Defining reference conditions and determining change using discrete vegetation attributes at landscape scales
Spatially mapping change and trends in the state of vegetation can gauge where and how much of the landscape has been modified. This information can inform conservation investment, identify areas for restoration or to evaluate conservation and land management policies. Mapping change in vegetation can be challenging because quantities of long-term, site-based data are limited; vegetation communities can be structurally and compositionally complex; defining reference states from which to measure change is context dependent and different vegetation attributes respond differently to disturbance. These issues influence how we define reference states and measure change.
We addressed these challenges by assimilating archived site-based floristic data (n = 9752). Each site contained an inventory of species (including both native and non-native) with estimates of foliage cover (%). We transformed the floristic data into six discrete vegetation attributes (overstorey cover, midstorey cover, total groundcover, grassy groundcover, other growthforms in the groundcover and native species richness). To predict the reference state, we identified sites that represented least modified, relatively intact vegetation. Using only these exemplars, combined with a set of abiotic explanatory variables, we trained ensembles of artificial neural networks (ANN).
Models were validated using a subset (30%) of withheld data. Model performance varied. Correlation coefficients (r) ranged from 0.80 (overstorey cover) to 0.43 (native species richness). Trained networks were used to predict the reference state for each attribute over 11.5 million hectares. Raster-based analyses compared the predicted contemporary state of each vegetation attribute to predicted reference state for each vegetation attribute resulting in spatially explicit maps of change at the landscape scale.
Information linking history, ecology and conservation requires long-term data, but relatively few ecological studies can provide data needed to build spatially explicit, multi-attribute reference states. Conservation planning and land management are inherently spatial processes, and to be effective they need to be implemented over extensive geographic regions. We have employed predictive ecological modelling to recreate spatially explicit reference conditions at a landscape scale. Our work supports conservation practitioners, planners and policy makers with evidence needed to make informed decisions.