Accelerating cadastral mapping with CadastreVision and CadNet

The lack of a large, open benchmark dataset that combines Earth observation imagery with accurate cadastral references has been holding back progress in the land administration community. This article presents a dataset and deep learning model developed in the Netherlands to meet this need. Comparison against baseline architectures shows that this approach delivers higher accuracy and improved boundary continuity, with boundary lines that are cleaner, better connected and more closely aligned with the true cadastral boundaries.

Secure land rights are vital to combating poverty, hunger and inequality, and are central to achieving several United Nations Sustainable Development Goals (SDGs). While high-income countries typically have robust systems for recording land ownership, many people in lower-income regions lack access to formal land registration. It’s estimated that 70 to 75% of the global population do not have official land rights documentation. The fit-for-purpose land administration (FFPLA) approach was introduced by FIG and the World Bank to address this challenge. It leverages that many property boundaries correspond to visible features such as fences, roads or buildings – features often identifiable in aerial or satellite imagery. As Earth observation (EO) data becomes increasingly accessible and suitable for automated analysis, this approach has gained momentum as a scalable solution for mapping land rights remotely.

Recent deep learning (DL) developments have shown strong potential for cadastral boundary extraction from EO imagery. These methods reduce the number of manual interactions needed to digitize boundaries, significantly improving speed and efficiency. However, most current studies rely on small, localized datasets. This limits both the scalability of these methods and their generalization to new geographic areas. At the same time, the field has yet to fully adopt cutting-edge artificial intelligence (AI) models, such as transformer-based architectures, which require large and diverse datasets to perform effectively. Without access to such data, developing and benchmarking next-generation DL methods for cadastral mapping remains constrained.

What’s missing and what’s next?

One of the main bottlenecks is the lack of a large, open benchmark dataset that combines EO imagery with accurate cadastral references. This contrasts with other remote sensing domains, like building detection, land cover classification or road mapping, where benchmark datasets have been pivotal in driving progress. A standardized, publicly available dataset for cadastral boundary extraction would allow the community to test and compare models consistently, helping to bring land administration into the digital era. To push the limits of what’s possible in this field, we must explore how far DL can go when powered by benchmark data and advanced model architectures. Testing these models in a controlled, standardized environment is key to understanding their full potential and developing reliable, scalable tools for real-world applications.

Introducing CadastreVision

To meet this need, the authors developed CadastreVision, a large-scale benchmark dataset that combines Dutch cadastral reference data with high-quality aerial and satellite imagery. Designed to support the training and testing of deep learning models, CadastreVision offers a standardized resource that can help accelerate cadastral mapping in data-rich and data-scarce regions. It also supports research into model transferability, evaluating whether models trained on Dutch data can be adapted to other parts of the world. Notably, CadastreVision includes a detailed classification of cadastral boundaries as visible or non-visible in the imagery. This distinction is critical, as most DL models today are limited to detecting only visible boundaries. By enabling separate evaluation on visible and invisible boundaries, the dataset supports more nuanced and realistic assessments of model performance.

Building the dataset

The Netherlands was divided into 10x10km tiles, from which 90 were selected to reflect the country’s diverse cultural landscapes, including areas shaped by human activity such as field patterns, dikes and waterways (see Figure 1). For each tile, RGB imagery was collected from a variety of EO sources: high-resolution aerial photos (8 and 25cm), available via Publieke Dienstverlening op de Kaart (PDOK), and satellite imagery from SuperView (50cm), PlanetScope (approximately 3m) and Sentinel-2 (10m). While aerial images were ready to use, due to cloud cover and image extent limitations the SuperView satellite data required manual selection, mosaicking and clipping to align with the vector tiles. PlanetScope and Sentinel-2 images were extracted from cloud-free mosaics provided by the Netherlands Space Office. For each EO image, a corresponding binary mask of cadastral boundaries was created using Basisregistratie Kadaster (BRK) parcel data from spring 2022. These boundaries were buffered by 0.4m to account for positional uncertainty and rasterized to match the spatial resolution of the imagery.

Visible vs non-visible boundaries

CadastreVision identifies which cadastral boundaries are visible in the EO data and which are not. This was done using open Dutch geospatial datasets (BGT, BRT, BRP and BAG) to extract topographic features, which were then compared with the BRK cadastral boundaries. Boundaries overlapping with topographic objects were classified as visible; the rest were classified as non-visible. This allows researchers to independently assess model accuracy on both types, which is crucial for fair evaluation and practical use in land administration systems.

A full pipeline run across all 90 tiles showed that 72.2% of the total cadastral boundary length could be matched to visible topographic features. This demonstrates the viability of nationally automated visibility classification. A visual example is shown in Figure 1, where visible boundaries (green) and non-visible boundaries (red) are overlaid on aerial imagery. Thematic classifications of the matched segments, such as vegetation, water, built-up area and buildings, are also shown in the image. Analysis of the thematic classification reveals that more than 53% of the matched segments were matched with vegetation, 35% with water, 24% with road, 10% with buildings and 33% with others (due to multi-label classification, the percentages do not add up to 100%).

