Bioregion classification using model-based clustering: a case study in North Eastern Queensland

The Interim Biogeographic Regionalisation of Australia (IBRA; Environment Australia, 2000) is a planning framework defining land areas comprised of interacting ecosystems repeated across the landscape. In many states these bioregions are currently arrived at by consensus of an expert panel (Neldner et al. 2004) and well accepted as a spatial unit for planning and environmental management. This work was motivated by the need to define these regions in a scientifically defensible way to justify any decisions made on the basis that they are representative of broad environmental assets. The present bioregional boundaries of Queensland version 4.3 are shown in Figure 1. The case study is situated in North Eastern Queensland in an area where a boundary change was proposed. This research investigates at the meso-scale the success of broad climate and soil variables in identifying patterns and processes. We report results based on three bioclimate and soil variables, suggested through exploratory analysis and model sensitivity: temperature seasonality (bc04); annual precipitation (bc12); and B horizon available water holding capacity (baw). This paper compares a range of statistical methods for bioregion classification, within a continuum of data-driven to expert-driven, including Bayesian methods. Model-based clustering moves away from traditional methods which delineate boundaries, instead assessing similarity between and within geographic regions and environmental envelopes. Bayesian statistical modelling enables explicit input of expert prior knowledge during development of bioregions. We assessed two alternative prior knowledge bases: vegetation communities or existing bioregion boundaries. In data poor areas expert defined boundaries are feasible, but subjective. Vegetation-based priors can be considered more objective, although they require subjective identification of communities, and are a useful alternative to expert boundaries. This study confirmed that experts contribute knowledge beyond what is currently mapped on bioclimate and soils. The Bayesian model-based approach has significant benefits in assessing impact of different types of expert knowledge for bioregions-either mapped communities or boundaries-as well as for quantifying precision of modelled regions. Practically we found that the Frequentist modelbased approach was useful in initial stages of modelling. The distance-metric based approaches to clustering though relatively simple to implement provide qualitatively different boundaries, and require an unwieldy process for obtaining predictions, for which no assessment of uncertainty is available. For bioregionalisation of new areas, expert-defined boundaries may still play a role, although this has now been demonstrated to be more useful when combined with bioclimate and soils datasets in a Bayesian framework. In data-rich areas, the Frequentist model-based approach may suffice. (Figure Presented)