Improvement of land-use classification using object-oriented and fuzzy logic approach
Traditional classification approaches are all
pixel-based and do not utilize the spatial and context
information of an object and its surroundings, which has
potential to further enhance digital image classification.
Instead of pixel-based, pixels groupings and object segmentation offers more innovative techniques to image
classification. In this study, land cover types in the Klang
valley, Malaysia were analyzed to compare classification
accuracy between the pixel-based and the object-oriented
image classification approaches. Landsat 7 ETM+ with six
spectral bands was used for the land cover classification. In
the pixel-based image classification, supervised classification was performed using the maximum likelihood classifier. On the other hand, the object-oriented image
classification was performed using the combination of
object segmentation using fuzzy dimension techniques.
The selected parameters for image segmentation were:
scale parameter 15, homogeneity composition criterion
(color 0.7 and shape 0.3), shape criterion (smoothness 0.9
and compactness 0.1). Fuzzy dimension functions were
devised to classify the segmented image objects. The
classification results showed that the object-oriented cum
fuzzy logic approach was superior to that of the pixel-based
supervised classification. The former has achieved higher
overall, producer and user accuracies for most of the land
cover classes compared to those of the latter. In addition,
the accuracy of the former has met the requirements of
international standard for digital mapping with overall
accuracy exceeding 85%; Kappa value above 0.85 while
accuracy differences among the classes were kept minimal.