We employ a multi-stage modelling approach to assess the impact of climate change on tree species distribution in the eastern United States. In the first stage, we  adopt an empirical-statistical approach to determine the potential suitable habitats for 134 tree species under 3 GCM scenarios (HadCM3, GFDL and PCM) and 2 emission scenarios (A1FI and B1) at a 20 km resolution.  In particular we adopt a tri-model approach called DISTRIB where we use decision-tree ensembles to predict species abundance using a robust technique called RandomForest; then, we assess the reliability of the model with bagging trees, and in cases where we deem the model to be stable and therefore reliable, we use a single decision tree to understand and map the predictors driving the distribution of abundances spatially.

The individually predicted species abundances were assembled to forest types showing large increases in the oak-hickory type in the northeast for the high emission scenario.

In the second stage, we employ a spatially explicit cellular model called SHIFT, which calculates the colonization probabilities in each 1km cell using the potential suitable habitats from DISTRIB and the percent forest in each cell. The long distance dispersal mechanism is captured with an inverse power function of the distance between occupied and unoccupied cells.  Essentially, the SHIFT model restrains the colonization of species suitable habitats spatially - because not all suitable habitats can be expected to be colonized under current fragmented landscapes. 

The SHIFT outputs for several species revealed that only a small portion (<~15%) of the newly suitable habitat predicted by DISTRIB would have even a small probability of getting colonized within 100 years, meaning large lag times will inhibit non-assisted migration.

The DISTRIB and SHIFT models do not address several biological and disturbance factors that could ultimately determine the species ability to respond to climate change.  Therefore in the third stage, in order to maximize the usefulness of our models, we devised a scoring system called MODFAC that weights species-specific biological and disturbance factors (gleaned from various sources) to come up with a better potential future tree habitat model at regional and local scales.

Results/Conclusions

We believe this multi-stage modelling approach is an effective strategy to get a better handle on the numerous uncertainties that beset empirical-statistical species distribution models under future climates.  It should provide managers grappling with these uncertainties a better assessment tool for implementing management actions on the ground.