Unsupervised Selection of Color Factor Weight and Segmentation Scale Parameter for Successful Segmentation of Urban Land Use/Land Cover

Abstract. Image segmentation is a crucial step in object-based image analysis of urban remote sensing data. Its primary goal is to divide a digital image into meaningful objects that are internally homogeneous and clearly distinguishable from neighboring segments. While the segmentation scale parameter helps limit the size of these image segments, it alone cannot guarantee optimal intra-segment homogeneity or inter-segment separability. Many existing segmentation quality assessment methods rely on spatial autocorrelation measures, which often lead to irregular variations in the global objective function. As a result, segments representing spectrally distinct but spatially large objects may be incorrectly merged during the final stages of segmentation, leading to significant over- and under-segmentation errors. This paper presented an unsupervised hybrid segmentation evaluation strategy that combines the Moran’s index and the image standard deviation measures. The proposed segmentation evaluation strategy was tested on a color aerial image of the Cape Town metropolitan area. Experimental results show that the proposed approach successfully identified optimal combinations of scale parameters and color factor weights that minimize over- and under-segmentation of the image. The approach achieved very promising over- and under-segmentation (OS and US), as well as area fit index (AFI) error magnitudes, in comparison to some of the existing state-of-the-art approaches available in the literature. It was also found that associating small weights of color factor with medium-range scale parameters resulted in optimal segmentation outcomes, while larger segmentation scale parameters required large weights of the color factor to produce meaningful segmentation outcomes. Furthermore, it was uncovered that the spatial autocorrelation curve achieved stability at optimal segmentation parameters and a near horizontal fluctuation shape, describing a drop of image variance to values very close to zero at optimal segmentation parameters.