Paredes et al. (2025) rsofun v5.1: a model-data integration framework for simulating ecosystem processes

Identification

• Journal: Geoscientific model development
• Year: 2025
• Date: 2025-12-10
• Authors: Josefa Arán Paredes, Fabian Bernhard, Koen Hufkens, Mayeul Marcadella, Benjamin D. Stocker
• DOI: 10.5194/gmd-18-9855-2025

Research Groups

• Institute of Geography, University of Bern, Hallerstrasse 12, 3012 Bern, Switzerland
• Oeschger Center for Climate Change Research (OCCR), University of Bern, Hochschulstrasse 6, 3012 Bern, Switzerland

Short Summary

This study introduces and evaluates rsofun v5.1, an R package providing a computationally efficient implementation of the P-model for simulating ecosystem photosynthesis and trait acclimation, coupled with Bayesian model-data integration. Multi-target calibration using ecosystem fluxes and leaf traits demonstrated robust parameter estimation and unbiased predictions that generalize well across diverse environmental conditions.

Objective

• To demonstrate and evaluate the rsofun R package for computationally efficient P-model simulations and Bayesian model-data integration, specifically assessing if the P-model can be consistently calibrated to multiple observational targets (ecosystem gross primary productivity, leaf carbon-13 isotopic discrimination, and the ratio of maximum biochemical rates of carboxylation to electron transport) to yield unbiased parameter estimates and robust predictions on independent test data.

Study Configuration

• Spatial Scale: Site-scale simulations conducted across globally distributed sites, comprising 193 training sites and 231 test sites, covering diverse vegetation types and biomes.
• Temporal Scale: Daily for GPP simulations and meteorological forcing, multi-day for acclimation processes, and growing season averages for leaf traits (Δ and VJ). GPP training data included over 12 years of observations per site.

Methodology and Data

• Models used:
  • P-model (optimality-based light use efficiency model for gross primary production and photosynthetic trait acclimation)
  • rsofun R package (implementation of the P-model and a model-data integration framework)
  • Farquhar–von Caemmerer–Berry (FvCB) model for C3 photosynthesis
  • Bucket model for soil water content (based on the SPLASH model)
  • Bayesian calibration using Markov chain Monte Carlo (MCMC) sampling with the DREAMzs algorithm, implemented via the BayesianTools R package.
• Data sources:
  • Gross Primary Productivity (GPP): Daily time series from FluxDataKit (FDK v3.4.2), combining consistently processed eddy-covariance data from multiple regional networks (e.g., FLUXNET2015, ICOS).
  • Ratio of maximum biochemical rates (VJ): Compilation of top canopy VJ observations from Smith et al. (2019) and other sources.
  • Leaf carbon-13 isotopic discrimination (Δ): Global dataset of C3 plant leaf material isotopic signatures from Cornwell et al. (2018).
  • Meteorological Forcing (GPP): Daily meteorological measurements (mean daytime air temperature, vapor pressure deficit, daily minimum/maximum temperatures, daily sum of precipitation, daily means of net radiation, atmospheric pressure, CO2 concentration) from FluxDataKit.
  • Meteorological Forcing (Δ, VJ): Growing season average climate conditions derived from the global WorldClim dataset (monthly averages), ETOPO-1 digital elevation model (for atmospheric pressure), and Mauna Loa CO2 record.

Main Results

• The rsofun R package provides an efficient framework for P-model simulations and Bayesian model-data integration, enabling robust parameter estimation and uncertainty quantification.
• Multi-target Bayesian calibration, combining ecosystem GPP fluxes and leaf traits (Δ and VJ), yielded robust parameter estimates and unbiased predictions that generalized well across diverse environments, with similar prediction-observation residuals for both training and independent test data.
• Specific observational targets provided distinct constraints: Δ data primarily constrained the unit cost ratio (β), while VJ data constrained β and the unit cost of electron transport (c*), though with strong correlations between them.
• A step-wise Bayesian calibration approach (using posteriors from Δ and VJ calibration as priors for GPP calibration) was essential to mitigate parameter biases, particularly for β, which otherwise tended to unrealistically low values.
• The step-wise approach (Setup S6) resulted in parameter estimates consistent with previous literature: β ≈ 208 (unitless), c* ≈ 0.58 (unitless), and φ0* ≈ 0.05 mol mol⁻¹.
• The acclimation time scale (τ) was estimated to be 86400 seconds (1 day), suggesting instantaneous acclimation, which deviates from prior estimates of 1.21 x 10⁶ to 1.30 x 10⁶ seconds (14 to 15 days).
• Remaining correlations between φ0* and c* indicate potential equifinality or model structural errors, particularly concerning the representation of water stress effects.
• While overall predictions were unbiased, the model showed specific shortcomings at certain sites, such as underestimation of GPP during dry conditions (e.g., US-Var) and early growing season overestimation (e.g., US-MMS, US-PFa).

Contributions

• Development and open-source release of rsofun v5.1, an R package for computationally efficient P-model simulations and flexible Bayesian model-data integration.
• Demonstration of a robust methodology for parameterizing a mechanistic vegetation model (P-model) by integrating diverse, multi-level observational data (ecosystem fluxes and leaf traits) within a comprehensive Bayesian framework.
• Elucidation of the complementary information content of different observational data types for constraining specific model parameters and understanding their interdependencies.
• Identification and resolution of challenges in multi-target model calibration, highlighting the necessity and effectiveness of a step-wise Bayesian approach to address structural uncertainty and unbalanced data.
• Pinpointing specific model structural deficiencies (e.g., representation of soil moisture stress, acclimation time scale) through the calibration process, thereby guiding future model improvements.
• Provision of a user-friendly, efficient, and flexible blueprint for model-data assimilation in terrestrial photosynthesis modeling.

Funding

• Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung (grant no. PCEFP2_181115)
• LEMONTREE (Land Ecosystem Models based On New Theory, obseRvations and ExperimEnts) project, funded by Eric and Wendy Schmidt by recommendation of the Schmidt Futures program

Citation

@article{Paredes2025rsofun,
  author = {Paredes, Josefa Arán and Bernhard, Fabian and Hufkens, Koen and Marcadella, Mayeul and Stocker, Benjamin D.},
  title = {rsofun v5.1: a model-data integration framework for simulating ecosystem processes},
  journal = {Geoscientific model development},
  year = {2025},
  doi = {10.5194/gmd-18-9855-2025},
  url = {https://doi.org/10.5194/gmd-18-9855-2025}
}