Modeling Southeast Asian bat distributions: Assessing the effect of ecology and spatial biases on model accuracy



Journal Title

Journal ISSN

Volume Title



Species distribution models (SDMs) describe the probability of a species being present, given locality records and environmental characteristics of the study area. SDMs guide conservation decisions about reserve selection, species translocation, habitat priorities, and predicted areas susceptible to biological invasions However, the ability of any SDM to aid conservation is affected by data biases and model accuracy. Failure to minimize data biases and maximize model accuracy can lead to a misrepresentation of species distributions, compromising the conservation decision process. Spatial biases in data cause models to assign a high occurrence probability to an area primarily because of sampling methodology and effort rather than species’ ecology. To combat spatial sampling biases, survey effort should be systematic; however, surveying for a species across its entire range can be logistically difficult due to the inability to access sites across a species’ range, lack of funding or trained personnel, and differences in survey methodology. Consequently, the most common spatial bias in occurrence data is an over-sampling of easy-access sites such as protected areas, urban areas, and sites near roads, and at the same time under-sampling in hard-to-access sites. Researchers have advocated for the use of virtual species frameworks for assessing modeling methods because the mechanisms governing species distributions can be manipulated, modeled distributions can be compared to a simulated known distribution (referred to as the “true” distribution), and data quality and availability issues are avoided. Bats are a good model for exploring distribution model accuracy and biases but they are understudied in species distribution modeling research. A diverse taxon, with over 1300 species world-wide, bats fit into a variety ecological trait ensembles, which means that accuracy and bias hypotheses can be tested on the same taxonomic order but with species that have different niche requirements. Many bat species also have widespread ranges that are difficult to evenly sample, meaning that biases will be present in locality data, and given bat diversity, the sampling biases may be caused by different mechanisms for different species or ensemble groups. A lack of regional data on bat species impedes conservation efforts of most bats in Southeast Asia. Maximizing accuracy in regional bat distribution models would aid conservation decision-making such as where to place protected areas and identification of research gaps. Therefore, the aim of my dissertation was to assess spatial data biases in Southeast Asian bat locality and examine the effects of ecology and spatial bias correction on distribution models.

My results demonstrated that roosting and foraging ecology traits influenced spatial bias, but distance to protected area was the most significant contributor to spatial bias. Bats were mostly heavily sampled in or near a protected area, and over time, locality records are getting closer to protected areas. Therefore, despite increased sampling effort over time, bat localities are severely biased towards protected areas, which may not characterize the entirety of a species’ niche or equate to equal sampling across bat species and families. Furthermore, in contrast to the current literature, virtual generalist species had higher distribution model accuracy than the specialist species. And while spatially biased models had lower accuracy than unbiased models, bias correction methods did not significantly improve accuracy. Finally, I matched real Southeast Asian bat species with my virtual species to assess if the patterns of ecology, bias correction, and model accuracy were similar. Real species distribution models showed no significant difference in accuracy scores between specialists and widespread generalists, and both groups had higher accuracy than the restricted generalists. Accuracy scores showed no correlation between real and virtual species, though there were differences when data were split into ecological niche groups. Different accuracy metrics (TSS and AUC) had a weak positive correlation, and there were no significant correlations for individual ecological niche groups. The use of TSS versus AUC scores changed the accuracy order of the ecological niche groups, which changes the interpretation of how ecology affects model accuracy. These results highlight the need for continued research on if model inconsistencies are caused by true ecological effects, statistical artifacts, or difference among modeling algorithms.



Chiroptera, Species Distribution Model, MaxEnt, SDM