Bioclimatic similarity between species locations and their environment revealed by dimensionality reduction analysis

Abstract

Species distribution modeling is an active research topic with applications in conservation management, pest risk assessment, and population ecology. Several machine-learning methods have been applied to estimate species distribution. Non-linear dimensionality reduction techniques aim to preserve the similarity among objects at a reduced dimension for visualization, clustering, and feature selection. We propose a framework that uses Uniform Manifold Approximation and Projection (UMAP) to analyze bioclimatic variables associated with environmental (background) and species samples. Our objective was to identify geographic areas similar to those inhabited by the species. We hypothesize that the similarity between species locations and their environment in the reduced dimension will reflect similarity in the multivariate bioclimatic space. We estimated the probability of background points near a species point utilizing the latent nearest neighbor distance distribution. We tested this procedure with ten insect pest species of global importance and found that UMAP was able to generate a gradient of similarity between geographic areas and species occurrence. We also found that background-species latent distance tends to have a convergent non-linear relationship with the mean value of bioclimatic variables, thus supporting our key assumption. The performance of UMAP as a binary classifier and comparison with MaxEnt supports its use in modeling of species distribution. Potential applications are discussed for multi-species and multi-scenario analysis, as well as projection to new regions.