The effects of constraining variables on parameter optimization in carbon and water flux modeling over different forest ecosystems

• We conducted an inverse modeling analysis in carbon and water flux modeling.
• EC observations and LAI from MODIS were used as constraining variables.
• Model simulated results were mostly affected by photosynthesis-related parameters.
• The model performance depended largely on the combinations of constraining variables.

The ability of terrestrial biogeochemical models in predicting land-atmospheric carbon and water exchanges is largely hampered by the insufficient characterization of model parameters. The direct observations of carbon/water fluxes and the associated environmental variables from eddy covariance (EC) flux towers provide a notable opportunity to examine the underlying processes controlling carbon and water exchanges between terrestrial ecosystems and the atmosphere. In this study, we applied the Metropolis simulated annealing technique to conduct parameter optimization analyses of a process-based biogeochemical model, simplified PnET (SIPNET), using a variety of constraining variables from EC observations and leaf area index (LAI) from MODIS at three ChinaFLUX forest sites: a temperate mixed forest (CBS), a subtropical evergreen coniferous plantation (QYZ) and a subtropical evergreen broad-leaved forest (DHS). Our analyses focused on (1) identifying the key model parameters influencing the simulation of carbon and water fluxes with SIPNET; (2) evaluating how different combinations of constraining variables influence parameter estimations and associated uncertainties; and (3) assessing the model performance with the optimized parameterization in predicting carbon and water fluxes in the three forest ecosystems. Our sensitivity analysis indicated that, among three different forest ecosystems, the prediction of carbon and water fluxes was mostly affected by photosynthesis-related parameters. The performances of the model simulations depended on different parameterization schemes, especially the combinations of constraining variables. The parameterization scheme using both net ecosystem exchange (NEE) and evapotranspiration (ET) as constraining variables performed best with most well-constrained parameters. When LAI was added to the optimization, the number of well-constrained model parameters was increased. In addition, we found that the model cannot be well-parameterized with only growing-season observations, especially for those forest ecosystems with distinct seasonal variation. With the optimized parameterization scheme using both NEE and ET observations all year round, the SIPNET were able to simulate the seasonal and inter-annual variations of carbon and water exchanges in three forest ecosystems.