Terrestrial ecosystem models play an important role in ecological forecasting, but their predictive accuracy is often limited by poorly constrained parameters that cannot be directly observed. These parameters are typically inferred by calibrating model outputs against observations such as leaf area index (LAI) and aboveground biomass (AGB), but traditional Markov chain Monte Carlo (MCMC) methods can be computationally prohibitive because they require many expensive forward model evaluations, and gradient-based methods are often impractical because model derivatives are unavailable or costly to compute. We present a scalable parameter calibration framework based on ensemble Kalman inversion (EKI) with likelihood tempering. We apply this approach to the SIPNET ecosystem model across >200 FLUXNET and NEON sites, calibrating a influential parameters identified through sensitivity analysis, against a combination of bottom-up (e.g., NEE, LE) and remotely-sensed (e.g., LAI, AGB) observations. In our experiments, accurate calibration was achieved with relatively small ensembles (~50). Relative to the prior ensemble, calibrated parameter ensembles improved representation of seasonal canopy dynamics and aboveground biomass across sites. EKI results also showed increasing agreement with posterior distributions obtained by MCMC over tempering iterations. Across sites, calibration reduced median seasonal amplitude error in LAI by 43%, indicating improved representation of the annual canopy cycle. For AGB, calibration reduced median CRPS by 67% and RMSE by 78%. Calibrated ensembles were also sharper, with median interval width reduced by 66% while maintaining coverage of the nominal 90% interval. These results demonstrate that our framework provides a scalable approach for parameter inference in complex ecosystem models.

Keywords: Data assimilation, Bayesian calibration, Uncertainty quantification, Terrestrial carbon cycling