Bayesian belief networks as a versatile method for assessing uncertainty in land-change modeling

Abstract

AbstractLand-use and land-cover change modeling helps us to understand the driving factors and impacts of human-induced land changes better, and depict likely future development paths. Uncertainty associated with various steps in the modeling process substantially influences the reliability of the results, but until now it has only rarely been addressed. In this study, we explore uncertainty in land-change modeling using a probabilistic approach based on Bayesian belief networks. We apply this approach to a case study of deforestation in the Brazilian Amazon and identify three modeling steps as sources of uncertainty: model structure, variable selection, and data preprocessing. For these three steps, we quantify the uncertainty and the respective impact on the outcome accuracy. The results indicate remarkable uncertainties in each of the steps. We demonstrate that a higher uncertainty in the land-change modeling process does not necessarily lead to a lower accuracy of the modeling outcome. Moreover, we show that the different uncertainty sources only slightly influence the ratio between quantity disagreement and allocation disagreement for the modeling outcome. We conclude that uncertainty is inherent in land-change modeling, and that future studies should address this uncertainty more explicitly to improve the robustness of modeling outcomes for science and decision-making.Keywords: uncertainty assessmentland-change processeserrorprobabilityspatiotemporal modeling AcknowledgmentsWe thank the participants of the land-change modeling workshop of CarBioCial (‘Carbon sequestration, biodiversity and social structures in Southern Amazonia’, see http://www.carbiocial.de) for their willingness to share their expert knowledge. The authors would also like to thank Dennis Funke for data preparation work, and particularly Florian Gollnow, Ulf Leser, and Daniel Müller for their helpful comments and fruitful discussions that helped to improve this manuscript. We additionally thank the anonymous reviewers which helped us improving an earlier version of the article by giving a variety of beneficial remarks and suggestions.Additional informationFundingThe presented work is funded by Deutsche Forschungsgemeinschaft (German Research Foundation) [GRK 1324/1, METRIK].