Species Distribution Modeling with Spatial Point Process: Comparing Poisson and Zero Inflated Poisson-Based Algorithms

Abstract

Spatial point pattern is randomly arranged collection of points distributed over space, such as the locations of a tree species in a forest. Such a study is also commonly known as Species Distribution Modeling (SDM), where the main concern is to relate the distribution of tree species and environmental variables. Within spatial point process framework, SDM is closely related to modeling the intensity of spatial point process. The standard technique for parameter estimation of the intensity is by method of Maximum Likelihood Estimation (MLE) employing Berman-Turner Approximation, resulting in Poisson-based regression. However, this technique could raise an issue due to a large number of dummy points required in the approximation since large number of dummy points relates to excessive zeroes in response variable. Previous studies suggest the application of Zero Inflated Poisson (ZIP) regression over Poisson regression to model response variable with excessive zeroes. This study compares Poisson and ZIP-based method for modelling the distribution of Beilschmiedia Pendula tree with respect to environmental covariates. We compared both techniques by Bayesian Information Criteria (BIC) and concluded that the ZIP-based method performs better mainly due to excessive zeroes from dummy points. In addition, elevation and gradient affect significantly the distribution of Beilschmiedia Pendula tree.