Improving species distribution model quality with a parallel linear genetic programming-fuzzy algorithm.

• Ano de defesa: 2016
• Autor(a) principal: Bieleveld, Michel Jan Marinus
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Algoritmos genéticos; Applied and specific algorithms; Bioclimatologia; Ecological niche models; Fuzzy logic; Genetic algorithms; Inteligência artificial; Lógica Fuzzy; Species distribution modelling
• Link de acesso: http://www.teses.usp.br/teses/disponiveis/3/3141/tde-26012017-113329/

Biodiversity, the variety of life on the planet, is declining due to climate change, population and species interactions and as the result f demographic and landscape dynamics. Integrated model-based assessments play a key role in understanding and exploring these complex dynamics and have proven use in conservation planning. Model-based assessments using Species Distribution Models constitute an efficient means of translating limited point data to distribution probability maps for current and future scenarios in support of conservation decision making. The aims of this doctoral study were to investigate; (1) the use of a hybrid genetic programming to build high quality models that handle noisy real-world presence and absence data, (2) the extension of this solution to exploit the parallelism inherent to genetic programming for fast scenario based decision making tasks, and (3) a conceptual framework to share models in the hope of enabling research synthesis. Subsequent to this, the quality of the method, evaluated with the true skill statistic, was examined with two case studies. The first with a dataset obtained by defining a virtual species, and the second with data extracted from the North American Breeding Bird Survey relating to mourning dove (Zenaida macroura). In these studies, the produced models effectively predicted the species distribution up to 30% of error rate both presence and absence samples. The parallel implementation based on a twenty-node c3.xlarge Amazon EC2 StarCluster showed a linear speedup due to the multiple-deme coarse-grained design. The hybrid fuzzy genetic programming algorithm generated under certain consitions during the case studies significantly better transferable models.