Urban digital twins rely on accurate rooftop geometry, yet airborne lidar point clouds are frequently sparse and incomplete, leading to substantial information loss in building reconstruction. This thesis investigates diffusion--based learning as a remedy for high-fidelity roof recovery under severe data corruption.

This thesis proposes a two-stage framework that operates on 2.5D height-map representations. Stage~I introduces a dual-task diffusion model that jointly performs roof height-map completion and roof-line prediction. A novel Bidirectional Control Module enables reciprocal conditioning between the two tasks, enforcing geometric consistency during the denoising process. Stage~II employs a patch-based diffusion upsampler equipped with positional embeddings and a domain-specific global context encoder to synthesise high-resolution height maps while remaining computationally tractable for large and variably-sized buildings. A rigorous preprocessing pipeline further yields two challenging benchmarks, \textsc{S80\_i30} and \textsc{S80\_i80}, derived from 160k real-world building samples.

Extensive experiments conducted on these datasets demonstrate the effectiveness of the proposed approach. Under moderate corruption (\textsc{S80\_i30}), the completion model attains an \textit{RMSE} of \textbf{0.89}~m and a Chamfer distance of \textbf{0.06}, improving upon the state-of-the-art RoofDiffusion baseline by 13.2\% and 17.3\%, respectively. In the severe setting (\textsc{S80\_i80}), the method sustains a 13.5\% \textit{RMSE} reduction. The upsampling stage delivers an additional 10\% \textit{RMSE} gain over the best classical interpolator, and the end-to-end pipeline achieves \textit{RMSE} values of 0.91~m (moderate) and 1.42~m (severe).

The thesis contributes: (i) a structurally-aware diffusion framework for roof completion, (ii) a scalable patch-based upsampler, and (iii) public benchmarks that reflect real lidar degradation. Collectively, these advances close a critical gap between theoretical research and practical generation of LOD2.2 building models, facilitating more reliable urban analytics and planning applications.

Bathymetric depth for shallow water regions is essential for coastal management and research. The measurements using echo sounding and LiDAR leave data gaps because vessels cannot reach nearshore waters or the green laser unable to penetrate specific areas. Satellite-Derived Bathymetry (SDB) is an alternative to extract shallow water depths that is able to overcome these problems using multispectral imagery. There are two approaches of SDB: analytical and empirical. The analytical method requires several water properties, which might not be known. The empirical method relies on the linear relationship between reflectances and depths, but the relationship may not be entirely linear due to bottom types variation, water column, and noise. Machine learning approaches have been used to address nonlinearity, but it treats pixels independently, whereas there is a spatial correlation that influences SDB computations since adjacent pixels are correlated to depth. This characteristic of the local connectivity can be captured by Convolution Neural Networks (CNN). Therefore, this thesis conducts a study of SDB using CNN.

This research focuses on the following questions: (i) what kind of preprocessing is needed for the data sets; (ii) what kind of CNN architecture can be used; (iii) what is the accuracy of the method; and (iv) to what extent the pretrained model in certain areas can be reused in other areas. In order to represent a variety of depth, bottom type, turbidity, and water column properties, this study chooses six areas of interest in three different coastal regions: Puerto Rico, Key West, and Hawaii.

With several CNN configurations, the optimum accuracy is obtained using three convolutional layers, a window size of 9x9, and the RGBNSS bands. Based on the experiment and comparison to the previous studies, the accuracy of SDB using the CNN approach outperforms the linear transform, the ratio transform, Random Forest, and the radiative transfer model. The results show that the accuracy decreases as the depth increases and in more turbid water. Comparison between different image preprocessing indicates another benefit of CNN: removing the need to preprocess images since suitable corrections can be automatically performed by CNN given adequate training.

The use of multi-temporal images enhances the variety of training data and thus improves SDB accuracy. However, data variation should be equally distributed to avoid abnormality in the result. Transfer model analysis indicates several limitations of SDB results at particular depths or when implemented to a different water condition, making the coastal water characteristics considered when reusing a pretrained model from one area to another.

In summary, CNN does not require additional image preprocessing and features specifications for training. CNN can produce better SDB accuracy than several other methods. The accuracy improves by increasing the variety of training data. However, SDB using the transfer model still need to be further investigated. A thorough identification of the proportion of sample data is needed to obtain balanced training data. In this way, it is more likely to produce a more reliable and more stable CNN model for extracting shallow water depths in the new data.

