A comparative study applying GARP and their parallel versions for ecological niche modelling
Fabiana Soares Santana, Tereza Cristina Giannini, Antonio Mauro Saraiva, Isabel Alves dos Santos
Last modified: 2008-09-13
In order to address the preservation of natural resources, an approach is to study the geographic distribution of species from spatial, ecological and evolutionary perspectives. Ecological niche modelling combines species ocurrence data with environmental layers to generate models that represent the probabilistic distribution of species using specific algorithms. These models have been used, for example, to propose scenarios for sustainable use of the environment and to evaluate climatic changes impacts on biodiversity. Models can be generated by applying several algorithms, but GARP (or GARP with Best Subsets implementation, named GARP BestSubsets) is probably the most applied algorithm for ecological niche modelling worldwide. GARP has an inherent sequential characteristic. However, parallel solutions for GARP were proposed for cluster implementations based on MPI, named P-GARP, P-BestSubsets and HighP-BestSubsets, as part of the openModeller project. P-GARP is a parallel version of GARP, breaking the sequential characteristic but maintaining the genetic algorithm approach. P-BestSubsets is a parallel version of GARP BestSubsets. HighP-BestSubsets is a parallel version of GARP BestSubsets using P-GARP. This paper compares the results obtained with parallel and original GARP and GARP BestSubsets, from a correctness and performance standpoint. Species taxonomy, ecosystem, #points, #layers and granularity are furnished to improve the quality of the study. Species of Peponapis bees and Cucurbita (squashes and pumpkins) were considered in the study. They are highly associated to each other due to the pollination dependency of Cucurbita flowers on the visit of these bees. Performance results indicate a speedup of about three for P-GARP related to GARP, which means that P-GARP is three times faster, for a defined range of processors. Other parallel algorithms speedups are even better. P-BestSubsets is as accurated as GARP BestSubsets, since parallel version implies in minimum differences from original implementation. The correctness of P-GARP and HighP-BestSubsets is relatively dependent on the number of iterations: as this number increases, the algorithms correctness also increases and differences tend to be irrelevant. The same results are observed when convergence is possible. Therefore, for a number of iterations high enough or for converging models, experiments allowed us to conclude that P-GARP and HighP-BestSubsets algorithms are also as correct as the original ones, for the studied species. This paper illustrates that parallel versions of GARP may be, in fact, faster than and as correct as original algorithms. Besides, positive results may represent a relevant step in order to encourage parallel algorithms design for ecological niche modelling.