Integrating external knowledge into SVM classification - Fusing hyperspectral and laserscanning data by kernel composition

Roof materials are important sources of pollutants within cities. To monitor and quantify polluted roof runoff, a precise classification approach of various roof material classes is needed. Within urban environments, different geometries of roofs exist, e.g. sloped roofs and flat roofs, which can be distinguished by using ALS data. To precisely classify different roof material classes, e.g. brick, slate, gravel, hyperspectral datasets can be utilized. Thus, exploitation of both hyperspectral and ALS data is helpful. In order to exploit these data sources, data fusion needs to be performed. A novel approach for data fusion is possible with kernel composition methods. Support vector machines (SVMs) have proven to be capable classifiers for hyperspectral and ALS data separately, but also for combined datasets [Camps-Valls et al., 2006]. Kernel functions are used to find the solutions of SVM classifiers. The kernel composition [Camps-Valls et al., 2006] takes account of the fact, that kernel functions can be combined (e.g. by addition) to form new kernels. This combination offers a novel option for data fusion. An application for the fusion and classification of hyperspectral and ALS data is given in [Braun et al., 2011].