Potential distribution models of Sechium tacaco (Pittier) C. Jeffrey in Costa Rica

Abstract

Objective: To determine the current and potential distribution of S. tacaco in Costa Rica with seven machine learning models, to optimize the management of phytogenic resources of S. tacaco, aimed at identifying patterns of geographic distribution and possible climatic adaptations that allow to have perspectives on their conservation and breeding.

Design/Methodology/Approach: 21 occurrence records together with 19 bioclimatic variables and the altitude variable were used to evaluate seven machine learning models and an assembly of them. Open-source libraries executed in Rstudio were used.

Results: The distribution models were inferred by the variables bio1, bio2, bio3, bio4, bio12, bio13, bio14, bio18 and bio19. The GAM model obtained the highest AUC (0.96) and TSS (0.9) values, however, the seven evaluated models and the assembly showed adequate performance (AUC> 0.5 and TSS> 0.4). The bioclimatic variables related to temperature turned out to be the ones with the greatest contribution to the models and the main limitations in the distribution of S. tacaco.

Limitations/Implications: A greater number of occurrence records may be required to evaluate distribution models.

Conclusions: The areas with high suitability for the potential distribution of S. tacaco are found in the central valleys of Costa Rica, covering regions of the provinces of Alajuela, Cartago, San José and Puntarenas. These can be sources of germplasm for future conservation and breeding studies.

Keywords: Machine learning, germplasm, conservation, breeding.