rks
not
DS
1an
>ct-
ade
and
DT
en.
by
nds
one
nts
her
eria
ised
ture
one
ised
lage
Roeland de Kok
Classification
In order to be able to compare different object features like color and size as well
as uncertain statements, fuzzy logic functions are used for classification (Figure
2). This allows classification of very complex tasks on one hand, and makes
classification transparent and adjustable in detail on the other. Fuzzy logic is a
mathematical approach to quantify uncertain statements. Basic idea is to replace
the two strict logical statements 'yes' and 'no' by the continuous range of [0...1],
where 0 means exactly no and 1 means exactly yes. All values or expressions
between 0 and 1 represent a more or less certain state of yes and no. To translate
the range of most different features into fuzzy logic expressions, the software uses
two kinds of classifiers, membership functions and next neighbor classifier. All Figure2: ^ membership function
expressions of one class have to be combined to produce a result. This is done using logical operators such as and
(max), and (mean), or, if and else (MAYER MECHLER SCHLINDWEIN WOLKE, 1993).
The features used for classification can be divided into three categories:
Object Features like color, texture, form and area
Classification related features like relations to sub-objects, super-objects and neighbor objects
Terms like nearest neighbor classifier or similarity to other classes
6 RESULTS
6.1 Pixel based classification of the SPOT data
The first step was to analyze the separability of forest from non-forest as well as the two Nothofagus forest types using
a maximum likelihood classifier without NULL class as featured in the PCI ImageView software.
The following 4 classes were separated:
e Water
e Non-Forest
e Nothofagus Pumilio
e Nothofagus Antarctica
As one can see on the scatter plot, there are significant overlaps between the classes Nothofagus Pumilio and
Nothofagus Antarctica as well as Non-Forest (Fig. 3). This is due to the strong inhomogeneity especially in the near
infrared channel inherent to the two forest and the non-forest classes. This inhomogeneity is caused by the distribution
of the pixel values in feature space, which is not normally distributed as expected by maximum likelihood classifiers.
Table 1 shows the spectral mean values and standard deviation for the five classes. Given the small number of classes
the classification result was nevertheless satisfactory. The confusion matrix based on the training areas (not on
independent test areas) gave an average accuracy of 96.9% and a Total accuracy of 98.8% (LILLESAND & KIEFER
1994, RICHARDS & JIA, 1999).
A
X Class Channel Mean Std dev
Green 72,5 4,76
Non-forest Red 52,9 5,12
NIR 77,0 17,72
Green 60,7 5,92
Non-Forest t
Nothofagus antarctica N. Pumilio Red 34,9 4,65
| i: NIR 95,9 17,82
| / Nothofagus pumilio
Water \ / Green 62,4 3,43
; N. Antarctica Red 37,9 3,27
NIR 87,4 5,83
Green 62,4 3,43
| ow : DE Water red 37,9 3,27
NIR NIR 87,4 5,83
Figure 3: Scatter plot of the maximum likelihood Table 1: Spectral values of the different classes
classification with five classes.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 217
