Detection of anomalies in land usage using aerial photography, geographic information system and machine learning : Identifying unlawful land use and unwanted constructions relevant for municipal city monitoring

Sammanfattning

During recent years, land management has become a crucial issue as cities and natural environments face pressures from unlawful or unintended use of land. Identifying and monitoring these activities at scale is challenging and requires sophisticated tools. This thesis explores how ML in conjunction with aerial imaging and GIS is able to automate the detection of unlawful land usage in the form of illegally constructed or expanded buildings, unapproved placement of piers, and settlements. To address this problem, YOLOv11 OD models are trained and evaluated on high spatial resolution aerial imagery with GIS data of buildings from Botkyrka municipality, and manually labeled piers and settlement-resembling objects. A series of experiments were conducted to investigate how tile size and GSD of the aerial images affect detection performance, training time, and memory requirements. Test datasets including different sets of objects and labels are utilized to uncover the influence of class granularity on detection performance. The possibility and plausibility of detecting building expansions is examined through the use of varying threshold parameters. The results show that a spatial resolution of 0.13 m/pixel yields the best detection performance, while also exhibiting signs of diminishing returns such that a higher GSD can be used with restricted computational resources. A tile size of around 25-50 m shows the best performances, limiting the number of pixels and complexity of samples, while allowing for the buildings to be discernible within the images. Division of classes into subclasses, or utilizing cross-class detection did not improve performance consistently. Detection of piers showed the best performance, shortly followed by buildings, while the settlement objects, being the smallest objects, struggled. While the detection of building expansions prove to be possible, it is not deemed to be at a sufficient level, and other methods are therefore proposed for future research, such as the utilization of oriented BBs or image segmentation. Overall, the study demonstrates that an ML-based approach can provide a strong basis of automation with a human-in-the-loop in early-stage landuse monitoring. Future work could extend the system for improved expansion detection, and integrate it into existing GIS-based workflows for systems used in production.