Using Maximum Entropy modeling to predict the potential distributions of large trees for conservation planning

Abstract

Large trees, as keystone structures, are functionally important in savanna ecosystems, and low recruitment and slow growth makes their conservation important. Understanding factors influencing their distribution is essential for mitigation of excessive mortality, for example from management fires or large herbivores. We recorded the locations of large trees in Hluhluwe-Imfolozi Park (HiP) using GPS to record trees along 43 km of 10 m-wide transects. Maximum entropy modeling (MaxEnt) uses niche modeling to predict the distribution of a species from the probability of finding it within raster squares, based on environmental variables and recorded locations. MaxEnt is typically applied at a regional spatial scale, and here we assessed its usefulness when predicting the distribution of species at a small (local) scale. HiP has variable topography, heterogeneous soils, and a strong rainfall gradient, resulting in a wide variety of habitat types. We used locations of 179 Acacia nigrescens and 106 Sclerocarya birrea (large trees ≥ 5m), and raster environmental layers for: aspect, elevation, geology, annual rainfall, slope, soil and vegetation. A. nigrescens was largely restricted to the Imfolozi section, while S. birrea had a wider distribution across the reserve. Understanding the interaction of environmental variables dictating tree distribution may facilitate habitat restoration, and will assist planning decisions for persistence of large trees within reserves, including options to reduce fire frequency or herbivore impacts. Though the AUC (Area Under the Curve) values used to test model predictions were high for both species, the ground truthing test data showed that distribution for A. nigrescens was more accurate than that for S. birrea, highlighting the need for independent test data to assess model accuracy. We emphasize that MaxEnt can be used at finer spatial scales than those typically used for species occurrence, but models must be tested using spatially independent test data.