High resolution land use forecasts for Switzerland in the 21st century

We modelled the spatial suitability for each LULC class with a neural network (NN) using >200 predictors and accounting for climate and policy changes. We used data augmentation to increase performance for underrepresented classes, resulting in an overall quantity disagreement of 0.053 and allocation disagreement of 0.15, which indicate good model performance.

These class-specific spatial suitability maps outputted by the NN were then merged in a single LULC map per time-step using the CLUE-S algorithm, accounting for LULC demand for the future and a set of LULC transition rules. As the first LULC forecast for Switzerland at a thematic resolution comparable to available LULC maps for the past, this product lends itself to applications in land-use planning, resource management, ecological and hydraulic modelling, habitat restoration and conservation.