bayesian classification one. Further work is 
still needed to optimize developed software to increase clas 
sification speed. 
REFERENCES 
Peterka,V.,19Sl.Bayesian approach to system identification. 
In. Trends and Progress in System Identification. Ed. P. 
Eykhoff , Pergamon Press, Oxford. 
Tubbs,J.D.,1980. Effect of Autocorrelated Observation on 
Confidence Sets Based upon Chi-Square Statistics. IEEE 
Trans. Syst. Man. and Cybern., vol. SMC'-10(4): 177-180. 
class 
Contextual 
Bayesian c. 
Bayesian 
classifier 
1 
1 
1 
2 
0.8 
0.72 
3 
0.9 
0.82 
4 
0.95 
0.S1 
5 
0.98 
0.93 
6 
1 
1 
7 
0.92 
0.79 
8 
1 
0.93 
9 
0.77 
0.56 
10 
0.97 
0.97 
11 
1 
0.89 
12 
0.99 
0.91 
13 
0.85 
0.6 
14 
0.92 
0.88 
15 
0.95 
0.68 
P 
0.93 
0.85 
We have chosen the autoregressive model of order one (N = 
1) and the smallest possible thematic map window of nine 
pixels (n = 9). In such a case, the contextual classifier 
needs h(*) — 2010, h( + ) = 1479 operations, while the 
conventional approach only /?.(*) = 525, h{ + ) = 435 ones. 
6 CONCLUSION 
The tested example shows improvement in classification 
342