2024-10-04, 14:45–15:30, Wodak Room (IIASA)
In this workshop, we will leverage hierarchical land cover legends and pixel-based uncertainty to create wall-to-wall land cover maps with a guaranteed level of accuracy.
Participants will join Martijn Witjes to balance detail and accuracy in land cover and crop mapping, leveraging pixel-based uncertainty and hierarchical legends.
We will work with previously predicted probabilities for 52 LUCAS land cover classes, and through binned multi-level validation, aggregate likely errors to more general classes. For example, a pixel where the model had difficulty discerning between wheat or barley will be aggregated to 'Cereal', which is more likely to be correct, although less exact. This sacrifice of detail will target the worst predictions first, until a desired level of accuracy is attained.
Participants proficient in Python are encouraged to bring their laptops: a Google Colab notebook will be shared for them to explore the trade-off between detail and accuracy in a (European) area of their interest.
Open-Earth-Monitor Cyberinfrastructure (Grant agreement ID: 101059548)
Martijn is a junior researcher at OpenGeoHub. He likes to play with probabilities to make maps more useful.