of 
ed into 
and 
als of 
(4.1) 
viations 
^n gd 
  
Fig. 8 shows the classification maps by both methods. The several shades of grey 
should be interpreted as follows: 
- pure white patches: rejected (unclassifiable) pixels 
- dull white : water (lake) 
- very light grey : built-up areas 
- light grey : forests 
- dark grey : farms 
- not very black 3. grass 
- black : marsh 
Both methods produced similar distributions of the object classes. However, if 
the two photographs are compared, four situations stand out quite clearly in 
the photograph of the Maximum Likelihood classification (Fig. 8 (a)): 
- reject.(or unclassifiable) pixels are every where on the photograph; and 
to be exact, there are 2,571 rejects. On the other hand there are only 
50 in the Separating Hyperplanes classification, and these, by another 
visit, have been found to represent large groups of sail boats at the 
banksof the lake adjacent to the built-up areas, as well as an airfield 
near the east-central edge of the photograph. 
- boundaries between classes whose variance levels do not differ much have 
their mixed pixels rejected. 
- boundaries between classes whose variance levels differ considerably have 
their mixed pixels assigned to the class of higher level of variance, 
with absolutely no rejects. 
- an island near the south-west end of the lake (classified as marsh by the 
Separating Hyperplanes) was completely rejected as unclassifiable. 
A conclusive demonstration of the extend to which the statistical proper- 
ties (means and variances) of Landsat MSS data could degrade classifications 
by the Maximum Likelihood method is found in a comparison of the classification 
results (Fig. 9) by both methods of the data of an area near the city of Jos, 
Nigeria, with 22, 374 pixels. The path of a river which runs westwards into a 
lake in the south-western corner of the area is better defined in Fig. 9(b) 
(the Separating Hyperplanes classification) than in Fig. 9(a) (the Maximum 
Likelihood classification). The variance levels of the different classes are 
shown in Table 3. 
The area is mainly farmland. The river path has been classified as bare 
ground, because, but for the lake, all water in the area had dried out at the 
time of the acquisition of data by Landsat in December 1975. Considering only 
the two classes, farmland and bare ground, it should be said that these are 
represented in the right proportions in Fig. 9(b) (Separating Hyperplanes 
classification). This situation is confirmed by the fact that the river path 
(i.e. bare ground), which exists within farmland, is quite clearly defined. 
Fig. 9(a) (Maximum Likelihood classification) shows much more bare ground 
because of the higher level of variance of bare ground in comparison with farm- 
land (21.6 against 3.4). Table 4 shows that bare ground has 18.56 % more of 
the area than it should have, and that farmland has 10.67 2 less. 
Other advantages of the Separating Hyperplanes method over the Maximum 
Likelihood method include the following: 
(1) There is no minimum number of training pixels. In other words, a pixel per 
class are adequate to carry out a classification (Ekenobi 1981). This can 
be advantageous when a rare class must be represented in a classification 
67 
aud o EE PE A te tt ra C Er A