Integrating GeoAI and Multisource Remote Sensing with Cadastral Data for Greenbelt Compliance Screening

Greenbelt monitoring requires both the detection of physical change and its interpretation against parcel and administrative records. This study develops a multistage GeoAI–GIS workflow for that purpose and examines its application in Siheung, Republic of Korea. Satellite, aerial, and unmanned aerial vehicle imagery support regional change screening and detailed mapping, while cadastral attributes and municipal registers provide the context for selecting parcels for review. A metric-learning change module trained with triplet loss, a YOLOv8-based structure segmentation module, and a SegFormer land-cover module address separate analytical tasks. Their internal validation F1-scores were 87.3%, 88.4%, and 84.8%, respectively. Class-specific results showed stronger performance for forests, buildings, and greenhouses than for installations, stockpiles, farmland, and bare land. Spatial integration converted the predictions into parcel-linked geometries, attributes, and reasons for administrative examination. For the building- and greenhouse-related comparison, the workflow produced 2486 candidate parcels, whereas the manual aerial-photo interpretation dataset contained 1836 cases. The difference of 650 records, or 35.4%, describes candidate volume rather than additional confirmed violations or an increase in measured recall. The study demonstrates a practical connection between image analysis and parcel-based review, and identifies class reliability, spatial dependence, and the distinction between physical change and legal status as central to its interpretation. Automated screening supports the preparation of evidence; permit assessment and decisions about enforcement remain administrative responsibilities.