A Framework for the Evaluation of Multi-spectral Image Segmentation

A general framework for testing the quality of the segmentation of a multi-spectral satellite image is proposed. The method is based on the production of synthetic images with the spectral characteristics of the image pixels extracted from a signature multi-spectral image. The knowledge of the exact location of objects in the synthetic image provides a reference segmentation, which allows for a quantitative evaluation of a segmentation algorithm applied to the image. The Hammoude metric and the external similarity indices Rand, Corrected Rand and Jaccard are used. A practical application was carried out to illustrate the value of the proposed method. Two satellite images, from SPOT HRG and Landsat TM, were used to extract the spectral signature of 8 land cover types. Six test images were produced using all 8 land cover classes and with two different sub-sets with 5 classes. The segmentation results provided by a standard algorithm were compared with the reference or expected segmentation. An evaluation of the parameters used in the eCognition software segmentation algorithm was also carried out, using the proposed indices.