Alexandre Bry
From Points to Prints
This thesis proposal explores methods for generating accurate building rooftints and footprints from Airborne Laser Scanning (ALS) point clouds, with application to France’s LiDAR HD data. The research investigates how point cloud characteristics influence the extraction of both rooftop and facade boundaries, combining deep learning for semantic segmentation of roof and facade points with geometric processing to reconstruct outlines. The work aims to improve upon existing datasets like BD TOPO, which lacks harmonization and precise georeferencing, and supports France’s initiative to develop a nationwide digital twin with detailed 3D building models.
Supervisors: Hugo Ledoux and Ravi Peters
(company involved: IGN)

Andika Hadi Hutama
Deep Learning Classification Leveraging Multi-Temporal Historical Maps for Enabling Urban Analysis in The Past
This thesis investigates automated urban growth analysis from Dutch historical topographic maps using deep learning. It compares patch-based classification (for nation-wide urban extent extraction) and pixel-based semantic segmentation (for regional building block extraction) applied to multi-temporal Bonnebladen map series. The research aims to evaluate the effectiveness, accuracy, and limitations of each method for quantifying urban expansion patterns over time. Leveraging the MapReader framework and IIIF for data access, the study will implement an iterative training workflow and perform object-based change detection to characterize urban growth types, contributing to scalable computational methods for historical cartographic analysis.
Supervisors: Martijn Meijers and Ana Pereira Roders

Anne den Hartog
Thermally comfortable urban mobility: operationalizing urban microclimate data into a pedestrian routing tool
This thesis proposal aims to develop a fully open-source, scalable pedestrian routing tool that integrates urban microclimate data to improve thermal comfort during heatwaves. Using UMEP (specifically SOLWEIG and URock models), it will generate high-resolution data on mean radiant temperature and wind conditions to assess heat stress. The research focuses on operationalizing this data into a web-based routing application, employing adaptive algorithms that balance route length and heat exposure without predefined weights. The tool will be tested in Rotterdam, emphasizing reproducibility, scalability, and user accessibility to support climate-aware mobility decisions.
Supervisors: Clara Garcia Sanchez, Daniela Maiullari and Lukas Beuster

Carlo Cordes
Wildfire Risk Assessment using Spatiotemporal Transformers
This thesis proposes a spatio-temporal transformer model for global wildfire risk assessment. It aims to integrate multi-modal data—including satellite imagery, meteorological, and topographic datasets—to generate probability maps across multiple time scales. The approach addresses limitations of existing models in capturing long-range dependencies by leveraging self- and cross-attention mechanisms. Key research questions explore temporal scope effects, prediction accuracy over extended lead times, and uncertainty quantification via Bayesian methods. Preliminary work includes data harmonization for CONUS using FIRMS hotspots, ERA5, and land cover data. The study intends to advance wildfire prediction through scalable, data-driven deep learning techniques.
Supervisors: Azarakhsh Rafiee and Justin Schembri

Carmem Félix Aires
A Comparison of Wind Velocity Predictions from Wind-Only and Thermally Coupled OpenFOAM Solvers in Complex Terrain
This thesis proposal investigates the feasibility of using advanced urban microclimate simulations (urbanMicroclimateFOAM) to analyze heat stress in complex, non-flat urban terrain. It compares wind-only and thermally-coupled CFD models against field measurements from Carnegie Mellon University during a heatwave, evaluating accuracy, computational demands, and numerical stability. The study also integrates automated 3D city reconstruction from LiDAR data to assess its suitability for detailed simulations. The goal is to identify when high-fidelity thermal coupling is necessary for reliable urban climate analysis, thereby supporting evidence-based urban planning for heat resilience.
Supervisors: Clara Garcia Sanchez and Miguel Martin Fehlmann
(company involved: Haskoning)

Daan Schlosser
Investigation on the data model requirements for Digital Building Logbooks and Building Renovation Passports
This thesis investigates the data model requirements for Digital Building Logbooks (DBLs) and Building Renovation Passports (BRPs) in the EU. It identifies a research gap due to closed-source, poorly documented, and non-interoperable data models in current initiatives. The study proposes developing an open, formal data model based on geospatial standards, specifically CityGML and its Energy ADE 3.0 extension. Through systematic comparison and mapping of existing initiatives, it aims to define a minimal viable product (MVP), assess the standard’s coverage, and propose necessary extensions to support building renovation data needs and enhance interoperability in the built environment sector.
Supervisors: Giorgio Agugiaro and Hiba Doi

