866 
node of the layered classifier was also 
tested. 
Finally, the single-step classifier 
contained also a version with a 
two-channel set including ND and TM5, 
which yielded excellent results in a 
previous exercise (De Wulf and 
Goossens 1989). Table 2 lists the dif 
ferent classification versions. 
The minimum distance algorithm was 
used in all tested classification 
methods. 
sification accuracy. This parameter is 
an estimate of kappa, which adjusts the 
classical "percentage correct" measure 
by subtracting the estimated 
contribution of chance agreement 
(Campbell 1987). 
From the khat values in Table 3 it can 
be concluded that the layered clas 
sification based on spectral groupings 
using TM channels, and the ND • at the 
first node (methods 3 and 4) yields 
Table 2 List of the different classification versions. 
Layered classification 
Spectral classes 
TM channels (2-7) 
* Single channels (1) 
* Subsets of channels (2) 
* ND at first node + remaining single channels 
at other nodes (3) 
* ND at first node + subset remaining channels 
at other nodes (4) 
SPOT-l-equivalent channels (TM 2, 3 and 4) 
* Single channels (5) 
* Subsets of channels (6) 
* ND at first node + remaining single channels 
at other nodes (7) 
Information classes 
TM channels (2-7) 
* Single channels (8) 
* Subsets of channels (9) 
* ND at first node + remaining single channels 
at other nodes (10) 
* ND at first node + subset remaining channels 
at other nodes (11) 
SPOT-l-equivalent channels (TM 2, 3 and 4) 
* Single channels (12) 
* Subsets of channels (13) 
* ND at first node + remaining single channels 
at other nodes (14) 
Single-step classification 
TM channels (2-7) 
* One channel (6) (15) 
* A subset of channels (2, 3, 4 and 7) (16) 
* Normalized difference (17) 
* ND + TM5 (18) 
SPOT equivalent channels (TM 2, 3 and 4) 
* One channel (2) (19) 
* A subset of channels (3 and 4) (20) 
4. RESULTS AND DISCUSSION 
Table 3 presents excerpts from the 
error matrices of the 20 clas 
sifications tested out. The results 
refer to classifications of a set of 
test pixels independent from the trai 
ning pixel set. 
Only the results of classification of 
natural vegetation classes, which are 
of major concern in this exercise, 
have been included in Table 3. However, 
to evaluate the discrimination between 
agriculture classes and natural vegeta 
tion classes, the khat value was cal 
culated as a measure of overall clas- 
the best results for Level II and Level 
III classes. 
Classification into natural vegetation, 
agriculture and water is best performed 
using the single step classifier with 
ND and TM5 (method 18) . The good re 
sults of the latter combination 
confirms earlier results obtained for 
temperate forests (De Wulf and Goossens 
1989). 
Closer examination of the clas 
sification results for single natural 
vegetation classes reveals that no 
single classification method stands out 
as the best.