A technique for optimal selection of segmentation scale parameters for object-oriented classification of urban scenes

Abstract

Multi-scale image segmentation produces high level object features at more than one level, compared to single scale segmentation. Objects generated from this type of segmentation hold additional attributes such as mean values per spectral band, distances to neighbouring objects, size, and texture, as well as shape characteristics. However, the accuracy of these high level features depends on the choice of segmentation scale parameters. Several studies have investigated techniques for scale parameter selection. These proposed approaches do not consider the different objects’ size variability found in complex scenes such as urban scene as they rely upon arbitrary object size measures, introducing instability errors when computing image variances. A technique to select optimal segmentation scale parameters based on image variance and spatial autocorrelation is presented in this paper. Optimal scales satisfy simultaneously the conditions of low object internal variance and high inter-segments spatial autocorrelation. Applied on three Cape Town urban scenes, the technique produced visually promising results that would improve object extraction over urban areas.