Maximum entropy modeling of mature hardwood forest distribution in four U.S. states

In the eastern United States, mature hardwood forest provides habitat for many species of native flora and fauna, but is much less common now than historically. This study examined the utility of maximum entropy modeling and spatial application to identify ecosystem types like mature hardwood forest. I performed pilot modeling in Charles County, Maryland, where I compared fine-scale geographic data available locally to coarse-scale data available nationally. As expected, a model constructed with the best locally available data, including LiDAR-derived canopy height and fine-scale soil maps, outperformed a model constructed with nationally consistent data. However, the model using national data nevertheless accurately identified most mature hardwood forest sites and excluded most young forest. I then applied the coarse-scale approach to four states: Pennsylvania, Ohio, Kentucky, and Tennessee. Average test AUC (area under the receiver operating curve) based on 10 replicates varied from 0.76 to 0.80 when comparing mature hardwood forest locations to general forest locations. The maximum training or test sensitivity plus specificity threshold, depending on the state, captured 78-79% of positive locations while rejecting 74-81% of negative locations. The maximum entropy approach is versatile, and can be applied to other ecosystems and species.