What is the reason for this project?

'We started this project together with professor of Historical Cartography Bram Vannieuwehuyze and Dr Melvin Wever, because researchers increasingly need image recognition on digitised maps. This allows them to use images on maps for their research more quickly, accurately and efficiently.

As a result, the Allard Pierson wishes to add this kind of image recognition to its digitised maps. The Library is also focusing on further development in the field of AI. Both parties therefore aim to publish the project results online, so that they can be used by researchers.'

How does it work?

'We use the Hasty software, although in the first instance it can’t really do anything. To train the software, students manually mark all the objects to be recognised on a number of maps. For each map, we teach Hasty that a certain image is a mill, for example. If we mark these images often enough, the software learns: “When I come across this kind of image, it really is a mill.” The more images you add, the more accurate it becomes. After the training, Hasty can automatically recognise objects on other, similar maps. We can also export what Hasty learns, which means that others can use the AI model. Ultimately, we will make the data available in such a way that researchers can automatically establish links with other data.'

Why did you choose the Atlas of the Netherlands as a pilot?

'First of all, this atlas is a highlight of our collection. In addition, the content involves a varied collection of maps that lends itself well to machine learning. This atlas contains early modern historical maps, and these kinds of maps form the core of the Allard Pierson map collection.

The display of objects on maps from this period differs only slightly from one cartographer to another. By using these maps to train Hasty, we are developing a model that allows us to process all our early modern maps in an automated way. Other institutions will also be able to enrich similar map collections in this way if they use the freely accessible AI model.'

What was achieved in 2021?

'Last year, we carried out a pilot. During this pilot, we trained Hasty using an initial set of maps with input from students, the professor and the project leader. The researchers rated the results very positively and are happy to contribute to a follow-up project. The model needs to be further trained with the help of volunteers.

As we were able to determine the keywords, we know which objects to mark with the corresponding terms. We also know what we shouldn’t mark: for example, a line on a map could be a road or a national border, but only a human can see what it means. As the software can’t make this kind of interpretation, it make no sense to mark these types of objects.'

What are your ambitions for the future?

'We first want to process the areas for improvement flagged in the evaluation. In the follow-up project, together with volunteers, we will then mark the six hundred or so remaining maps. We will then publish all the raw data as publicly accessible, linked open data. We also want to export the AI model from Hasty and make it available as a model. Anyone will be able to copy, modify and distribute it free of charge.

Our ambition for the future is to create a freely available interface that allows users to search for objects on maps and enables a faithful representation.

Finally, we would like to digitise and enrich all the early modern maps of the Allard Pierson collection with AI-controlled image recognition.'

This interview is part of the Annual Report 2021 of the UvA/AUAS Library. Curious what else the Library has achieved in 2021? View the entire annual report.