Modeling Abiotic Niches of Crops and Wild Ancestors Using Deep Learning: A Generalized Approach

Abstract

Introduction Understanding what interactions and environmental factors shape the geographic distribution of species is one of the fundamental questions in ecology and evolution. Insofar as the focus is on agriculturally important species, insight into this is also of applied importance. Species Distribution Modeling (SDM) comprises a spectrum of approaches for establishing correlative models of species (co-)occurrences and geospatial patterns of abiotic environmental variables. Methods Here, we contribute to this field by presenting a generalized approach for SDM that utilizes deep learning, which offers some improvements over current methods, and by presenting a case study on the habitat suitability of staple crops and their wild ancestors. The approach we present is implemented in a reusable software toolkit, which we apply to an extensive data set of geo-referenced occurrence records for 52 species and 59 GIS layers. We compare the habitat suitability projections for selected, major crop species with the actual extent of their current cultivation. Results Our results show that the approach yields especially plausible projections for species with large numbers of occurrences (>500). For the analysis of such data sets, the toolkit provides a convenient interface for using deep neural networks in SDM, a relatively novel application of deep learning. The toolkit, the data, and the results are available as open source / open access packages. Conclusions Species Distribution Modeling with deep learning is a promising avenue for method development. The niche projections that can be produced are plausible, and the general approach provides great flexibility for incorporating additional data such as species interactions.