Recently, the application of machine learning and data fusion techniques on hyperspectral imagery have demonstrated potential for ground cover classification at material level. Hereby, specific locations of resources enclosed in cities (e.g. roof materials) can be identified, which is critically relevant within the field of urban mining. A limitation of this approach is the so-called 'pepper and salt effect', the oversensitivity of the classifiers to spectral variations within a pixel (e.g. chimneys, roof windows). Identifying and correcting affected pixels can be done statistically (e.g. using a majority filter), but not in cases where spectral variations affect a majority of pixels characterizing a surface. A solution to this limitation would be the usage of 3D city models containing the objects inducing the spectral variations. However, such highly detailed 3D city models are often unavailable as they cannot be produced automatically yet. An alternative covered by this research is to use a less detailed 3D city model and semantically enrich it with the required data. As 3D city models are usually produced using a point cloud, such a point cloud is used to perform the enrichment. The main research question addressed is therefore: How can a CityGML LOD2 model be semantically enriched in order to improve material classification performed on roof surfaces?.      To address this, an existing LOD2 model was compared to a point cloud acquired by Ligth Detecation and Ranging and 'deviation' points were identified. This identification uses a distance check for seed selection and performs a region growing with an orientation check. In a subsequent step, 'deviation' point regions were translated into a geometric shape by usage of their Voronoi diagram and fused with the pixels of hyperspectral imagery. Part of this research is also a nominal validation analyzing a total of 41 buildings and 831 pixels located in the south of Rotterdam (Netherlands). Overall kappa values of up to 0.7 and commission errors as low as 10% (for the class 'clean' pixels) were obtained, showing potential of the chosen method. Additionally, a rational validation was performed to assess the impact of potential tolerance of classifiers for 'spectral deviations'. This one only included 10 buildings, but took into account 328 pixels located up to 30% outside the roof surface A main outcome is the recommendation on settings to use depending on the specific user needs. To accurately quantify materials, relatively 'loose' settings are recommended. In contrast, to identify presence of materials, stricter settings are recommended. Beyond this, recommendations to data suppliers and potential applications of the method to other fields are formulated.

There is a paradigm shift from two- to three-dimensional data, from maps to information dense models. Self-driving cars, digitization of historic buildings or maintenance of highway infrastructure are a small selection of many applications that use laser scanning to acquire three-dimensional data of our physical surroundings. Most of these applications require more than the shape of their surroundings. For example, a self-driving car needs to identify pedestrians, road signs and traffic lights in order to navigate safely. Therefore the point cloud acquired by laser scanning needs to be enriched with additional information. Automatic assignment of the object type a point belongs to is called classification.

This research focuses on deep learning for point cloud classification, because it revolutionized classification of imagery. PointNet is used to enable deep learning directly on point cloud data sets. To date PointNet is proven for indoor point clouds, this research explores the application of PointNet on an outdoor highway scene. The methodology creates point cloud training data efficiently by reusing known 2D object locations. Different spatial representations and sampling methods for the points are tested.

On average the method classifies 50% of the points correctly in four object classes. In combination with clustering of the point-wise predictions, the method predicts 60% of the 2D object locations successfully. The performance is comparable to the 47% average class accuracy PointNet achieves for 13 classes on the indoor data set.