Predictive models of ecosystem productivity in water-limited conditions

Accurate predictions of environmental controls on ecosystem photosynthesis are essential for understanding the impacts of climate change and extreme events on the carbon cycle and the provisioning of ecosystem services. Widely used machine learning models for simulating ecosystem photosynthesis do not consider temporal dependencies in the data, even though process-understanding suggests these should exist through effects such as soil moisture stress. Here, we investigate the impact of accounting for temporal structure in modelling ecosystem photosynthesis.

Using time-series measurements of ecosystem fluxes paired with measurements of meteorological variables from a network of globally distributed sites and remotely sensed vegetation indices, we train three different models to predict ecosystem gross primary production (GPP): a mechanistic, theory-based photosynthesis model, a straightforward multilayer perceptron (MLP) and a recurrent neural network (Long-Short-Term Memory, LSTM). Through comparisons of patterns in model error, we assess the ability of these models to account for temporal dependencies that arise through effects such as soil moisture stress and cold acclimation. We further investigate the influence of different environmental factors on the generalisability across space.

We find that both deep learning models outperform the mechanistic model, and that the LSTM performs best with an R2 of 0.74. In particular, model skill is consistently good across moist sites with strong seasonality. Model error tends to increase with increasing potential cumulative water deficit, in particular in ecosystems with evergreen vegetation. Generalisation patterns reveal that the LSTM tends to be more successful than the MLP in adapting to arid environments and to ecosystems with seasonal dryness, suggesting an advantage of recurrent models for GPP modelling in those conditions. However, there remains a large variability in model skill across arid sites.

Our findings reveal the impacts on model error due to unknown effects of water limitation when predicting fluxes across different ecosystems. Due to climate change, temporal dependencies such as water limitation are becoming more prevalent, making an accurate representation of such processes increasingly important for modelling ecosystem function.