Yao Zhang

Grazing lands, including both pastures and rangelands, are among the most widespread land uses on Earth, covering ~67% of the agricultural land [1]. They are estimated to store between 10% and 30% ...

An abundant amount of different data is required to calibrate soil organic carbon (SOC) models to represent ecosystems at large-scale. However, due to challenges related to model state projections, this calibration becomes very computationally heavy with traditional calibration methods. Here, we test 4-Dimensional Ensemble Variational data assimilation (4DEnVar) method to parameterize the MEMS v1 SOC model using data from the LUCAS network and compare its performance against MCMC calibration. Additionally, we performed an experiment where we adjusted the litter input calculation to see if the two calibration methods react differently to the change. The total SOC projections from both parameterizations showed similar improvements though the produced parameter sets differed. A thorough analysis revealed that the detailed SOC states differed from each other, but we also lacked information to determine which parameter set was closer to the truth. Furthermore, changing the litter input partition highlighted how much that assumption affects the calibration results with both methods. Our results here establish 4DEnVar as an applicable calibration method for SOC models but also highlight the need for more nuanced validation methods, as well as careful examination on how different data sets affect the model calibration.

An abundant amount of different data is required to calibrate soil organic carbon (SOC) models to represent ecosystems at large-scale. However, due to challenges related to model state projections, this calibration becomes very computationally heavy with traditional calibration methods. In this work, we test 4-Dimensional Ensemble Variational data assimilation (4DEnVar) method to parameterize the MEMS v1 SOC model using data from the LUCAS soil sampling network and compare its performance against MCMC calibration. Comparing the total SOC projections from both parameterizations to the validation datasets showed similar improvements even though the produced parameter sets differed. A thorough analysis revealed that the detailed SOC states were not similar to a degree that is meaningful for future predictions, but we also lacked information to determine which parameter set was closer to the truth. Our results here establish 4DEnVar as an applicable calibration method for SOC models but also highlight the need for more nuanced validation methods as well careful examination how different data sets affect the model calibration.

An abundant amount of different data is required to calibrate soil organic carbon (SOC) models to represent ecosystems at large-scale. However, due to challenges related to model state projections, this calibration becomes very computationally heavy with traditional calibration methods. In this work, we test 4-Dimensional Ensemble Variational data assimilation (4DEnVar) method to parameterize the MEMS v1 SOC model using data from the LUCAS soil sampling network and compare its performance against MCMC calibration. Comparing the total SOC projections from both parameterizations to the validation datasets showed similar improvements even though the produced parameter sets differed. A thorough analysis revealed that the detailed SOC states were not similar to a degree that is meaningful for future predictions, but we also lacked information to determine which parameter set was closer to the truth. Our results here establish 4DEnVar as an applicable calibration method for SOC models but also highlight the need for more nuanced validation methods as well careful examination how different data sets affect the model calibration.

An abundant amount of different data is required to calibrate soil organic carbon (SOC) models to represent ecosystems at large-scale. However, due to challenges related to model state projections, this calibration becomes very computationally heavy with traditional calibration methods. In this work, we test 4-Dimensional Ensemble Variational data assimilation (4DEnVar) method to parameterize the MEMS v1 SOC model using data from the LUCAS soil sampling network and compare its performance against MCMC calibration. Comparing the total SOC projections from both parameterizations to the validation datasets showed similar improvements even though the produced parameter sets differed. A thorough analysis revealed that the detailed SOC states were not similar to a degree that is meaningful for future predictions, but we also lacked information to determine which parameter set was closer to the truth. Our results here establish 4DEnVar as an applicable calibration method for SOC models but also highlight the need for more nuanced validation methods as well careful examination how different data sets affect the model calibration.

