American Marten (Martes americana) Habitat Suitability Change Models

Download: Vertebrate Habitat Suitability (829 MB)

We built habitat suitability models for 237 bird, 117 mammal, and 12 amphibian species. Species were chosen for inclusion in the study based on a simple set of criteria. For a species to be included in the study, it had to be primarily associated with terrestrial habitats, have a digital map of its current range, and have some portion of its current distribution intersect with the study area extent. In addition, we restricted the list of species used in the study to those for which a well-performing continental-scale model could be built. Digital species range maps were converted from polygons into 50 square kilometer resolution grid cells representing species presences. Although using point-based occurrence data to represent species presences is preferred when building correlative niche models, comprehensive occurrence data sets that adequately represent entire species ranges are generally unavailable, particularly for wide-ranging species with distributions extending into the subarctic and artic regions of North America. We used maps representing species ranges (Ridgley et al. 2003, Patterson et al. 2003, IUCN 2013), that at a coarse, continental scale were deemed to be adequate for representing species’ climatic niches.

Values represent projected habitat suitability changes. Each cell is attributed with one of four values, which represent:

• 10 - not suitable habitat
  
• 11 - not suitable habitate historically (1961-1990), suitable in the future (2070-2099), i.e. "expansion"
  
• 20 - suitable habitat historically, not suitable in the future (2070-2099), i.e. "contraction"
  
• 21 - suitable habitat historically, present in the suitable in the future (2070-2099), i.e. "stable"
  

The tens digit represents presence/absence historically with 1 = not present and 2 = present. The ones digit represents the projected future presence/absence (for the future time period) with 0 = not present and 1 = present.

Climate suitability models were developed for each of the 366 species using the 50-km resolution dataset, and we used random forests classifiers to model species distributions as a function of the bioclimatic variables. Random forest is an ensemble-based machine-learning algorithm used for both classification and regression analysis producing relatively accurate predictions based on the combined results of multiple classification trees. The random forest models produce predictions ranging from 0–1. To convert these values to a binary prediction representing suitable or unsuitable conditions, we selected a threshold value for each species model based on the receiver operator characteristic (ROC) curve, and an equal weighting of the importance of false positives and false negatives. To project potential future changes in climatic suitability for the 366 species, we used projected climate from two coupled atmospheric-oceanic general circulation models (AOGCMs), the Hadley CM3 model, and the Canadian Centre for Climate Modeling and Analysis CGCM3.1 model. The Hadley CM3 model simulates future climatic conditions that are warmer and drier relative to CGCM 3.1 model projections. We used projections for one greenhouse-gas emissions scenario—the A2 scenario, as described by the IPCC Special Report on Emissions Scenarios. The A2 scenario represents a mid-high emission scenario. Projected future values for the 23 bioclimatic variables were applied to the same 30 arc-second grid used for the historical data. The future projections were averaged across a 30-year period, spanning the years 2070 to 2099.

The models were built using 75% of the presences and absences for each species. We used the remaining 25% to test the models. We calculated the proportion of correctly predicted presences and absences for each species. The species for which the models correctly predicted at least 80% of the presences and at least 95% of the absences were used in the study. We then applied the 50-km resolution climate suitability models to the downscaled 1-km2 resolution climatic data to produce fine-scaled versions of the baseline and future projected climate suitability maps for each species To account for the impacts of vegetation on species distributions and the fact that our climate-based models, built with coarse resolution data, were unable to capture finer scale patterns, we refined our model projections with projected changes in habitat suitability based on climate-driven changes in biome distributions. To assess potential impacts of biome shifts on species distributions, we used projected biomes from Rehfeldt et al. (2012).

Terrestrial habitat associations described in the NatureServe Explorer online database records to develop the species-biome relationships using the biome classifications developed by Rehfeldt et al. (2012). With these relationships as a guide, we classified each biome type as either suitable or unsuitable for each species. We then generated maps of biome-suitability for each species based on these classifications and the projected future biome distributions. For each species, we combined the map of projected biome-suitability with the map of projected climate suitability to produce a projection of habitat suitability. As a final refinement to these projections, for all non-synanthropic species, we reclassified areas dominated by urban, suburban, exurban and agricultural land-uses as being unsuitable. We classified species as being synanthropic or non-synanthropic based on habitat associations recorded in NatureServe Explorer.