Spatial-statistical downscaling with uncertainty quantification in biodiversity modelling

Accurate downscaling with uncertainty quantification and its inclusion in fitting biodiversity models to data are essential for accurate, valid inferences and predictions. Here, we provide a general framework for spatial modelling of biodiversity that involves downscaling environmental covariates. We derive downscaling for ecological data based on a spatial-statistical model that accounts for spatial change-of-support. Through a simulation study, we demonstrate that our statistical downscaling provides accurate uncertainty quantification. With the Monte Carlo samples of a downscaled covariate, we develop a two-stage protocol that propagates downscaling uncertainty to a generalised linear model (GLM), commonly used in biodiversity modelling. We call the implementation of the protocol CORGI (Change Of Resolution in GLM Inference). A simulation study shows that this framework for downscaling covariates improves the quantification and propagation of uncertainty for use in biodiversity modelling when compared to existing methods. The two-stage protocol is of broad utility given the routine use of environmental covariates available at spatial scales different from those of species population or diversity metrics in biodiversity models. Moreover, the protocol is readily implemented with the aid of standard software packages. Extensions of the protocol that include accounting for measurement errors and missing values in the covariate data, non-Gaussian covariate data, fusing multi-source data, adding spatial random effects and imposing physical constraints, are discussed.