Blog post

“We want to make land cover data that is always up to date”

Nov 12, 2021

by Helena Åström (helena@scalgo.com)


Last year we released an imperviousness map for Denmark in high resolution and with high accuracy. The map was created using machine learning which allows for nation-wide cost-effective mapping.

The map was well received and has been used extensively in city planning and surface water management across Denmark.

Figure 1. Last year we released an imperviousness map for Denmark in high resolution and with high accuracy.

Since then, we have been working on improving the method further.

"In the first map we used orthophotos from 2015 and 2016 because those photos were most similar to the available Lidar point cloud data in our training areas" say Jonas, machine learning expert at SCALGO.

The fusion of different input features, orthophotos from 2015-2016 and Lidar data, enabled ground-breaking performance in automated imperviousness mapping. But this fusion also introduced some problems.

“In Denmark, Lidar data is collected with a 5-year cycle. This means that some of the Lidar data doesn’t match the chosen orthophotos simply because urban areas develop fast”, Jonas explains.

“This confuses the model and reduces performance in some areas. Also, a 5-year update in data is not good enough if we want to create up-to-date land cover maps”.

Figure 2. Ørestaden, large development site in Copenhagen. In the map initial map (on the left), a mismatch between lidar data and photos resulted in a poor assessment. In the updated map (on the right) imperviousness is mapped with high accuracy.

Throughout the last year Jonas has worked on creating a new map, relying on only a single orthophoto source while improving performance compared to the previous model.

“Orthophotos are, in Denmark, updated every year. If we can become independent from lidar data we can make maps that are much more up to date”, says Jonas.

Figure 3. Urban areas develop fast. Lidar data collection cannot keep up with this development.

The new machine learning model is trained using orthophotos from varying years, which allows for training the model to understand different kinds of shadows, brightness in the pictures, colours, and other details in the photos. By doing this, Jonas has developed a model that is much more robust.

The new model has improved the performance in many locations, despite using less input data. An example is suburban areas where the new model maps driveways better than before.

Figure 4. Example to the left is from the old model while example to the right is from the new model. Despite using less input data, performance has improved in many areas of the new model.

An updated imperviousness map that uses orthophotos from 2020 has now been released and is available for all SCALGO Live users in Denmark. Our goal is to continue developing our model and release imperviousness maps to users in other countries too.

Want to learn more about how SCALGO Live uses machine learning to map land cover? Contact us (info@scalgo.com) to hear more.