Frederick Auer
Testing and enhancing the Scenario ADE in the context of building performance simulation in Rotterdam
This thesis aims to develop and implement a Scenario Application Domain Extension (ADE) for CityGML 2.0 to manage and compare multiple future energy scenarios for urban districts in Rotterdam. Focusing on the Digitwins4PEDs project, the research will test the ADE’s ability to store simulation results for different climate, renovation, and technology scenarios—such as varying window-to-wall ratios, insulation upgrades, and PV adoption—within a 3DCityDB database, avoiding data redundancy. Using SimStadt for energy demand simulations and Python for analysis, the work will validate the framework’s utility in supporting structured scenario analysis and identifying intervention priorities through unsupervised learning. The goal is to enhance decision-making for Positive Energy Districts and the Dutch energy transition.
Supervisors: Camilo León Sánchez and Giorgio Agugiaro

Heiko Rotteveel
Automatic water (level) detection using ICESat-2 measurements
This thesis proposes an automated method to detect inland water bodies and estimate their surface elevation using ICESat‑2’s raw photon data (ATL03 product). It aims to overcome limitations of existing water masks and official products like ATL13, which rely on predefined masks and miss smaller or dynamic water bodies. By combining multiple water‑photon interaction features—such as photon density, standard deviation, after‑pulse patterns, bottom reflectance, and surface flatness—in a Random Forest classifier, the approach seeks to globally identify water stretches without external masks. The workflow will produce georeferenced water‑body line segments with derived elevations, validated against datasets like HydroLAKES, offering a scalable, real‑time solution for monitoring inland water resources.
Supervisors: Maarten Pronk and Hugo Ledoux

Hongyu Ye
From Point Clouds to Wind Risk: A ForestGALES–CFD Framework for Tree-Level Urban Wind Damage Assessment
This thesis proposes a physics-based framework to assess wind damage risk to individual urban trees in Dutch coastal cities. It integrates three key components: detailed tree structure from airborne LiDAR point clouds, computational fluid dynamics (CFD) simulations to resolve complex urban wind fields, and the mechanistic ForestGALES/TMC model to calculate tree failure probability. The research aims to translate point-cloud-derived morphology and species information into aerodynamic and structural parameters, extend ForestGALES for urban broadleaf species, and couple it with CFD outputs to produce tree-level, interpretable risk maps. This unified workflow seeks to identify vulnerable trees by quantifying the joint influence of tree architecture and street-canyon geometry on wind loading, moving beyond statistical methods to support targeted urban forest management.
Supervisors: Azarakhsh Rafiee and Xuanchen Zhou

Julia Pille
Seabed Fingerprinting
This thesis proposes a seabed fingerprinting workflow for GNSS-denied underwater navigation. It develops a method to extract compact terrain fingerprints from high-resolution multibeam echosounder (MBES) data, stored in an indexed database. During real-time operation, onboard MBES measurements are matched against these fingerprints within an INS-defined search region to correct inertial drift. The approach aims to enable accurate, infrastructure-free navigation for both crewed and uncrewed marine vehicles. Preliminary results using EMODnet data show promising position estimation. The research addresses knowledge gaps in terrain-aided navigation and focuses on applicability in the Dutch North Sea.
Supervisors: Liangliang Nan and Roderik Lindenbergh
(company involved: CGI)

Lars van Blokland
Air temperature estimation through thermal satellite imagery using uncertainty-integrated transformers
This thesis proposes a deep learning model to estimate high-resolution (70m) air temperature across the Netherlands using thermal satellite imagery (ECOSTRESS LST) and ancillary data. The core research explores how well a spatio-temporal transformer architecture can leverage land surface temperature and its associated uncertainty to predict near-surface air temperature. The model will integrate multiple data sources—including wind, vegetation, land use, and sky view factor—while explicitly incorporating input uncertainty to provide confidence estimates for its predictions. The study will evaluate the model's performance across seasons and urban environments, and investigate architectural choices like cross-attention mechanisms and temporal embedding strategies. The goal is to produce a fine-scale, uncertainty-aware temperature map to better monitor urban heat islands and support climate adaptation.
Supervisors: Azarakhsh Rafiee and Roderik Lindenbergh

