IMPACTS OF A RESOLUTION PYRAMID ON GIBBS RANDOM FIELD CLASSIFICATION

摘要

In this paper we describe how to extend the Markov/Gibbs random field model used for texture classification to a multi-resolution approach in the context of land cover analysis. Our method allows a scale-invariant analysis of image data by applying single scale classification on different resolution levels of the input image. Finally, the independent but highly correlated outcomes are joined. Evaluating a set of four different landcover classes, sampled in five different resolutions leads to 15000 combination results. Based on ground truth data of land coverage, all results were analysed for their correctness. We present an appropriate strategy, using reclassification results to dynamically choose one combination. Our experiments expose a clear tendency towards certain combinations which support the proposed approach. By priotarising those resolutions with a better reclassification result we achieve a median correctness ratio of 99percent when incorporating all regions of our test scene.