Process-based models are versatile tools for understanding and monitoring the carbon cycle in agricultural ecosystems, but their large-scale application is often hindered by their computational cost. In an article recently published in the European Journal of Agronomy, Viivi Aakula, Istem Fer and Julius Vira (Finnish Meteorological Institute) studied how machine learning can be used to emulate a dynamic grassland model to reduce computational costs, particularly in a demanding process such as calibration.
Making direct measurements on all relevant processes, stocks and fluxes at scale is both costly and susceptible to large uncertainties, making the quantification of carbon dynamics in agroecosystems a challenging task. However, when combined with measurements, process-based models of agroecosystems that simulate carbon dynamics can greatly help predicting carbon exchange. This is the case for instance of BASGRA (the Basic Grassland Model), which incorporates environmental conditions and management practices for grasslands to simulate daily greenhouse gas fluxes, harvest yield, and other parameters.
In their study, which can be found here, the authors developed an emulator for this model, enabling its scalability to large data sets (e.g., for calibration) and the more efficient exploration of different scenarios. Concretely, the authors trained a neural network in learning relationships between meteorological data, plant-soil properties and carbon exchange in agroecosystems, optimising it for time-dependent emulation of the BASGRA model. The emulator’s potential for low-cost model calibration was later assessed, comparing the performance with other methodologies.
As the publication shows, the neural network emulator presented high predictive accuracy, explaining over 95% of the variation of the process-based model for different output variables over a year, such as leaf area index (LAI), gross and net primary productivity (GPP and NPP), soil moisture and harvest yield, using weekly meteorological data, soil properties and a subset of model parameters as input. Moreover, the emulator was applied for calibrating model parameters against data from three Finnish sites, improving the predictive performance of GPP while also informing LAI and harvest carbon estimation. Overall, this study shows one of the first successful applications of emulating temporal dynamics of agroecosystem models to facilitate large-scale calibration, pointing towards its replication with more complex agroecosystem models with other crops and management practices.