Luc Jonker
Automated Generation of Tree-Aware Urban Shade Maps using Deep Learning
This thesis proposes using deep learning, specifically a pix2pix-style conditional generative adversarial network, to automatically generate urban shade maps that include the shading effects of trees. Current geometric methods are computationally expensive. The research aims to reproduce tree-aware shadow maps comparable to those from UMEP with high accuracy but significantly faster inference times. The methodology involves data preparation from existing urban DSM and shadow map tiles, iterative model design, and performance evaluation using metrics like RMSE and SSIM. The goal is to demonstrate the viability of deep learning for efficient, vegetation-inclusive urban shade simulation to support heat mitigation and urban planning.
Supervisors: Hugo Ledoux and Lukas Beuster

Michel Beeren
Optimizing 3D Alpha Wrapping for CFD Applications; Repairing and Simplifying 3D Building Models
This thesis proposes extending CGAL's 3D Alpha Wrap algorithm to generate valid, watertight, and CFD-ready building meshes that better preserve sharp architectural features while minimizing unnecessary mesh complexity. The core objective is to develop a method that uses locally adaptive α parameters—smaller values near sharp/small features for detail, larger values in smooth regions to reduce triangle count—guided by feature detection from techniques like adaptive octree refinement or the Medial Axis Transform. The workflow includes pre-processing for feature detection, in-algorithm modification for local α-adaptation, and post-processing to sharpen rounded edges. The improved algorithm aims to serve as a robust fallback in automated reconstruction pipelines like City4CFD, producing ISO 19107-compliant meshes efficiently without manual repair.
Supervisors: Clara Garcia Sanchez and Hugo Ledoux

Ming Chieh Hu
Gaussian Splatting to Piecewise-Planar Surfaces For 3D Building Reconstruction
This thesis proposes a novel pipeline for generating watertight, piecewise-planar 3D building models using 3D Gaussian Splatting (3DGS). It addresses the limitations of current methods that produce dense, unstructured meshes unsuitable for urban analysis applications. The research integrates semantic segmentation (e.g., Segment Anything Model) and geometric constraints to filter non-structural clutter and enforce planar priors during Gaussian optimization. The resulting Gaussian representation is then fed into a piecewise-planar surface reconstruction algorithm to extract lightweight, manifold building geometry. The method will be evaluated against state-of-the-art baselines across multiple datasets, aiming to bridge the gap between photogrammetric efficiency and the structural utility required for downstream urban analysis tasks.
Supervisors: Liangliang Nan and Michael Weinmann

Neelabh Singh
Identifying Sidewalks From Crowdsourced SVI for OSM Enrichment
This thesis proposes an automated workflow to extract sidewalk geometries from open, crowdsourced Street View Imagery (e.g., Mapillary) for enriching OpenStreetMap. Using semantic segmentation models (like DINOv3 and SAM) combined with monocular depth estimation (Depth Anything V2), the research aims to transform 2D pixel masks into georeferenced vector polygons or centerlines with estimated widths. The goal is to bridge the critical data gap in pedestrian infrastructure mapping by providing a scalable, license-compliant method to generate reliable sidewalk datasets suitable for accessibility assessment and network routing.
Supervisors: Hugo Ledoux and Lukas Beuster

Sara Brakelé
Extracting building typology parameters from 3D city models for earthquake risk assessment
This thesis proposal investigates how 3D city models can be used to automatically extract building typology parameters for earthquake risk assessment. Focusing on the seismically active Groningen region in the Netherlands, the research aims to determine which structural characteristics (e.g., dimensions, shape) can be derived from 3D geometry. Using a validation dataset of about 300 buildings, it will apply statistical tests and machine learning to predict non-visible attributes like construction type and foundation. The goal is to develop a methodology that enhances individual building-level seismic vulnerability analysis, offering a faster, scalable alternative to traditional on-site surveys.
Supervisors: Jantien Stoter and Giorgia Giardina

