Big and Useful Maps: Land use / land cover classification with high thematic depth

Recent rapid increases in the quality, volume and accessibility of Earth Observation data has spawned a plethora of ambitious global and continental land use / land cover mapping initiatives like ESA’s WorldCover and Google’s DynamicWorld. While such projects are revolutionary in both their scale and their spatial and temporal resolution, they often limit their mapping efforts to a handful of abstract land cover classes like forests, urban areas, and cropland. While this makes it much easier to make accurate global maps and large-scale analyses, it begs the question on how useful these maps are to decision makers who have specific questions and may need more detailed categories, such as distinguishing between different vulnerable nature types and tree plantations or intensively grazed pastures. 

This creates several challenges, however: Firstly, to what extent are commonly used accuracy metrics suitable for classification tasks with several, often highly imbalanced classes, some of which might occur within the same pixel? Secondly, how do you serve these much more complex datasets to users in a way that actually helps them answer their questions? This project investigates several techniques that leverage the use of sets of predicted probabilities to 1) assess and compare model performance, 2) provide reliable map uncertainty, 3) analyse long-term trends.