Developing a Transfer Learning-Based Deep Learning Model for Visible Cadastral Boundary Delineation in Nepal

Abstract

This study investigates the potential of transfer learning-based deep learning models to complement ground-based cadastral data acquisition in Nepal. Multiple transfer learning strategies were evaluated to extract visible cadastral boundaries. The evaluation was conducted using 30 cm-resolution Pleiades Neo 4 satellite imagery acquired over a heterogeneous study area comprising predominantly agricultural land and sparsely built, semi-urban settlements. A pre-trained U-Net model was used as a classifier, feature extractor, and fine-tuning target, and its performance was compared with a U-Net trained from scratch and a U-Net with a ResNet-101 encoder (ResNet-101 U-Net) pre-trained on ImageNet. Among the experiments, the fine-tuned U-Net model pretrained on an agricultural dataset achieved the best performance, attaining an F1 score of 0.56 when evaluated against digitized visible land boundaries. The predicted boundaries were further compared with the official cadastral boundaries, applying multiple buffer distances to quantify spatial accuracy. At a buffer distance of 2 m, the highest F1 score obtained was 44.81%, reflecting both spatial discrepancies in the 1600-hectare study area’s official cadastral datasets and the challenges associated with delineating legally defined visually indistinct boundaries. The study highlights the potential of transfer learning-based models for visible cadastral boundary delineation tasks, while acknowledging their limitations in delineating invisible legal cadastral boundaries, as corroborated by stakeholder feedback and field observations.