Improving object-oriented land use/cover classification from high resolution imagery by spectral similarity-based post-classification

To classify an image, traditional classifiers depend mainly on the spectral and/or textural distinctions between different land-cover units, while this study attempts to explore the properties of statistical distinction. Using the historical classification results, we present a novel algorithm for imagery classification that achieves high accuracy, automation and efficiency. Based on object-oriented image analysis, it exploits the advantages of d(ch) (the Chaudhuri's metric) using a multi-step approach, and the objective is not to reclassify an image, but to refine or update the existing land-use/-cover classification results by comparing the pairwise d(ch) value (namely similarity) between different image segments. Finally, the similar/homogeneous segments will be confirmed as their original class labels, while the inhomogeneous/dissimilar segments will be masked out with an appropriate threshold on the similarity image and be relabelled. We have systematically evaluated the algorithm by running it on the basis of the existing GIS base maps, which indicated the good performance of it.

Automated summary

This study proposes a novel algorithm for image classification that explores statistical distinctions and refines or updates existing land-cover classification results by comparing the similarity between different image segments. The algorithm, based on object-oriented image analysis, demonstrates good performance when evaluated on existing GIS base maps.