GeoAI-Based Land-Use Compliance Monitoring for Land Rights Administration Using Sentinel-2 and Orthophotos

Large-scale monitoring of land rights remains constrained by substantial time requirements, limited spatial coverage, and considerable human resource demands associated with conventional image interpretation and field verification. This study develops an operational Geospatial Artificial Intelligence (GeoAI) workflow for indicative monitoring of land-use compliance by integrating semantic segmentation of land cover within registered land rights (HAT) boundaries with spatial planning information. U-Net and SF-Net architectures were evaluated for temporal monitoring using two models: M3 based on 10 m Sentinel-2 imagery from 2017–2023 and M4 based on 0.07 m orthophotos from 2022–2023. Data augmentation increased the number of M3 samples from 445 to 3560 and M4 samples from 1106 to 8848. The U-Net M3 model achieved an F1-score of 0.961 and an IoU of 0.927, whereas the SF-Net M4 model achieved an F1-score of 0.969 and an IoU of 0.940. At Keera Plantation, the 28–29% difference in estimated oil palm area between M3 and field verification reflects inter-method variation in a heterogeneous landscape rather than model error alone. M3 processed the entire 680,380.99 ha administrative area of East Luwu Regency in only 4 min 31 s to assess operational scalability and monitoring efficiency. The proposed GeoAI workflow is effective for large-scale initial screening and prioritization. However, GeoAI outputs are intended as decision-support evidence and do not replace field verification or legally accountable administrative decision-making.