Classifying roof materials using data fusion through kernel composition — Comparing ν-SVM and one-class SVM

Many roof materials are considerable sources of pollutants. A reliable classification approach is required to identify these materials. It is beneficial to fuse the information of hyperspectral and laserscanning data. As the resulting feature space is high dimensional an advanced classifier is needed. Support vector machine classifiers based on kernel composition provide a reasonable mean to cope with high dimensionality. Futhermore, they combine information on the feature level. We evaluate the capabilities of two types of SVMs - ν-SVM and One-class SVM. They use the same mathematical foundations, but are employed conceptually different. While ν-SVM classifies relatively, One-Class SVM is an absolute classifier. Therefore, it can be used to avoid an overly intricate training procedure which would incorporate all classes present in the scene. A comparison classifying roof materials in a combined dataset of HyMap and TopoSys data is presented.