Detecting an Optimal Scale Parameter in Object-Oriented Classification

Avoiding spatial autocorrelation is the key to many research questions especially for field design, remote sensing data selection, and maximum spatial variation caption. Spatial variation across land cover types as well as the gradients inherent in ecotones can be captured in reflectance which is a spatially continuous variable. The spatial variation between reflectance values of any two pixels will depend on the lag distance beyond which pixels are no longer spatially autocorrelated. This paper demonstrates the utility of semivariogram for determining the lag distance in which pixels will be spatially autocorrelated. According to sampling theorem, objects should be sampled at half their width such that spatial resolution should be half of the semivariogram lag distance. As object-oriented classification is now the most broadly accepted classification method, scale parameter determination is the foremost important decision for determining the size of image objects. The scale parameter was adjusted during image segmentation to test how the size of image objects changed. The optimal scale parameter was chosen when the average distance between neighbouring image object centroids was near to the lag distance of the semivariogram. Results showed that the size of image objects reached a scaling threshold as the scale parameter was increased. When the scale parameter was adjusted to create image objects that exceeded this threshold, the segmentation was not able to accurately represent the spatial variation observed on the ground.