Modelagem preditiva da distribuição potencial de espécies arbóreas na bacia hidrográfica do Rio Grande, MG

• Ano de defesa: 2015
• Autor(a) principal: Carvalho, Mônica Canaan
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: UNIVERSIDADE FEDERAL DE LAVRAS; DCF - Departamento de Ciências Florestais; UFLA; BRASIL
• Palavras-chave em Português: Fitogeografia; Random forest; Árvore de decisão; Redes neurais artificiais; Maxent; Phytogeography; Decision tree; Artificial neural networks; CNPQ_NÃO_INFORMADO
• Link de acesso: http://repositorio.ufla.br/jspui/handle/1/9580

The present study had the main objective of comparing the performance of four machine learning algorithms (Decision Tree, Random Forest, Artificial Neural Networks and Maxent) in modeling the distribution of tree species in the state of Minas Gerais, Brazil. To this end, we used data from 197 inventoried fragments in Minas Gerais and 25 variables related to climate, topography and soil. The predictive capacity of the algorithms was evaluated by measuring the Area Under the Curve (AUC) obtained by cross-validation (10%) and by a set of independent test data (30%). The results obtained by the cross-validation were tested by the T-matched statistical test, with 95% confidence. We evaluated two sets of abiotic attributes in the modeling, the first was formed by all 25 abiotic variables available and the second by 10 variables with the highest information gain. The species Casearia sylvestris, Copaifera langsdorffii, Croton floribundis and Tapirira guianensis were selected according to their high abundance and wide distribution in the state. For all these species, the algorithms showed no significant improvement in performance when modeled. According to the cross validation, most species showed no significant difference between the predictive capacity of the Decision tree, Random Forest and Artificial Neural Networks. However, Random Forest demonstrated numerically superior AUC in most cases. The Random Forest was the superior of all tested algorithms, including the Maxent, when the validation set was run. The area predicted by the Random Forest was smaller than that predicted by Maxent when using the minimum limit of environmental suitability present in the training set; and smaller when the environmental suitability is reclassified, adopting the limit of 0.5. According to the evaluation metrics and maps obtained for each species, the Random Forest algorithm showed great potential for modeling species distributions.