Near-term ecological forecasting: A Bayesian framework for modelling spatio-temporal population dynamics in extreme environments

Abstract

Ecological forecasting plays a crucial role in anticipating ecosystem dynamics and guiding conservation decisions, particularly in arid ecosystems where species respond to extreme environmental variability and complex interactions. This thesis focuses on forecasting the population abundance of small mammals in central Australia, integrating long-term monitoring data with advanced statistical methods to improve model accuracy, ecological relevance, and practical use. Using a combination of Multivariate Autoregressive State-Space (MARSS) models and Multivariate Generalized Additive Models (MVGAM), the thesis explores how model structure, missing data, training length, and forecast horizon influence predictive performance. While MARSS models are effective in capturing structured population dynamics across space, they struggle with computational limitations and nonlinear processes. Simulated experiments evaluating imputation methods show that although differences in predictive accuracy are small, multiple imputation better preserves underlying ecological patterns and uncertainty. Forecasting performance is shown to depend more on model flexibility and the ability to represent latent ecological processes than on the quantity of training data alone. Short-term forecasts using projected climate data reveal divergent species responses across sites and climate scenarios, underscoring the importance of near-term ecological forecasting. Overall, this work demonstrates the value of combining ecological understanding with flexible, interpretable modelling approaches to improve forecasts in variable environments and inform conservation planning under climate change.