Roeland de Kok
7. DISCUSSION & CONCLUSION
The results of the pixel based and the object-oriented classification of the SPOT images were compared, using a
set of test areas distributed randomly over the image. A confusion matrix was built, giving the pixel numbers of
the test areas of each class and the percentages of correctly and wrong classified pixels of each class. As the
result shows, (Tables 2 & 3) the classifications do not differ considerably. This could be due to the similarity of
the classifiers used and the generally rather simple task.
Class Pixel Water Non-forest N. Pumilio N. Antarctica
Water 0,0 0,02 -
Non-forest 0,05 -
N. Pumilio 0,96 003
N. Antarctica 0,03 017 1 08
General Accuracy 90,95
Average Accuracy 93,21
Table 2: Classification accuracy of the pixel based classification
Class Pixel Water Non-forest N. Pumilio N. Antarctica
Water 1459 | 100° Dew : :
Non-forest 1606 = 097 - 0,03
N. Pumilio 2442 - 0,02 0,96 0,02
N. Antarctica 2617 - 0,08 0,05 0,87
General accuracy 94,47
Average accuracy 96,09
Table3 Classification accuracy of the object oriented classification
What the confusion matrix does not show are the different appearances of the two classification results. Whereas
the pixel-based map (Figure 6 B) shows a kind of salt & pepper effect with single pixels being distributed in the
two forest classes, the object-based classification (Figure 6 A) rather resembles a manually digitized map.
In this case the result of the object-oriented classification is preferred because it comes closer to the situation in
the field. In reality, the two Nothofagus species do not build mixed stands so the small dots seen in the forest
classes of the pixel-based classification are mostly misclassified pixels.
TR .
cn MEE E
5o SE
ve
+ £5. 4 A7 ;
© Water N. Pumilio o N. antarctica CO Non-forest
Figure 7: Results of object-oriented (A) and pixel-based (B) classification
220 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
