Incorporating abundance information and guiding variable selection for climate-based ensemble forecasting of species' distributional shifts

Abstract

Ecological niche models (ENMs) have increasingly been used to estimate the potential effects of climate change on species' distributions worldwide. Recently, predictions of species abundance have also been obtained with such models, though knowledge about the climatic variables affecting species abundance is often lacking. To address this, we used a well-studied guild (temperate North American quail) and the Maxent modeling algorithm to compare model performance of three variable selection approaches: correlation/variable contribution (CVC), biological (i.e., variables known to affect species abundance), and random. We then applied the best approach to forecast potential distributions, under future climatic conditions, and analyze future potential distributions in light of available abundance data and presence-only occurrence data. To estimate species' distributional shifts we generated ensemble forecasts using four global circulation models, four representative concentration pathways, and two time periods (2050 and 2070). Furthermore, we present distributional shifts where 75%, 90%, and 100% of our ensemble models agreed. The CVC variable selection approach outperformed our biological approach for four of the six species. Model projections indicated species-specific effects of climate change on future distributions of temperate North American quail. The Gambel's quail (Callipepla gambelii) was the only species predicted to gain area in climatic suitability across all three scenarios of ensemble model agreement. Conversely, the scaled quail (Callipepla squamata) was the only species predicted to lose area in climatic suitability across all three scenarios of ensemble model agreement. Our models projected future loss of areas for the northern bobwhite (Colinus virginianus) and scaled quail in portions of their distributions which are currently areas of high abundance. Climatic variables that influence local abundance may not always scale up to influence species' distributions. Special attention should be given to selecting variables for ENMs, and tests of model performance should be used to validate the choice of variables.

Fig 1. Model performance metrics used in determining the best variable selection approach to estimate potential distributions for temperate North American quail through the Maxent algorithm.

Variable selection approaches included a biologically relevant suite (black bar), the top performing random suite (grey bar), and a correlation/variable contribution (striped bar) suite. The random variable suite included in this figure is the random suite that had the best standardized omission error of the five random variable suites included in our analysis. Significant differences in model performance metrics are indicated by letter groupings from post hoc Tukey HSD pairwise comparison test results from a significant one-way ANOVA.

Fig 2. Test omission errors from regularization multiplier tuning experiments of Maxent models for temperate North American quail.

Test omission errors were based on a 10^th percentile presence threshold and are averaged across 100 replications. Significant differences in test omission estimates are indicated by letter groupings from post hoc Tukey HSD pairwise comparison test results from a significant one-way ANOVA.

Fig 3. Estimated percentages of distribution shifts for temperate North American quail based on ensemble projections of Maxent models into 2070.

Ensemble forecast model agreement is indicated as followed: 75% (black), 90% (gray), and 100% (striped).

Fig 4. Future predicted changes in distributions of California quail (Callipepla californica; a) and Gambel’s quail (Callipepla gambelii; b) projected to 2070 and based on ensemble forecasts (estimated through Maxent) at 90% agreement.

Major rivers of North America (blue lines) are included for geographic reference. Full descriptions for possible distribution conditions are given in Table 2. In short, distribution conditions represent: condition 1 (distribution expansion from current to 2050 and remaining suitable from 2050 to 2070), condition 2 (suitable at current and through all time periods), condition 3 (unsuitable from current to 2050 but expanding from 2050 to 2070), condition 4 (distribution contraction from current to 2050 but expanding from 2050 to 2070), condition 5 (distribution expansion from current to 2050 but contracting from 2050 to 2070), condition 6 (suitable from current to 2050 but contracting from 2050 to 2070), condition 7 (unsuitable at current and through all time periods), and condition 8 (distribution contraction from current to 2050 and remaining unsuitable from 2050 to 2070).

Fig 5. Future predicted changes in distributions of scaled quail (Callipepla squamata; a) and northern bobwhite (Colinus virginianus; b) projected to 2070 and based on ensemble forecasts (as estimated through Maxent) at 90% agreement.

Major rivers of North America (blue lines) are included for geographic reference. Full descriptions for possible distribution conditions are given in Table 2. In short, distribution conditions represent: condition 1 (distribution expansion from current to 2050 and remaining suitable from 2050 to 2070), condition 2 (suitable at current and through all time periods), condition 3 (unsuitable from current to 2050 but expanding from 2050 to 2070), condition 4 (distribution contraction from current to 2050 but expanding from 2050 to 2070), condition 5 (distribution expansion from current to 2050 but contracting from 2050 to 2070), condition 6 (suitable from current to 2050 but contracting from 2050 to 2070), condition 7 (unsuitable at current and through all time periods), and condition 8 (distribution contraction from current to 2050 and remaining unsuitable from 2050 to 2070).

Fig 6. Future predicted changes in distributions of Montezuma quail (Cyrtonyx montezumae; a) and mountain quail (Oreortyx pictus; b) projected to 2070 and based on ensemble forecasts (estimated through Maxent) at 90% agreement.

Major rivers of North America (blue lines) are included for geographic reference. Full descriptions for possible distribution conditions are given in Table 2. In short, distribution conditions represent: condition 1 (distribution expansion from current to 2050 and remaining suitable from 2050 to 2070), condition 2 (suitable at current and through all time periods), condition 3 (unsuitable from current to 2050 but expanding from 2050 to 2070), condition 4 (distribution contraction from current to 2050 but expanding from 2050 to 2070), condition 5 (distribution expansion from current to 2050 but contracting from 2050 to 2070), condition 6 (suitable from current to 2050 but contracting from 2050 to 2070), condition 7 (unsuitable at current and through all time periods), and condition 8 (distribution contraction from current to 2050 and remaining unsuitable from 2050 to 2070).