An AI model that keeps learning, creating ever-improving land cover maps

This enables the creation of precise high-resolution land cover maps from local datasets, with each new country adding challenges, training data, and opportunities for the AI model to learn.

As Morten Revsbæk, CEO of Scalgo, explains:

“We have deliberately positioned ourselves to map land cover continuously across the world. On one hand, we have developed a powerful AI approach that can identify land cover from aerial images with high accuracy. On the other hand, expanding into new countries gives us unique insights into local datasets and we can use these to train the AI model. The result is land cover maps that improve continuously.”

This development means that, when we return to countries to update land cover maps, the results are improved both in accuracy and with new land cover classes. Now it is Sweden’s and Great Britain’s turn to get enhanced land cover maps.

Better land cover classification in the mountains

When the inital Scalgo land cover map was launched in Sweden in 2022 it was the first detailed land cover map that covered the entire country. Classifying 2 trillion pixels was no small task, and the map has been a popular part of Scalgo Live with users working with the map more than 10.000 hours.

Since then we have been developing the underlying AI-model and with the newest release in Sweden, the classification performance in mountainous regions is significantly improved. Where early versions of the model had difficulties distinguishing between bare land, bare rock and snow, the newest map performs with high accuracy. 

This development also improves the land cover map in Great Britain: In Scotland, hillside and mountainous areas are now mapped more precisely with bare rock. Also, alpine grasslands and moorlands are distinguished as shallow vegetation, not bare land.

“Every time we encounter a new landscape or a new type of terrain, we can add training data and refine the model.”

Jonas Tranberg, Head of Machine Learning at Scalgo

From shallow vegetation to fields

Another important development has been the ability of the model to recognise agricultural fields directly from aerial imagery.

Until now, the Scalgo approach has relied on auxiliary datasets such as government-provided cropland boundaries. These worked well in countries like Denmark, France, and Germany, but in many countries - such as the UK, Poland, and the USA - farmland data is not always available nationwide in a form that is both consistent and precise.

To solve this, we trained the AI model to recognise farmland using training data from countries where reliable field boundaries exist.

The new Great Britain land cover map has been serving as a main test case, and the results are great: The new model now consistently achieves more than 90% accuracy in identifying fields across England, Wales and Scotland. The approved model is expected to scale well to other countries where auxiliary data is limited.

We have also continued to improve the model’s capability to distinguish between dense and shallow vegetation in, for example, forested areas.

More to come

The new model results represent a substantial step forward compared to the previous versions.

As Jonas Tranberg, Head of Machine Learning at Scalgo, puts it:

“Our AI method improves continuously. Every time we encounter a new landscape or a new type of terrain, we add training data and refine the model. These refinements then result in improvements to all maps when they are updated. So, the more countries we add, the better the model becomes in the countries we already map.”

The progress in Sweden and the UK illustrates how the Scalgo AI model develops: each new country provides new challenges, new training data, and new opportunities for the model to learn. Over time, this knowledge feeds back into all countries, improving accuracy everywhere.

Follow us for more news on AI and how we continue to improve land cover mapping across the world!