When events, such as presence of individuals or particular soil properties (e.g. contaminants), are highly clustered in space or are truly rare and infrequently observed, large spatial areas are devoid of events. As a result, sampled data often exhibit large numbers of zero values, a condition characterized statistically as zero-inflated frequencies. Because of the excess of zeroes, modeling of spatial distributions and estimation of population parameters can be problematic. Associated with the imperfect modeling of the spatial distribution of the rare events is that models of the impact of anthropogenic effects or other environmental changes on those events are also imperfectly captured. In this talk we compare prediction performances of several alternative modeling approaches for zero-inflated data including: 1) a spatially-explicit, zero-inflated model that uses soil characteristics as explanatory variables for tree species distributions; 2) two-stage fitting procedures in which a binary model is first fitted to determine probability of presence based on the spatial soil parameters and then, conditional on the first outcome, a model for the non-zero data is fitted; and 3) a Bayesian hierarchical approach that incorporates spatial autocorrelation in the model.  These models are applied to the spatial distribution of particular tree species in North Central US as a function of soil characteristics available digitally from the USGS.