Cadastral Boundary Delineation Using Transfer Learning Approach

Abstract

The adoption of Fit-for-Purpose Land Administration aims at providing a flexible and low-cost alternative for cadastral boundary delineation that emphasizes visible boundaries derived from earth observation imagery. Recent advances show that deep learning can support the automatic delineation of visible cadastral boundaries. However, applications remain constrained by limited, small-scale datasets and weak understanding of which cadastral characteristics are identifiable. This study investigates the use of a transfer learning-based U-Net model to extract visible cadastral boundaries from high-resolution UAV imagery in a study area covering 200 hectares in Banepa Municipality, Nepal. The U-shaped model, pre-trained on agricultural field boundaries in Cambodia and Vietnam, and fine-tuned on 30 cm satellite imagery of the Terai region, was applied as a fixed classifier to UAV orthophotos at 3 cm, 15 cm, and 30 cm resolutions. A patch-based pipeline splits imagery into 512 X 512 tiles, predicts boundary masks, and converts them into vector polylines via morphological operations, skeletonization, and graph-based line extraction. Results indicate that coarser resolutions produce more generalized and cadastral-like parcel shapes, while finer resolutions capture detailed man-made features but suffer from over-segmentation. The 15 cm resolution yields the best overall performance, particularly in sub-urban areas, though the model struggles to form closed and complete polylines. Predicted boundaries align closely with visible features such as roads, fences, and cultivation edges. The outcomes suggest that deep learning-based workflows can generate preliminary boundary maps that can assist image-based cadastral mapping, reducing reliance on manual digitization and field visits, thus supporting the concept of fit-for-purpose approach in countries with incomplete and outdated cadastres.