Kernel Composition with the one-against-one Cascade for Integrating External Knowledge into SVM Classification

Summary: This work focuses on two main questions. How can data fusion be performed before SVM (support vector machine) classi›cation? And secondly: how can the one-against-one cascade be exploited to use information selectively thus integrating human knowledge? Kernel composition represents a specialized method for fusing data on the feature level. Its main advantage is given by the fact that it reduces the Hughes phenomenon (performance decrease due to high dimensionality) because it abstains from raising dimensionality in the feature space. Since the paper focuses on hyperspectral data, a specialized kernel based on the spectral angle is employed and evaluated. Two application schemes are presented. At ›rst, hyperspectral data are fused with laserscanning data by taking into account explicit knowledge on roof geometries. Secondly, a spectral-spatial framework forhyperspectraldataispresentedwhichintegrates implicit knowledge on the relevance of spatial context into classi›cation. Both approaches are promising as they obtain higher classi›cation accuracies when integrating external knowledge. The innovation of the contribution is that data fusion with a second source of data via kernel composition is combined with a modi›cation of the one-againstone cascade which allows integration of human