Up-to-date 3D data is essential for urban planning, building inspections, and monitoring changes in the built environment. While 2D aerial imagery is widely used, it lacks height information and is sensitive to shadows and seasonal effects. In contrast, 3D point clouds provide detailed spatial information and enables better interpretation.

This thesis presents a method for detecting structural building changes using bitemporal airborne laser scanning (ALS) data from the national height model of the Netherlands (AHN) and the Rotterdam municipality. These datasets are pre-aligned in the stelsel van de rijksdriehoeksmeting (RD)-normaal Amsterdams peil (NAP) coordinate system and include building classifications, which allows the focus of this research to be placed directly on detecting change.

Comparing point clouds from different time epochs is challenging due to differences in density, noise, occlusion, and scan geometry. To address this, a random forest (RF)-based classifier is trained on synthetically generated urban scenes that simulate realistic change scenarios. These synthetic scenes are made with different scanning parameters, incorporating diversity in the training dataset. A certainty index is introduced that combines the model’s probability output with occlusion visibility across both epochs, providing a confidence measure for each prediction.

The method is applied to real AHN and Rotterdam datasets. Since no labelled ground truth is available, results are evaluated visually. The method successfully identifies structural changes such as dormers and extensions, and also detects moved or temporary objects such as sunshades or picnic tables. When combined with aerial imagery, the approach helps distinguish static from dynamic changes.

This work is innovative in its integration of occlusion-aware certainty scoring, visual certainty feedback, and the automated generation of synthetic training data for change detection