Global High-resolution Land-use Change Projections: A Bayesian Multinomial Logit Downscaling Approach Incorporating Model Uncertainty and Spatial Effects

Using econometric models to estimate land-use change has a long tradition in scientific literature. Recent contributions show the importance of including spatial information and of using a multinomial framework to take into account the interdependencies between the land-use classes. Few studies, however, agree on the relevant determinants of land-use change and there are no contributions so far comparing determinants on a global scale. Using multiple 5 arc minute resolution datasets of land-use change between 2000 and 2010 and taking into account the transitions between forest, cropland, grassland and all other land covers, we estimate a Bayesian multinomial logit model, using the efficient Polya-Gamma sampling procedure introduced by Polson et al. (2013). To identify and measure the determinants of land-use change and the strength of spatial separation, our model implements Bayesian model selection through stochastic search variable selection (SSVS) priors and spatial information via Gaussian Process (GP) priors. Our results indicate that spatial proximity is of central importance in land-use change, in all regions except the pacific islands. We also show that infrastructure policy, proxied by mean time to market, seems to have a significant impact on deforestation throughout most regions. In a second step we use aggregate, supra national land-use change results from the partial equilibrium agricultural model GLOBIOM as a framework for projecting our model in ten-year intervals up to 2100 on a spatially explicit scale along multiple shared socioeconomic pathways.