Predicting vegetation change for adaptive management, under complex disturbance regimes, using monitoring data and expert judgements

Predictions of how vegetation will respond to management are important, allowing managers to simulate alternate management strategies, and compare their likely benefits and costs. Ideally, such models are connected to monitoring programs, allowing updates from measured ecosystem responses, in a learning cycle. Currently this remains rare, and there is often a gulf between models and the needs of managers. We demonstrate a tractable data-driven approach to modelling change in yearly increments, and apply it to grasslands in Australia. Our approach is novel in three ways: 1) It provides a method for integrating empirical field data (i.e. field-based monitoring) and expert-derived data (allowing models to be augmented with training scenarios that have not yet been observed). 2) It provides a framework for encoding complex disturbance events that explicitly accounts for the complexities of seasonal timing, duration, intensity and temporal legacies. 3) It provides a structure for testing predictions against independent field observations using the concept of a null model. We demonstrate successful year-by-year predictions of the foliage cover and species richness of selected plant species-groups, influenced by complex management regimes including fire, grazing, herbicide, and active restoration. We evaluate the performance of several models against measured annual vegetation change at 24 field sites, over 7–10 years. Multi-target Artificial Neural Networks outperform several other approaches, and are more accurate than a naïve null model. Our approach can be applied to most ecosystems, and can support an adaptive management cycle, by integrating field monitoring and expert predictions to improve models.