Application of Fisher’s Discriminant Analysis to Classify Forest Communities in the Pampa Biome

Abstract

Fisher Discriminant Analysis (DA) seeks a linear combination of independent variables maximizing separation of predicted groups and also permits new observations for being classified in groups know a priori. We applied DA with eight structural attributes of vegetation obtained of systematic tree inventory surveys realized in five physiognomies types in the Brazilian Pampa biome. Later, 10 new samples were randomly selected from the same vegetation types to perform model validation. The DA generated four discriminant functions (DFs), where the first two had 88.4% power for discriminating groups (DF1 = 74.4% and DF2 = 14%). From the structural attributes used in the model, species richness, commercial height, and total height were related to DF1. Basal area and maximum stem diameter were related to DF2. The others DFs and structural variables have had less power of discriminating the groups. The DA classified 100% of the cases in their respective groups, showing a high efficiency of the chosen discriminating variables. The new forest samples inserted in the model were also classified with a small degree of error. The use of DA models should be enhanced because it is simple and more effective to express a forest classification model than the other descriptive multivariate methods. Keywords: forest physiognomy; forest structure; multivariate statistic; Rio Grande do Sul Continuous Forest Inventory. RESUMOA análise discriminante de Fisher (ADF) busca realizar uma combinação linear das variáveis independentes com objetivo de maximizar a separação de grupos preditos em um espaço reduzido bidimensional e ainda permitir que novas observações sejam classificadas ou não dentro dos grupos conhecidos a priori. Empregou-se a ADF utilizando oito variáveis estruturais obtidas de inventários sistemáticos do componente arbóreo (DAP>10 cm) realizados em cinco tipos florestais (total de 5 ha) distintos no bioma Pampa do sul do Brasil. Posteriormente foram sorteadas 10 novas amostras provenientes das mesmas fitofisionomias para realizar a validação do modelo. A AD gerou quatro funções discriminantes (FDs), sendo que as