Accurately classifying laser-scanned point cloud data remains a critical challenge in geospatial analysis, particularly due to the complexity and volume of the data. This thesis presents a novel, confidence-aware deep learning framework designed to improve the classification accuracy of point cloud data, specifically focusing on the Actueel Hoogtebestand Nederland (AHN) dataset. The framework integrates geospatial knowledge into the deep learning process, enabling the model not only to refine its predictions through iterative learning but also to enhance the training data along the way via iterative online learning, ensuring continuous improvement in both training data quality and model performance.

The preprocessing phase assigns confidence scores to each point in the point cloud based on local neighborhood properties, with additional input from multispectral imagery (MSI) to further enhance the confidence estimation. These confidence scores are central to the online learning process, where the model prioritizes high-confidence points for training while progressively updating lower-confidence points to improve accuracy. To test the hypothesis that confidence-aware learning can enhance point cloud classification, we selected the KPConv network due to its suitability for handling unstructured data and capturing complex geometric features.

Extensive experiments demonstrate that the proposed framework, particularly with the Online strategy, enables deep learning models to perform better when trained solely on native point cloud attributes (elevation and intensity) compared to models without this strategy. Importantly, the Online strategy qualitatively enhances the training data by refining labels and reducing noise, thereby supporting more robust model performance. While incorporating additional features from aerial imagery showed no overall improvement, specific classes, like High tension and others did see performance gains.

Automated reconstruction of detailed semantic 3D city models is challenging due to the need for high-resolution (HR) and large-scale input datasets, the ambiguous definition of the ensuing model, the intricacy of the processing pipeline, and its costs. Furthermore, existing methods mainly focus on geometry rather than semantics. Detailed semantic models may include roof installations whose size and function vary: dormers, windows, chimneys, etc. All elements visible on the roof from an aerial view are called ‘‘superstructures”. Deep Learning techniques can facilitate their modelization. This work inscribes itself in a project developed at the Technical University of Munich. The existing pipeline employs a convolutional neural network (CNN) on aerial images segmenting roof superstructures. These results can then be vectorized, extruded in 3D with their semantic description, and added to a simple 3D model.

This thesis demonstrates that building height data fused to a CNN on RGB aerial images improves the semantic segmentation of roof superstructures for classes with relief. Fusion of absolute and relative height data with different interpolation methods applied to LiDAR point cloud data is achieved through a fusion network from the state-of-the-art (FuseNet). First, experiments prove that detection accuracies increase by 11% on average for dormers and 12% for chimneys compared to U-Net output on the same dataset. Best performance is reached with the fusion of absolute height (rather than normalized) and IDW or NN interpolation technique (rather than none). However, although superstructure types are better recognized, their boundaries are fuzzier due to data input mismatches, and more background pixels are classified. Secondly, the predictions and modelization of both a Bavarian and Dutch test set prove the technique scalability. However, a training set annotated for Bavaria and applied to a test set in the Netherlands yields inaccurate results due to local architectural typologies and different input data characteristics.

Semantic segmentation of aerial images is the ability to assign labels to all pixels of an image. It proves to be essential for various applications such as urban planning, agriculture and real-estate analysis. Deep Learning techniques have shown satisfactory results in performing semantic segmentation tasks. Training a deep learning model is an expensive operation, while most of the time manually labelled images are required. Additionally, a bottleneck in semantic segmentation projects concerns the annotation of images. Consequently, synthetic data, which consists of images from a virtual world that simulates the real world, can be used as training data for segmentation tasks to improve the classification results. Therefore, this thesis aims to create a pipeline that generates synthetic images with semantic segmentation labels to be used in an existing deep learning model and discuss how the generated synthetic data improves the semantic segmentation of aerial images. In this research work, an existing model (FuseNet), which in previous works achieved satisfactory results, is trained with solely synthetic data and a mix of real data in different training and testing scenarios to classify true ortho imagery from Haaksbergen, Netherlands and Potsdam, Germany. In addition, a benchmark of domain adaptation techniques is performed to close the domain gap between the synthetic and real imagery. The semantic maps include building, road and other classes. Experiments are performed to test the performance of the synthetic data using 1) Different 3D models of the virtual world, 2) Different quantities of synthetic and real training data, 3) Different cross-geographical scenarios, and 4) Different domain adaptation techniques. The assessment is based on the (mean) intersection over union (IoU), F1 score, precision and recall and an extensive visual assessment. The virtual world is created through a pipeline in CityEngine using procedural modelling techniques and then rendered in Blender to create the training dataset. The results show that the synthetic data has a mIoU of 0.48, which is lower compared to cases when solely real data (0.75) are used, when the segmentation is performed in the same training and testing area. In addition, the 3D models partly affect the segmentation results. When using a mix of real and synthetic data, the results are maintained to a mIoU of 0.75. On the contrary, when training and testing in different areas, the use of synthetic data seems to improve the results on average by 21.5, 12.5, 1.5 and 2 percentage points on the mIoU, IoU for classes building, road and other respectively. Additionally, domain adaptation techniques such as Cycle GAN and Cycada improve the performance of synthetic datasets by 4 percentage points. Overall, this thesis shows that when the domain difference between the training and testing datasets is big, the addition of the synthetic data helps to improve the performance of the semantic segmentation of aerial images. Synthetic datasets improve the segmentation results by using a mix of existing labelled imagery from different geographical regions when a project lacks labelled imagery. In contrast, when labelled imagery is present in the same testing area, the real training data obtains robust results, thus the addition of synthetic data does not improve the segmentation results.

Storing accurate models of complex geometries in a compact way has become an increasingly challenging issue, especially when dealing with large datasets. One of such datasets is Cobra-Groeninzicht's database of all trees in the Netherlands. In the gaming industry, a new technique is being used to generate tree models: the L-system. An L-system stores a string representation of the structural model of a tree, with the added possibility for recursive modelling using growing rules. This format proves a promising alternative to more traditional methods of storing complex geometries. However, it remains unclear whether it can be an accurate enough representation for modelling and analysing real-life trees.

In this research project, the AdTree algorithm is used to reconstruct a skeleton from a point cloud of a single tree. This skeleton is then transformed to an L-System string format, as well as a CityJSON format (both in JSON structure). The L-system format comes with the advantage that it allows for several methods of increasing its compactness further (growing, generalisation). The overall size of these files also indicates fewer storage space is needed to store the tree geometry. The quality of the L-System skeleton is nearly equal to the input, the skeleton generated by. Assuming it can be read and drawn using a Turtle program, the L-system thus allows for storing the same geometric information more compactly than traditional storage formats, with sufficient accuracy, and the added possibilities of growing or generalising the model.