Cadastral map next

Land Registry

Using artificial intelligence, we rebuild the cadastral map. For this, we use scans of original historical fieldworks.

Last change on 4th of June 2024, at 11:14 (CET) | Publication Standard 1.0

• Publication category: Other algorithms
• Impact assessment: Field not filled in.
• Status: In use

General information

Theme

• Space and Infrastructure
• Work
• Living

Begin date

Field not filled in.

Contact information

https://formulieren.kadaster.nl/contact_kadaster_algoritmeregister

Responsible use

Goal and impact

The Cadastral map next is an improved Cadastral map. This map shows the boundaries as much as possible as they are in reality. We use the original measurement data, i.e. old fieldwork. We also use algorithms and artificial intelligence. This allows us to automate the process. That way we save costs and provide a more accurate Cadastral map: the Cadastral map next.

Considerations

Field not filled in.

Human intervention

Human on the loop. This means that AI is applied in an automatic process. Are there any deviations? Then humans check the result and adjust it if necessary.

Risk management

Minimal risk

Legal basis

Section 74 Land Registry Act states that the Land Registry may investigate whether data from the basic land registry are correct and complete. Article 23 Land Registry Decree states that a renewal survey may be carried out if data is found to be insufficiently correct and complete. Thus, there is no legal basis or foundation for the Cadastral Map Next yet. That is why we want to amend Article 23. This will allow us to go one step further and renew the Cadastral Map.

Operations

Data

We have used a set of field works that is characteristic of our archive. That set consists of different regions and time periods. Then there is a set for OCR: converting handwritten texts into information that the computer can read. This set consists of about 200,000 field works. The end-to-end set for line and object detection consists of 3,000 fieldworks.

Technical design

The algorithm uses separate neural networks for OCR, line detection and object detection: * for OCR, we use an NN, which consists of CNN and RNN layers * line detection is done based on a modified UNET * object detection is done with a MaskRCNN based on the SWIN transformer architecture These models are trained based on fieldwork (components) and annotations. Annotations are notes with a comment or explanation. We split the available dataset into train, validation and test sets. We augment the train data during training to create a versatile train set. During training, we determine the F-score on the validation set after each iteration. As soon as the performance on the validation set no longer improves, we stop training. Then we determine the F-score on the test-set. Is this F-score better than previous training sessions? Then we use the weights of the neural network for prediction.