One of the most wide-spread applications for multispectral images is ground cover classification. For the assessment of classification agreement between true and fusion-derived imagery the Iterative Optimization Clustering (or Migrating Means) k-means algorithm as described e.g. by [Richards (1993)] was used. The initial clusters in the spectral domain were chosen randomly and the pixels were assigned iteratively to the currently nearest candidate cluster on the basis of the Euclidean distance measure.
First, the original image truly sampled at 1 m was classified into a particular number of ground cover classes. The such found cluster centers were then used again to classify the fused images. This ensures that it is meaningful to compute a confusion matrix and a kappa coefficient [Congalton (1991)] in order to quantify the error induced by fusion.