Kernel composition with the one-against-one cascadefor integrating external knowledge into SVM classification

This work focuses on two main questions.How can data fusion be performed beforeSVM (support vector machine) classification? Andsecondly: how can the one-against-one cascade beexploited to use information selectively thus integratinghuman knowledge? Kernel compositionrepresents a specialized method for fusing data onthe feature level. Its main advantage is given by thefact that it reduces the Hughes phenomenon (performancedecrease due to high dimensionality) becauseit abstains from raising dimensionality in thefeature space. Since the paper focuses on hyperspectraldata, a specialized kernel based on thespectral angle is employed and evaluated. Two applicationschemes are presented. At first, hyperspectraldata are fused with laserscanning data bytaking into account explicit knowledge on roof geometries.Secondly, a spectral-spatial frameworkfor hyperspectral data is presented which integratesimplicit knowledge on the relevance of spatial contextinto classification. Both approaches are promisingas they obtain higher classification accuracieswhen integrating external knowledge. The innovationof the contribution is that data fusion with asecond source of data via kernel composition iscombined with a modification of the one-againstonecascade which allows integration of humanknowledge.