Optimal segmentation parameters prediction using a orthogonal decomposition approach for geographical object based classification of urban areas

Abstract

Good image segmentation produces objects that are internally homogeneous and distinct from their neighbors. Several approaches are used to assess the quality of segmentation parameters including visual observation of multiple segmentation results and techniques, which calculate measures of intra segment homogeneity and inter segment heterogeneity. Some of these techniques have been reported without considering under and over segmentation; occurrences and others exhibit some mathematical formulation instabilities when dealing with very heterogeneous areas. Moreover, the majority of segmentation assessment techniques focus on the evaluation of the scale parameter and give little attention to the influence of other parameters, such as compactness on the performance of the scale parameter. To the best of our knowledge, no existing approach can predict the optimal compactness thresholds that would enhance the performance of the scale parameters, or predict which scale parameter will perform best under certain compactness constraints. This paper proposes three spatial autocorrelation models, which identify the optimal parameters for single scale and multilevel segmentations of urban environments. Tested using a GeoEye satellite image of the City of Cape Town (South Africa), results reflect that the best combination of parameters to optimally segment the area and those that would poorly perform were successfully identified.