Open-source locality data are a key resource for species distribution modeling. Unfortunately, data are commonly spatially biased by variable collection methods and towards landscape features that promote researcher access such as protected areas and roads. Differential contribution of such factors has not been assessed nor if contributions vary across ecological trait groups, which might be expected as species traits can cause spatial variation in capture frequency. We assessed which variables contributed most to sampling bias and if contribution varied across ecological trait groups. MaxEnt software was used to model variable contribution and variable response curves for the effect of distance to protected areas, distance to roads, distance to universities, population density, land cover, and economic status on the presence of Southeast Asian bat sampling localities downloaded from GBIF (5840 records, 1980-2010). MaxEnt with 10-fold cross-validation was used for an overall model for all sampling localities and nine models for different ecological groups. Southeast Asian bats are an ideal model taxon as they are speciose with high variation in ecological traits that cause differential capture success. We used foraging and roosting ecology to group sampling localities as these traits can greatly affect spatial and taxon variability in captures.
Results/Conclusions
Across all models distance to protected areas and population density contributed most (>50%) to sampling locality presence, thus these features contribute most to spatial bias. Factor contribution varied across foraging groups. Insectivorous bats that forage within forested habitats were most biased towards protected areas (54%), whereas insectivorous species that forage in open areas were most biased by population density (54.4%). Model contributions for plant-visiting bats were more balanced across distance to protected areas (30.8%), population density (29%), and distance to universities (27.4%). Interestingly, distance to roads did not contribute highly to any model (<5.9%) although it is a common focus in the spatial bias literature. Across all models, variable response curves showed similar trends; sampling localities are more biased towards protected areas, roads, universities, and areas with a higher population density and higher economic status. Overall, bat sampling localities are most biased by distance to protected areas and population density, however percent contribution of these features vary across ecological trait groups. Currently spatial data biases are corrected without regard to taxa or origins of bias. To build more realistic species distribution models for ecological hypotheses and conservation decisions specific trait and bias origin information should be incorporated into bias correction.