Improving ecosystem model development with machine learning: a full hybrid approach

Making projections of ecological systems under environmental change is central to many disciplines. Process-based models aim to represent core ecological mechanisms governing ecosystem dynamics, which can then be valuable for projecting change under novel environmental conditions. Yet, as our ecological understanding evolves, updating parameter information can be challenging. In addition, classical statistical approaches to fitting functional relationships often miss the complexity of interacting, non-linear dynamics, which can limit the predictive capacity of models. As such, a growing body of work suggests that the integration of modern machine learning might help to improve the representation of key ecological dynamics within such process-based models.  Here, we present a case study, using machine learning to identify key relationships between relative growth, biomass and mortality compared to classical regression methods. Our results suggest that the inclusion of a deep neural network (DNN) into a “theory-driven” process based global vegetation model can greatly improve model predictions of patch level forest structure and vegetation dynamics. This hybrid approach offers both the benefits of interpretability and physically-realistic structure, combined with the depth of information contained in big datasets and the flexibility of model machine learning.