Combining the heursitic and spectral domains in semi-automated segment generation

Abstract

The generation of thematically accurate image segments or delineating land-cover elements is a common objective and challenge in Geographic Object Based Image Analysis (GEOBIA). A core notion to the applicability of segmentation algorithms for partitioning these land-cover elements is that said elements typically have some spectral or other homogeneity criteria that allow successful segmentation, to varying degrees. One approach that addresses this challenge models the parameterised segmentation process as a search problem. The search method is provided with a reference of optimal desired output. This idea is extended in this work by suggesting the encoding of spectral space transformations as additional variables to such a search problem. The automatic exploration and transformation of the spectral domain can allow for a closer correlation between thematic and spectral similarity of the land-cover elements of interest, thus aiding the segmentation process. Two simple spectral transformation methods functioning in conjunction with two scale-space constraint image segmentation algorithms are presented to illustrate this concept. A statistically significant improvement in segmentation results can be obtained consistently in acceptable time with this approach with off-the-shelf meta-heuristics for our test areas. It is also shown with the algorithms used in this study that the segmentation algorithm parameters (heuristic domain) are dependent on the spectral transformation parameters (spectral domain) to achieve optimal results. This necessitates simultaneous optimisation of these two domains.