AI-Enhanced Land Parcel Boundary Detection Using High-Resolution Satellite Imagery

Thesis Abstract

Accurate delineation of land parcel boundaries remains a critical challenge in land administration, urban planning, and resource management due to the limitations of conventional manual mapping techniques, which are often time-consuming, prone to human error, and insufficient for large-scale areas. The rapid proliferation of high-resolution satellite imagery offers unprecedented opportunities for automating boundary detection processes; however, existing methods frequently suffer from inaccuracies caused by spectral heterogeneity, shadow effects, and complex terrain features. This study aims to develop an innovative AI-driven framework that enhances the precision and efficiency of land parcel boundary detection using high-resolution satellite images. Specifically, the research seeks to integrate deep learning techniques, particularly convolutional neural networks (CNNs), with advanced image pre-processing and segmentation algorithms to improve boundary recognition accuracy. The study adopts a quantitative research design, employing a case study approach within a selected urban municipality covering approximately 500 square kilometers. A stratified random sampling technique was used to select 30 satellite images, each covering distinct land-use zones such as residential, commercial, and recreational areas, ensuring representativeness across diverse landscape conditions. Data collection involved acquiring high-resolution multispectral satellite images at 0.3-meter spatial resolution from a reputable remote sensing data provider, supplemented by existing cadastral boundary datasets for validation purposes. The core methodology includes developing a custom deep learning model trained on annotated datasets, with ground truth boundaries established through field verification and expert digitization. The model’s performance was evaluated using several metrics, notably the Intersection over Union (IoU), boundary displacement error, and overall accuracy, analyzed via statistical techniques such as regression analysis and paired t-tests to compare AI-detected boundaries with existing manual boundaries. The expected findings indicate that the developed AI model will considerably outperform traditional boundary detection algorithms, achieving a mean IoU score exceeding 0.85 and reducing boundary displacement errors to less than 1.2 meters in 90% of test cases. It is also anticipated that the model will demonstrate robustness across different land-use types and illumination conditions, thus establishing its viability for large-scale application. The study contributes to existing knowledge by demonstrating the applicability of deep learning in cadastral boundary extraction, providing a replicable framework that enhances objectivity, accuracy, and operational efficiency in land parcel mapping. It further extends the theoretical understanding of AI integration with remote sensing for land management and offers practical guidelines for implementing automated boundary detection systems in various geographical contexts. Main conclusions suggest that AI-based techniques, when properly calibrated and validated, can significantly improve cadastral boundary mapping processes, thus reducing dependency on manual surveys, decreasing costs, and expediting land registration workflows. The research recommends the adoption of AI-enhanced boundary detection tools by government cadastral agencies and private land developers to facilitate transparent and reliable land administration. Future studies are suggested to explore the integration of multi-source geospatial data, including LiDAR and temporal imagery, to further refine model accuracy, as well as to investigate machine learning explainability methods to improve user acceptance and operational interpretability of automated boundary detection outputs. This study underscores the transformative potential of artificial intelligence in modernizing land information systems and advancing sustainable land governance.