Evaluating Greedy and Shapley Approaches for Sequential Climate Envelope Modeling of Trees in the Conterminous United States

Understanding which species distribution modeling (SDM) methods and computational approaches are best suited to particular species, datasets, modeling horizons, and climate or land-use change scenarios is essential for developing robust models that may also yield mechanistic insights. In this study, we employ a data-intensive approach to identify the climatic variables associated with the spatial distributions of 25 tree species representing diverse taxonomic groups and ecological characteristics across the contiguous United States. We compare two approaches for ranking 19 climatic variables and sequentially constructing hyperrectangle-based climate envelope models using data from the USDA Forest Inventory and the WorldClim dataset. The first approach, based on Shapley values from cooperative game theory, ranks predictors according to their average marginal contributions across all possible predictor subsets, thereby accounting for interactions among variables. The second approach uses a greedy algorithm that iteratively selects variables according to their contribution to model performance, generating a ranking based on sequential importance. Across all species examined, we found that only five to seven climatic variables were sufficient to construct effective climate envelope models. Our results suggest that reducing the dimensionality of climate space is a nonlinear problem. Variable rankings differed markedly among species, likely reflecting species-specific differences in plant physiology and ecological traits that shape responses to climate. Overall, these findings demonstrate a fast and efficient framework for developing data-driven species distribution models and provide a foundation for future hypothesis-driven research aimed at uncovering the mechanistic basis of the observed variable rankings.