Extending site-based observations to predict the spatial patterns of vegetation structure and composition

Abstract Context Conservation planning and land management are inherently spatial processes that are most effective when implemented over large areas. Objectives Our objectives were to (i) use existing plot data to aggregate species inventories to growth forms and derive indicators of vegetation structure and composition and ii) generate spatially-explicit, continuous, landscape scaled models of these discrete vegetation indicators, accompanied by maps of model uncertainty. Method Using a case study from New South Wales, Australia, we aggregated floristic observations from 7234 sites into growth forms. We trained ensembles of artificial neural networks (ANN) to predict the distribution of these indicators over a broad region covering 11.5 million hectares. Importantly, we show spatially explicit models of uncertainty so that end-users have a tangible and transparent means of assessing models. Results Our key findings were firstly, widely available site-based floristic records can be used to derive aggregated indicators of the structure and composition of plant growth forms. Secondly, ANNs are a powerful method to predict continuous patterns in complex, non-linear data (Pearson’s correlation coefficient 0.83 (total native vegetation cover) to 0.42 (forb cover)). Thirdly, maps of the standardised residual error give insight into model performance and provide an assessment of model uncertainty in specific locations. Conclusions Spatially explicit, continuous representations of vegetation composition and structural complexity can add considerable value to conventional maps of vegetation extent or community type. This application has the potential to enhance the capacity for conservation planners, landscape managers and policy-makers to make informed decisions across landscape and regional scales.