Navigational charts play a vital role in a ship's safety while navigating the seas, rivers or lakes. With most of the features and obstructions being out of sight -below sea-level - these charts are more critical than e.g. topographic maps. For routing but also positioning, depth information is a key aspect on these charts. This depth information is available in either depth contours, coloured depth areas and individual soundings. However with the data originating from accurate but usually erratic survey data, a visualisation of raw data is not sufficient for use in a navigational chart directly. It would not clearly convey the information to a human operator in one sight, and thus this visualisation is in need of generalisation: a simplified representation of the same data with irrelevant details being omitted. This thesis gives new insights in the generalisation process for isobaths only and proposes a new framework to deal with those. We propose a framework based on a novel auxiliary data structure to link a triangulation to the resulting isobaths: the triangle region graph. It links the position of isobaths directly to individual triangles, as well as establishes relations between the isobaths themselves. With this structure we ultimately link the survey data to the final cartographic product and thus in theory we could integrate all information across the generalisation pipeline in one and the same process. We have successfully used this framework with a basic rule-based evaluation model: we isolate conflicting isobaths, triangles and vertices based on legibility requirements and target generalisation operators on those. With this approach we can successfully maintain more of the morphology while still yielding a finely legible chart. Especially at large scale charts the results are promising: narrow channels, pits and bends remain if legibility permits. With smaller scale charts the challenge now is to generalise beyond smoothness. More radical generalisation operators are needed to omit all irrelevant details. However the overall framework using the triangle region graph as integrating mechanism has potential to do so. It is easily extensible due to its modular approach and can incorporate most depth information: from survey accuracy to size of isobaths and even golden sounding selection in the future.

Semantic segmentation (or pixel-level classification) of remotely sensed imagery has shown to be useful for applications in fields as mapping of land cover, object detection, change detection and land-use analysis. Deep learning algorithms called convolutional neural networks (CNNs) have shown to outperform traditional computer vision and machine learning approaches in tackling semantic segmentation tasks. Furthermore, addition of height information (Z) to aerial imagery (RGB) is believed to improve segmentation results. However, discussion remains on the following: to what extent height information adds value; the best way to combine RGB information with height information; and what type of height information can best be used. This study aims to answer these questions. In this research work, the CNN architectures FCN-8s, SegNet, U-Net and FuseNet-SF5 are trained to semantically segment 10 cm resolution true ortho imagery of Haarlem, potentially augmented with height information. The outputted topographic maps contain the classes building, road, water and other. Experiments are conducted that allow for the comparison of 1) models trained on RGB and on RGB-Z, 2) models combining RGB and height information through data fusion and through data stacking, and 3) models trained using different types of absolute and relative height approaches. Performances are compared based on scores on the performance measure (mean) intersection over union (IoU) and through visual assessment of outputted prediction maps. The results indicated that on average segmentation performance improves by approximately 1 percent when absolute height information is added. The class building showed to benefit the most from the addition of height information. Furthermore, extracting features from height information in a separate encoder and fusing these into RGB feature maps, led to a higher overall segmentation quality than when height information is provided as a stacked extra band and processed in the same encoder as the RGB information. Finally, models using relative height delivered a higher quality segmentation than when absolute height approaches were used, especially for large objects. The best performing model; FuseNet-SF5 trained on RGB imagery and pixel-level, relative height, retrieved a mean IoU of 0.8427 and IoUs of 0.8744, 0.7865, 0.9131 and 0.7966 for the classes building, road, water and other respectively. This model was able to correctly classify over 90% of the pixels of 67% of all the objects present in the ground truth. Overall, this study showed that, when considering semantic segmentation of aerial RGB imagery, 1) height information can improve segmentation results, 2) adding height information through data fusion can result in a higher segmentation quality than when data stacking is used, and 3) providing relative height to a network, rather than absolute height, can improve semantic segmentation quality.

The use of 3D city models has increased the last decades due to the evolution of technology. Their use is related to the need for solutions on issues that correspond to the building environment. Modelling aspects, like geometric, semantic information and topology are necessary, in order to provide an integrated result of spatial analysis. The creation of a complete 3D city model that contains all the needed information for an application, is a time consuming and complex process. Furthermore, different 3D city model formats can contain different aspects of the same features. For example, a CityGML model can have the semantic information of a 3D object, while a Multi-View Stereo Mesh model can contain all the geometric and appearance information of the same 3D object.
This thesis is documenting the possible enrichments that can be done bidirectionally between these two different 3D city model formats, when both exist for the same entities. Moreover, it presents the exploration of possible automated enrichment methodologies that can enrich each 3D city model automatically with information from the other one. Finally, an automatic bidirectional enrichment methodology is proposed and is implemented on a testing area, provided in the two different 3D city model formats. This method is based on distance computations between the meshes of the two 3D city models, used to match corresponding features, or part of features, in order to segment semantically the Multi-View Stereo Mesh model (roof, wall, road, terrain, uncertain) and transfer texture to the surfaces of the CityGML model that correspond with the surfaces of the Multi-View Stereo Mesh model.
In addition, distance computations are performed for the validation of the absence of buildings and the shapes of the roofs in the CityGML model, with respect to the information given from the Multi-View Stereo Mesh model. After the implementation of the proposed methodology, it is found that both 3D city model formats can be used for the proposed enrichments of either 3D city model format.
Future improvements are presented for the achievement of using existing information of different formats of 3D city models of the same entities, in order to supplement each 3D city model format with useful information, enhance the performance of spatial analysis and boost the evolution on the use of such models on real life applications.