An abundant amount of different data is required to calibrate soil organic carbon (SOC) models to represent ecosystems at large-scale. However, due to challenges related to model state projections, this calibration becomes very computationally heavy with traditional calibration methods. In this work, we test 4-Dimensional Ensemble Variational data assimilation (4DEnVar) method to parameterize the MEMS v1 SOC model using data from the LUCAS soil sampling network and compare its performance against MCMC calibration. Comparing the total SOC projections from both parameterizations to the validation datasets showed similar improvements even though the produced parameter sets differed. A thorough analysis revealed that the detailed SOC states were not similar to a degree that is meaningful for future predictions, but we also lacked information to determine which parameter set was closer to the truth. Our results here establish 4DEnVar as an applicable calibration method for SOC models but also highlight the need for more nuanced validation methods as well careful examination how different data sets affect the model calibration.

An abundant amount of different data is required to calibrate soil organic carbon (SOC) models to represent ecosystems at large-scale. However, due to challenges related to model state projections, this calibration becomes very computationally heavy with traditional calibration methods. In this work, we test 4-Dimensional Ensemble Variational data assimilation (4DEnVar) method to parameterize the MEMS v1 SOC model using data from the LUCAS soil sampling network and compare its performance against MCMC calibration. Comparing the total SOC projections from both parameterizations to the validation datasets showed similar improvements even though the produced parameter sets differed. A thorough analysis revealed that the detailed SOC states were not similar to a degree that is meaningful for future predictions, but we also lacked information to determine which parameter set was closer to the truth. Our results here establish 4DEnVar as an applicable calibration method for SOC models but also highlight the need for more nuanced validation methods as well careful examination how different data sets affect the model calibration.

Grazinglands store substantial soil organic carbon (SOC), yet their potential to act as net carbon (C) sinks depends on management-driven net ecosystem CO2 exchange (NEE). Process-based models capable of representing contrasting management strategies are essential for evaluating mitigation potential. However, the uncertainty inherent in model outputs is often overlooked, which can limit the reliability of predictions. Here, we apply the MEMS 3.0 model to quantify uncertainty in SOC and NEE predictions and evaluate model performance across diverse U.S. grazing ecoregions. We conducted a Bayesian calibration with observations of SOC and its fractions, and annual NEE measurements from six NEON grazing sites. Model evaluation was performed with an independent dataset collected from four experimental sites located in Oklahoma, Michigan, and Wyoming USA under prescriptive and adaptive grazing management treatments. The model performed well in predicting baseline SOC and estimating weekly ecosystem fluxes across sites. Although annual NEE estimates exhibited some discrepancies relative to flux-tower observations, ~90% of measured fluxes fell within simulated posterior predictive intervals. Moreover, the model is consistent with the flux-observations, demonstrating no significant treatment differences between prescriptive and adaptive grazing treatments at the Oklahoma and Wyoming sites. We demonstrated that MEMS 3.0 can represent SOC dynamics and ecosystem fluxes in grazinglands across contrasting climates. Our results show that neglecting uncertainty in measured and simulated fluxes can lead to misleading model-data comparisons. These findings highlights the importance of uncertainty quantification for robust interpretation of predicted grazing management outcomes and for supporting credible climate change mitigation and C accounting frameworks.

An abundant amount of different data is required to calibrate soil organic carbon (SOC) models to represent ecosystems at large-scale. However, due to challenges related to model state projections, this calibration becomes very computationally heavy with traditional calibration methods. In this work, we test 4-Dimensional Ensemble Variational data assimilation (4DEnVar) method to parameterize the MEMS v1 SOC model using data from the LUCAS soil sampling network and compare its performance against MCMC calibration. Comparing the total SOC projections from both parameterizations to the validation datasets showed similar improvements even though the produced parameter sets differed. A thorough analysis revealed that the detailed SOC states were not similar to a degree that is meaningful for future predictions, but we also lacked information to determine which parameter set was closer to the truth. Our results here establish 4DEnVar as an applicable calibration method for SOC models but also highlight the need for more nuanced validation methods as well careful examination how different data sets affect the model calibration.