Calibration of an Agent Based Land System Model using Machine Learning to support Future Land Use Projections

Land system models for future simulation offers computational support for analysing and decision making under severe uncertainty. It is an exploratory modelling approach as it uses computational experiments to understand an uncertain and complex subject. Among different advanced exploratory land system models, Agent Base Modelling (ABM) provides flexible descriptions of land use processes and useful supports for decision-making without making excessive assumptions about human and land system behaviours, with the trade-off that model results are often highly uncertain in the sense of being non-predictive. Still, evaluation and calibration of ABM is a critical task to be done properly, prior to applying it to real-world problems. It often requires to consider various social and natural processes and heterogeneous data sources. Mathematically, it is finding solutions in high-dimensional mixed-space with multi-criteria and often categorical parameters. Determining acceptable degree of uncertainty is another dimension of the problem in an effort to capture variability within land system processes. We provide an overview of the calibration of ABM in this study and propose a practical framework using various machine learning algorithms to parameterise an ABM (CRAFTY-EU) toward simulating future land. For fifteen model parameters regarding land managers behaviour and ecosystem functions, parameter values are determined in accordance with process description together with heterogeneous data sources, from land statistics to satellite data. The results show improved representation of land system dynamics in arable and forest, while less in land types with the little information supplied. We discuss another fundamental issue unresolved that a better fitting to observation may be at the expense of the flexibility, which may be critical in simulating unseen future behaviours. We hope the presented framework as well as critical discussions on the findings holds promise for improving the utility and transparency of land use models and identify critical areas for further work.