hide
Free keywords:
Land cover, Area statistics, Post-processing, Uncertainty
Abstract:
Providing land cover estimates with both correct pixel-level class predictions and regional class area estimates is important for many monitoring and accounting purposes but rarely achieved by current land monitoring efforts. We propose a framework that uses class probabilities predicted by machine learning to guarantee that the mapped proportion of each class matches independent area estimates. We used CatBoost models trained on CORINE data to predict probabilities for 8 primary LUCAS land cover classes in five European countries. We then used the proposed algorithm to produce proportional class maps that match Eurostat class area estimates. We validate these proportional class maps and baseline highest likelihood class maps with LUCAS land cover observations and S2GLC validation points. Our results show that the framework and algorithms create maps that match area estimates, and that may also be more accurate than maps created with highest likelihood classification. This is especially the case with general-purpose models trained on data whose class proportions are not representative of the mapped area, which means that this algorithm can be used to localize such models for more accurate mapping of individual countries.
