New techniques continue to be developed to effectively employ different features inherent in remote
sensing and ancillary data for improving LULC classification and change detection results [14–17].
Compared to the pixel-based classification approaches, object-oriented classification methods can
effectively improve classification results and better express spatial information [4,18]. Typically,
object-oriented classification methods consist of two major steps: image segmentation and classification
of the meaningful segmentation objects [19]. Previous studies have proved that incorporation of
new remote sensing index bands (e.g., water bodies, vegetation and construction indexes), auxiliary data,
such as digital elevation model (DEM) terrain factors, and decision trees based on expert knowledge
can improve accuracy and solve image problems, such as “salt and pepper” problem in the
classification [16,18,20].