Modelling the joint species distribution: Simultaneous predictions of current distributions of three sympatric conifer species in the Central Rocky Mountains

Background/Question/Methods

The geographic distribution of a species is a function of abiotic conditions, biotic interactions, and dispersal ability. The relative influence of these factors varies across species, time and space, with biotic interactions emerging as particularly important in regions of sympatry among closely-related species. Species distribution models have been widely implemented to understand the factors determining species distributions, yet in most implementations these methods rely on correlation of species occurrence with environmental variables and fail to account for biotic interactions and dispersal.  In an effort to account for the effect of biotic interactions on species distributions, we developed multiple response models to simultaneously predict the current distributions of three sympatric conifer species in the Central Rocky Mountains: Pinus contorta, Pinus ponderosa and Pseudotsuga menziesii.  We used multivariate adaptive regression splines (MARS) and Conditional Forest classification tree ensembles with 30-year climate normals to predict probability of presence from occurrence data from the U.S. Forest Service F.I.A. database. We compared multiple response models with single response models to evaluate potential model improvement achieved by simultaneous prediction of multiple species distributions. Models were evaluated using standard indices of model performance and accuracy derived from confusion matrices, as well as overlap in predicted geographic distributions.

Results/Conclusions

Multiple response and single response MARS models both performed similarly when applied to validation data. All models showed strong predictive ability across all species with AUC values ranging from 0.85 to 0.90. Differences in predictions among single response and multiple response MARS models were most pronounced for P. ponderosa. Geographic similarity for single response and multiple response MARS models predicted to the study region was 88% for P. contorta, 83% for P. ponderosa, and 85% for P. menziesii.  Multiple response Conditional Forest models also displayed strong performance, with AUC values ranging from 0.88 to 0.94. Performance was similar for single and multiple response models, though accuracy was slightly improved by simultaneous prediction of all species. Our results indicate that directly accounting for species covariance within the Conditional Forest algorithm, rather than averaging responses across species as in the MARS algorithm, may more appropriately account for dependencies in species distributions. Additionally, our results offer no indication of competitive exclusion among the focal species, suggesting that, for some sympatric species, interactions may not always be an important consideration when predicting species occurrence.  Finally, preliminary results indicate that a Bayesian multiple-response approach may more appropriately characterize the joint species distribution.