From pixels to vectorized cadastral boundaries: Deep learning-based automated delineation of property boundaries in the Netherlands

Abstract

Cadastral mapping remains one of the most resource-intensive components of land administration, while large parts of the world still lack complete and up-to-date cadastral information. In this context, remote sensing and deep learning can support the rapid production of initial cadastral boundaries for semi-automated mapping workflows. We present CadNet , a deep learning architecture for extracting initial cadastral boundaries from very high-resolution aerial imagery. The model is evaluated at the national scale in the Netherlands using a large and geographically diverse subset of the CadastreVision benchmark and is compared with eight baseline networks under identical experimental conditions. CadNet consistently achieves the highest F1 scores across all evaluated landscapes (mixed: 0.502, rural: 0.468, and peri-urban: 0.556), outperforming the next-best models by 1.3%–1.7%, and produces the least fragmented vector output, reducing the connected-component ratio by 8%–26% relative to the best-performing baseline. Qualitative results further show more continuous boundary representations and better recovery of partially occluded visible boundaries. These findings demonstrate that connectivity-aware, multi-scale deep learning can improve the extraction of initial cadastral boundaries over large areas from aerial imagery.