Correcting environmental sampling bias improves transferability of species distribution models

Abstract

Sampling bias is an inherent problem in widely available biodiversity data, undermining the robustness of correlative species distribution models (SDMs). To some extent, subsampling occurrence data can account for uneven sampling efforts; yet, conventional approaches subsample in geographical space, while subsampling in environmental space remains underexplored. Here, we compared the effectiveness of subsampling methods that correct sampling bias either in geographical space (spatial gridding, spatial distance thinning) or directly in environmental space (environmental gridding), including two novel approaches introduced here: environmental clustering and environmental distance thinning. We hypothesised that environmental subsampling methods would be more effective in improving SDM performance across its three primary uses: explaining, predicting, and projecting. Using a virtual ecologist framework, we assessed SDM performance against four evaluation tests: replicating true species–environment response curves, predicting within the sampling region via internal cross‐validation and evaluation against independent data, and projecting outside the sampling region. Our findings demonstrate that environmental subsampling methods, especially environmental clustering and environmental distance thinning, outperformed other methods in yielding robust SDMs in almost all evaluation tests. Interestingly, cross‐validation favoured SDMs with no sampling bias correction, highlighting the inability of cross‐validation to identify unbiased models. Our findings emphasise a critical conceptual disconnect: SDMs appearing to perform well in predicting species' distributions may not reliably estimate species–environment relationships, nor transfer predictions onto novel environments. Environmental subsampling methods are reliable approaches for all uses, but are particularly suited for explaining species' niches and transferring predictions across space and/or time, such as when anticipating species' responses to climate change or assessing the risk of biological invasions. Conversely, geographic subsampling methods may suffice for predicting species' distributions within their current environmental context, as required in conservation planning. Our study firmly establishes the critical importance of correcting environmental sampling bias, while also providing reliable solutions for supporting biodiversity conservation in an ever‐changing world.