A new model: CadNet

Building on this foundation, the authors developed a U-shaped deep learning model – a convolutional encoder–decoder network widely used in image segmentation – to automatically extract cadastral boundaries from high-resolution imagery. To benchmark its performance, two baseline models were established using proven encoder-decoder architectures: one with a transformer-based encoder and the other with a ResNet backbone, both pre-trained on large-scale datasets. To enhance boundary connectivity, they introduced a method that examines neighbouring pixels in multiple directions, generating connectivity maps that help the model better capture how boundary segments link together. The authors also designed a novel multi-stage labelling system that integrates contextual information at different decoding stages. This allows the model to progressively refine its predictions, focusing on the most relevant features and improving the continuity of boundary lines.

The final model outputs both a segmentation mask and connectivity maps, which are combined to produce a clean, one-pixel-wide boundary line. This line is then converted into vector format, ready for use in cadastral mapping systems.

Results

Using a rural subset of the CadastreVision benchmark dataset, the model was evaluated against the two baseline architectures. The results show that the approach consistently outperforms both baselines across standard segmentation metrics, delivering higher accuracy and improved boundary continuity. Visual inspections confirm that the predicted lines are more complete and exhibit fewer breaks, particularly in areas with complex natural features such as vegetation edges or irregular field boundaries (see Figure 3). Compared to the baselines, the model produces boundary lines that are cleaner, better connected and more closely aligned with the true cadastral boundaries, making them more suitable for direct integration into mapping workflows without extensive manual correction.

Future work and applications

Building on the results obtained with CadNet and the CadastreVision dataset, several ways for further research and practical implementation can be pursued to enhance cadastral boundary extraction and mapping.

• Large-scale experiments
A natural next step is to perform experiments using the entire CadastreVision dataset. While previous work has focused on subsets, running the model on the full dataset will allow a more comprehensive evaluation of its performance and scalability, providing insights into strengths and limitations across diverse landscape types.

• Transferability assessment
To evaluate how well the model generalizes beyond the training regions, case study areas should be selected to test transferability. Assessing transferability ensures that the model can perform reliably in different geographic, social and environmental contexts. This is particularly important for supporting land administration in regions where cadastral data is limited or incomplete, contributing to equitable access to land information. By improving model applicability across diverse settings, this research directly supports several Sustainable Development Goals, including SDG1 (No Poverty), SDG2 (Zero Hunger), SDG5 (Gender Equality), SDG11 (Sustainable Cities and Communities) and SDG16 (Peace, Justice, and Strong Institutions), by enabling better land governance, resource management and legal recognition of land rights.

• Benchmark expansion
Expanding the benchmark dataset with more diverse imagery and cadastral reference data will further improve model robustness. Including a wider variety of rural and urban landscapes, as well as different geographic and climatic conditions, will help ensure the model can generalize globally.

• Human-in-the-loop integration
To improve accuracy and usability in real-world applications, a human-in-the-loop approach should be developed. In this framework, the model proposes candidate boundaries that human operators can review, refine and approve, combining automated efficiency with expert validation.

• Open-source GIS toolbox
Lastly, an open-source GIS toolbox should be created to implement these methods, making them accessible to practitioners, researchers and local authorities. This toolbox would integrate the proposed model, human-in-the-loop workflows and domain transfer methods, providing a practical platform for large-scale cadastral mapping and updating.
By pursuing these directions, future work can advance automated cadastral mapping, improve model transferability and support decision-making in regions with limited cadastral data, ultimately contributing to more efficient land administration and management practices worldwide.

Further reading

- PDOK, https://www.pdok.nl/datasets

- Crommelinck, S., Koeva, M., Yang, M.Y., Vosselman, G., 2019. Application of Deep Learning for Delineation of Visible Cadastral Boundaries from Remote Sensing Imagery. Remote Sensing 11.

- Enemark, S., Clifford Bell, K., Lemmen, C., McLaren, R., 2014. Fit-For-Purpose Land Administration. Technical Report. International Federation of Surveyors (FIG).

- Grift, J., Persello, C., Koeva, M., 2023. Cadastral Boundary Delineation using Deep Learning and Remote Sensing Imagery: State of the Art and Future Developments, in: FIG Working Week 2023.

- Grift, J., Persello, C., Koeva, M., 2024. CadastreVision: A Benchmark Dataset for Cadastral Boundary Delineation from Multi-resolution Earth Observation Images. ISPRS Journal of Photogrammetry and Remote Sensing 217, 91-100, https://www.sciencedirect.com/science/article/pii/S0924271624003150

- GitHub CadastreVision: https://github.com/jeroengrift/cadastrevision

- GitHub CadNet: https://github.com/jeroengrift/cadnet