Target-group backgrounds prove effective at correcting sampling bias in Maxent models

Aim Accounting for sampling bias is the greatest challenge facing presence-only and presence-background species distribution models; no matter what type of model is chosen, using biased data will mask the true relationship between occurrences and environmental predictors. To address this issue, we review four established bias correction techniques, using empirical occurrences with known sampling effort, and virtual species with known distributions. Innovation Occurrence data come from a national recording scheme of hoverflies (Syrphidae) in Great Britain, spanning 1983-2002. Target-group backgrounds, distance-restricted backgrounds, travel time to cities and human population density were used to account for sampling bias in 58 species of hoverfly. Distributions generated by bias correction techniques were compared in geographical space to the distribution produced accounting for known sampling effort, using Schoener's distance, centroid shifts and range size changes. To validate our results, we performed the same comparisons using 50 randomly generated virtual species. We used sampling effort from the hoverfly recording scheme to structure our biased sampling regime, emulating complex real-life sampling bias. Main conclusions Models made without any correction typically produced distributions that mapped sampling effort rather than the underlying habitat suitability. Target-group backgrounds performed the best at emulating sampling effort and unbiased virtual occurrences, but also showed signs of overcompensation in places. Other methods performed better than no-correction, but often differences were difficult to visually detect. In line with previous studies, when sampling effort is unknown, target-group backgrounds provide a useful tool for reducing the effect of sampling bias. Models should be visually inspected for biological realism to identify any areas of potential overcompensation. Given the disparity between corrected and un-corrected models, sampling bias constitutes a major source of error in species distribution modelling, and more research is needed to confidently address the issue.