Comparing data subsets and transformations for reproducing an expert-based vegetation classification of an Australian tropical savanna

Abstract

Mapping vegetation communities requires considerable investment in field data collection, analysis and interpretation. The methods for data collection and analysis can significantly affect field time and the accuracy of the classifications. We test the ability of field data subsets and data pre-treatments to reproduce an intuitively derived vegetation classification within the Australian tropical savanna biome. The data subsets include all strata, upper strata, ground strata, and tree basal area. A range of multivariate techniques were used to describe patterns in the datasets as they related to the a priori vegetation classification. We tested the degree of floristic correlation among the data subsets and the extent to which several data transformations (square root, fourth root, presence or absence) improved the level of agreement between the numerically and the intuitively derived mapping units. Our results implied high redundancy in sampling both basal area and upper strata species cover, and the ground stratum was poorly correlated with the upper stratum. Across all statistical tests, the groups derived from analysis of square root-transformed upper stratum cover data were closely aligned with the expert classification. We propose that a numerical approach using an optimal dataset will produce a meaningful classification for vegetation mapping in poorly known Australian tropical savanna.