This thesis explores the potential of AI-based building detection to support building database maintenance and regulatory supervision in Upplands Väsby municipality. By combining stakeholder input from surveys and semi-structured interviews with technical evaluations of three segmentation models—SAM2, LangSAM, and Mask R-CNN—the study assesses both organizational readiness and model performance. Input data included aerial imagery and LiDAR point clouds, and various geometric regularization techniques were tested. Performance was measured using mean IoU, precision, recall, and F1-score.

Findings highlight both organizational and technical challenges. While stakeholders expressed interest in AI adoption, concerns around GDPR compliance, limited internal expertise, resource constraints, and inconsistent data quality emerged as significant barriers. From a technical perspective, none of the models consistently achieved the precision and reliability needed for fully automated use. Performance varied based on urban context, input data, and regularization methods. Mask R-CNN showed the best overall performance, particularly when considering scalability, but still suffered from low recall. SAM2 offered high segmentation precision on individual buildings with point-prompts but lacked scalability. LangSAM was more scalable through text-prompts but struggled with consistency and recall at larger scales. For change detection, the best results were obtained by combining outputs from LangSAM and Mask R-CNN.

Overall, the evaluated models are more suitable for lower-precision internal tasks, such as general land use mapping. For high-precision applications, further model refinement, improved data quality, and seamless system integration are essential. With targeted development, AI segmentation tools like Mask R-CNN and SAM2 could play a valuable role in future municipal workflows.