There’s no best model! Addressing limitations of land-use scenario modelling through multi-model ensembles

Cellular automata models are popular tools for exploring future land change pathways. But simulation modelling approaches often focus too narrowly on calibration against historic reference maps, limiting the diversity of possible outcomes. We argue that, contrary to what is commonly believed, there is no ‘best model’, and that model specification and calibration accuracy depend on the objective of the research. We propose a multi-model ensemble approach, in which a wide range of models and calibration rules sets are systematically tested against multiple metrics. We apply our approach to a case study in Spain. No single model performed well for all statistics, illustrating the danger of cherry-picking statistics for best performance. In our case study, accounting for historic land changes in model design was useful for simulating compact urban development, but limited the variability of simulation outcomes. The accessibility model driver improved urban pattern replication, while suitability without accessibility was useful for simulating low-density development encroaching on natural areas. Rather than abandoning calibrations that show low agreement with reference maps based on a small number of metrics we should seek to understand what each metric is telling us and use this information to enrich the diversity of simulated outcomes.