Emulator-based calibration of a dynamic grassland model using recurrent neural networks and Hamiltonian Monte Carlo
2025
Aakula, Viivi | Fer, Istem | Vira, Julius | Ilmatieteen laitos | Finnish Meteorological Institute | 0000-0001-8236-303X | 0000-0003-2696-6885
Process-based models are versatile tools for understanding and monitoring the carbon cycle in agricultural ecosystems. Their large-scale application is often hindered by computational challenges, particularly in iterative processes such as calibration. Machine learning emulation offers a practical solution by substituting the process-based model with a computationally less expensive approximation. In this study, we build a neural network emulator for a dynamic agroecosystem model. The procedure included obtaining training data from model simulations, defining hyperparameter space for the neural network model, hyperparameter optimization and training the network and evaluation with 5-fold cross validation. We apply the method for BASGRA, a process-based model for managed grasslands, to predict weekly values of leaf area index (LAI) gross and net primary productivity (GPP and NPP), soil moisture and harvest yield over a year, with weekly meteorological data and soil properties, and a subset of model parameters as input. The emulator explained over 95% of the variation of the process-based model for each output variable, showing high predictive accuracy. We then apply the emulator for calibrating model parameters against GPP data from three Finnish sites and perform 71 calibration experiments with different calibration and validation data sets. The predictive performance of GPP is improved across all experiments, with 33%–64% reduction in RMSE, while also informing LAI and harvest carbon estimation. This study shows one of the first successful applications of emulating temporal dynamics of agroecosystem models to facilitate large-scale calibration.
显示更多 [+] 显示较少 [-]