Modeling habitat suitability with statistical algorithms (e.g., maximum entropy, regression tree ensembles, linear regression, etc.) often relies on field inventories or records of occurrence to produce acceptable results. However, these models may be influenced by issues of zero-inflated data, sampling biases, or a mismatch between the response and dependent variables that could affect performance statistics. To reduce the influence of these issues, we used the spatial density of the response variable, relative abundance of individual tree species, to derive the underlying modeling grid, an irregular lattice of 10×10 and 20×20 km grids. We model habitat suitability for 125 eastern United States tree species using this hybrid approach by combining uniform grids and irregular networks and show how other modeling efforts could benefit from a density-driven grain size. Using 84,204 annual records from the USDA Forest Service Forest Inventory and Analysis dataset and 45 environmental variables (climate, elevation, soil), individual species relative abundance was modeled across a hybrid lattice of 29,357 cells using the randomForest algorithm.
Results/Conclusions
The hybrid lattice approach provides a way to refine spatial resolutions where information is abundant, by reducing the mean and range of inventory plots among grid cells. The coarser 20-km cells mainly occupy the western portion, which is dominated by agricultural lands transitioning to prairie, where trees are confined to woodlots and riparian corridors. Model reliability among the 125 species resulted in 30, 47, and 48 species classed as high, medium, and low, respectively, influenced by the extent and abundance of the current distribution. Comparison with regular grid of 10 km showed that the hybrid approach captured the variation in the response better. Along with updated environmental conditions our hybrid approach to assess potential changes in habitat suitability provides better results that aids management decisions.