Fig 1.
Comparison of the mean accuracy (AUC) of SDM models over 1000 simulated taxa.
Sample clumping is caused by either biological or random processes. Panels show the predictive accuracy of data subsets binned into either high or low clumping and high or low coverage of the simulated true range. Points represent mean AUC scores from 1000 validation points per taxa and whiskers 95% confidence intervals around each mean, where scores less than 0.5 represent no accuracy gain over random chance. Spatial INLA—Bayesian SDM inferred using Integrated Nested Laplace Approximation with a spatial autocorrelation term, Non-spatial INLA—Bayesian SDM inferred using INLA without a spatial autocorrelation term, BRT—boosted regression trees based SDM, MAXENT—maximum entropy based SDM.
Fig 2.
Comparison of the mean accuracy (AUC) of SDM models over 1000 simulated taxa when altering the pseudo-absence (background) point configurations and the effects of spatial thinning of presence points, on four SDM methods and across 4 types of dataset with different clumping and spatial bias.
Panels show the predictive accuracy of data subsets binned into either high or low clumping and high or low coverage of the simulated true range. Where R represents random absence points, ST—spatial thinning, SW—spatially weighted absence points, B—both weighting and thinning (S1 Table) and Spatial INLA—Bayesian INLA model with spatial random effect, Non-spatial INLA—Bayesian INLA model without spatial autocorrelation term, BRT—boosted regression trees, and MAXENT—Maximum entropy based model. Points represent mean AUC scores from 1000 validation points per taxa and whiskers 95% confidence intervals around each mean, where scores less than 0.5 represent no accuracy gain over random chance.
Fig 3.
Comparison of mean accuracy (AUC) of spatially-explicit INLA SDM models on 1000 simulated taxa when varying the complexity of the underlying spatial mesh.
Colours show the predictive accuracy of data subsets binned into either high or low clumping and high or low coverage of the simulated true range. Points represent mean AUC scores across 1000 taxa and whiskers 95% confidence intervals around each mean, where scores less than 0.5 represent no accuracy gain over random chance.