Segher ter Braak
Downscaling ECOSTRESS Land Surface Temperature using Deep Learning
This thesis proposes a physics-guided deep learning framework to downscale ECOSTRESS land surface temperature (LST) data from ~70 m to 10 m resolution in heterogeneous Dutch landscapes. It integrates Sentinel‑2 imagery, land cover, and topography into a Swin Transformer‑based U‑Net architecture. The approach uses weakly supervised learning with physics‑informed loss functions—such as aggregation consistency and edge‑aware smoothness—to ensure physical plausibility and coherence across scales. The goal is to produce high‑resolution, physically consistent LST estimates suitable for detailed urban and agricultural analyses, addressing the limitations of traditional downscaling methods in complex environments.
Supervisors: Azarakhsh Rafiee and Roderik Lindenbergh

Sue Wang
Action-Aware Indoor Navigation with Agent-Guided Adaptive Voxelization and Scenario-Aware Replanning
This thesis proposes an action-aware indoor navigation system that adaptively voxelizes building models to balance path accuracy and computational cost. It introduces “probe-agent” sampling to refine only high-traffic regions and integrates semantic constraints (e.g., doors, stairs) to ensure realistic routing. The system supports scenario-aware replanning for dynamic changes like closures and exports results in standardized formats (IndoorGML/GeoJSON) for interoperability and visualization. The goal is to produce a reproducible, efficient navigation pipeline suitable for applications in accessibility, evacuation planning, and facility management.
Supervisors: Ken Arroyo Ohori and Martijn Meijers

Teo Mingjie
A Topology-Aware Multi-3DGS Viewer
This thesis proposes developing a "Topology-Aware Multi-3DGS Viewer" to enable seamless navigation between multiple 3D Gaussian Splatting (3DGS) scenes of indoor spaces. It identifies a research gap in linking discrete, high-fidelity 3DGS representations—which are computationally heavy as single scenes—using a topological model based on the IndoorGML standard. The project will extract geometric boundaries and door locations from 3DGS data to build a navigation graph, connect scenes via spatial transforms, and implement a web-based viewer that enforces these topological rules for immersive, context-aware exploration of multi-room indoor environments.
Supervisors: Edward Verbree and Martijn Meijers

Vincent Vanderheeren
Pillar of Morphology - Enhancing point-based mathematical morphology for processing applications in heritage point clouds
This thesis focuses on enhancing point-based mathematical morphology for direct processing of heritage point clouds. It aims to improve the existing algorithm by Balado et al. (2020) to address limitations in speed, variable density, orientation, and attribute preservation. The improved method will enable applications such as segmentation, object detection, and gap filling directly on unstructured point cloud data, aligning with the rich point cloud paradigm. The performance and accuracy of the enhanced algorithm will be evaluated on heritage datasets (e.g., ArCH dataset) and compared to existing methods, including machine learning and surface-based approaches, to demonstrate its viability for cultural heritage applications.
Supervisors: Martijn Meijers and Ken Arroyo Ohori

Xinya Bi
Multi-view Semantic Pruning for 3D Building Reconstruction using Gaussian Splatting
This thesis proposes a multi-view semantic pruning framework to improve 3D building reconstruction using 3D Gaussian Splatting (3DGS). It addresses the problem of floating geometric artefacts—spurious primitives that lack physical correspondence—by leveraging pre-trained vision foundation models (DINOv2) for offline semantic feature extraction. Instead of learning semantics during training, the method treats semantic identity as a physical invariant, computing multi-view semantic consistency scores to distinguish valid surfaces from artefacts in a post-processing stage. Low-confidence Gaussians are pruned, enhancing geometric accuracy while preserving photometric fidelity, thereby making 3DGS more reliable for precision-critical architectural reconstruction.
Supervisors: Liangliang Nan and Martijn Meijers

Yair Roorda
Pointcloud based intervisibility analysis for military airborne mission planning taking into account semi-transparent vegetation
This thesis proposes a novel method for 3D visibility analysis to support military airborne mission planning. It addresses key limitations in current approaches by processing LiDAR point clouds directly—rather than using intermediate models—and incorporating semi-transparent vegetation into non-binary, probabilistic line-of-sight calculations. The research aims to develop a fully 3D pipeline for querying, segmenting vegetation, and computing visibility using the Dutch AHN dataset. The outcome will be an improved, data-driven planning tool that enhances safety and operational effectiveness by providing more accurate visibility assessments in vegetated and complex terrains.
Supervisors: Peter van Oosterom and Martijn Meijers
(company involved: Nederlands Lucht- en Ruimtevaartcentrum)