Integrating occurrence data and expert maps for improved species range predictions

• Published in: Global ecology and biogeography Vol. 26; no. 1/2; pp. 243 - 258
• Main Authors: Merow, Cory, Wilson, Adam M., Jetz, Walter
• Format: Journal Article
• Language: English
• Published: Oxford John Wiley & Sons Ltd 01.02.2017; Wiley Subscription Services, Inc
• Subjects: biogeography; case studies; computer software; decision making; Ecological niche model; expert opinion; habitats; linear models; MACROECOLOGICAL METHODS; maximum entropy; Poisson point process; prediction; species distribution model; vegetation types
• DOI: 10.1111/geb.12539
• ISSN: 1466-822X

Aim: Knowledge of species geographical distributions is critical for many ecological and evolutionary questions and underpins effective conservation decision-making, yet it is usually limited in spatial resolution or reliability. Over large spatial extents, range predictions are typically derived from expert knowledge or, increasingly, species distribution models based on individual occurrence records. Expert maps are useful at coarse resolution, where they are suitable for delineating unoccupied regions. In contrast, point records typically provide finerscale occurrence information that can be characterized for its environmental association, but usually suffers from observer biases and does not representatively or fully address the geographical or environmental range occupied by a species. Innovation: We develop a new modelling methodology to combine the complementary informative attributes of both data types to improve fine-scale, large-extent predictions. We use expert delineations to constrain predictions of a species distribution model parameterized with incidental point occurrence records. We introduce a maximum entropy approach for combining the two data types and generalize it to Poisson point process models. We illustrate critical decision making during model construction using two detailed case studies and illustrate features more generally with applications to species with vastly different range and data attributes. Our methods are illustrated in the Supporting Information and with a new R package, bossMaps, that integrates with existing generalized linear modelling and Maxent software. Main conclusions: Our modelling strategy flexibly accommodates expert maps with different levels of bias and precision. The approach can also be useful with other coarse sources of spatially explicit information, including habitat associations, elevational bands or vegetation types. The flexible nature of this methodological innovation can support improved characterization of species distributions for a variety of applications and is being implemented as a standard element underpinning integrative species distribution predictions in the Map of Life (https://mol.org/).