Ecosystem model optimization using in-situ flux observations: benefit of monte-carlo vs. variational schemes and analyses of the year-to-year model performances

Diego Santaren, Philippe Peylin, Cédric Bacour, Philippe Ciais, Bernard Longdoz2013 · DOI

TL;DR

This paper explores optimizing complex ecosystem models using observational data. They found that a Monte Carlo approach (Genetic Algorithm) is more robust than traditional gradient-based methods for finding optimal parameters in models like ORCHIDEE. Crucially, using multiple years of data, especially those with varied weather, is essential for accurately simulating year-to-year ecosystem flux variations.

What Problem Does This Paper Address?

Terrestrial ecosystem models are vital for understanding how forests respond to climate change and rising CO2 levels. These models simulate complex processes like photosynthesis and respiration, but their accuracy is limited by unknown or poorly calibrated parameters. This makes long-term predictions of carbon and water fluxes unreliable. The challenge is to effectively constrain these model parameters using observational data to improve their predictive power, especially for inter-annual variability.

Methods

The authors developed a data assimilation framework to optimize the parameters of the ORCHIDEE land surface model. They compared two optimization approaches:

  • Variational Scheme (Gradient-based): This method uses gradients to find the minimum of a cost function. The paper specifically tested the L-BFGS-B algorithm.
  • Monte Carlo Approach (Genetic Algorithm - GA): This method uses a stochastic search over the parameter space, mimicking natural selection to find optimal solutions.

They also conducted "twin experiments" where synthetic data generated by the model itself was used to evaluate the algorithms' performance without the bias of real-world data uncertainties or model structural deficiencies.

Datasets

The study utilized eddy-covariance observations of CO2 (Net Ecosystem Exchange - NEE) and latent heat (LE) fluxes from a temperate beech forest site in Hesse, eastern France. The data spanned four years: 2001, 2002, 2003, and 2004. These years were characterized by different weather regimes, including wet (2001-2002) and dry (2003-2004) summers.

Key Insights and Ideas

  • GA vs. Variational: For complex models with many parameters, the Genetic Algorithm (GA) is significantly more robust than gradient-based methods in finding the global minimum of the cost function. Gradient-based methods are prone to getting stuck in local minima, leading to less reliable parameter estimates.
  • The Power of Multiple Years: Optimizing parameters using data from a single year is insufficient for capturing year-to-year variations. The model's predictive skill suffers when applied to years with different weather conditions.
  • Contrasted Weather is Key: Including years with diverse weather patterns (e.g., wet and dry summers) is crucial for optimizing parameters that control processes like photosynthesis, particularly their temperature sensitivity.
  • Model Structure Matters: While parameter optimization improves model fit, significant discrepancies can still highlight structural deficiencies in the model itself, especially in simulating processes like evapotranspiration during dry periods.

Key Results

  • The GA approach resulted in significantly lower normalized cost functions (mean of 0.02, std dev 0.01) compared to the BFGS algorithm (mean of 0.29, std dev 0.14) in twin experiments.
  • Optimizing with four years of data (2001-2004) led to a substantial improvement in the model's fit to NEE and LE fluxes, achieving a typical modeling efficiency of 92% (Nash-Sutcliffe coefficient).
  • A single-year optimization resulted in a mean reduction of the cost function by 68.5% when applied to other years, compared to 92% when applied to the optimization year.
  • The four-year optimized model (ORC4Y) simulated the mean seasonal cycle of NEE well, with a slight underestimation of cumulative NEE in the fall.
  • Only a few parameters (9 out of 28) were tightly constrained (error reduction > 90%) when using four years of data, indicating that many parameters remain uncertain.

Limitations and Future Work

  • Equifinality: The study acknowledges the "equifinality problem," where different combinations of parameters can lead to similar model fits, making it difficult to distinguish truly optimal parameters from those that compensate for model deficiencies.
  • Parameter Correlations: High posterior error correlations between parameters suggest that observations may not provide independent information to constrain all parameters.
  • Model Structure: The paper points out that even with optimized parameters, the model struggles to fully capture certain aspects of evapotranspiration, suggesting potential structural limitations in its representation of stomatal conductance and humidity gradients.
  • Future Work: The authors suggest investigating the GA's performance across multiple sites with different plant functional types. They also propose incorporating additional data types (e.g., soil carbon content, flux chamber measurements) to better constrain parameters related to annual and interannual processes. Further research is needed to understand if the observed biases are due to model structure or parameterization.

Summary updated

Abstract

Terrestrial ecosystem models can provide major insights into the responses of Earth's ecosystems to environmental changes and rising levels of atmospheric CO2. To achieve this goal, biosphere models need mechanistic formulations of the processes that drive the ecosystem functioning from diurnal to decadal time-scales. However, the subsequent complexity of model equations is associated with unknown or poorly calibrated parameters that limit the accuracy of long-term simulations of carbon or water fluxes and their inter-annual variations. In this study, we develop a data assimilation framework to constrain the parameters of a mechanistic land surface model (ORCHIDEE) with eddy-covariance observations of CO2 and latent heat fluxes made during the years 2001–2004 on the temperate beech forest site of Hesse, in eastern France. As a first technical issue, we show that for a complex process-based model such as ORCHIDEE with many (28) parameters to be retrieved, a Monte Carlo approach (genetic algorithm, GA) provides more reliable optimal parameter values than a gradient-based minimization algorithm (variational scheme). The GA allows finding the global minimum more efficiently whilst the variational scheme often provides values relative to local minima. The ORCHIDEE model is then optimized for each year, and for the whole 2001–2004 period. We first find that a reduced (<10) set of parameters can be tightly constrained by the eddy-covariance observations with a typical error reduction of 90%. We then show that including contrasted weather regimes (dry in 2003 and wet in 2002) is necessary to optimize few specific parameters (like the temperature dependence of the photosynthetic activity). Furthermore, we find that parameters inverted from four years of flux measurements are successful at enhancing the model fit to the data at several time-scales (from monthly to interannual) resulting to a typical modeling efficiency of 92% over the 2001–2004 period (Nash–Sutcliffe coefficient). This suggests that ORCHIDEE is able to robustly predict the fluxes of CO2 and latent heat of a temperate beech forest after optimisation. Finally, it is shown that using only one year of data does not produce robust enough optimized parameter sets in order to simulate properly the year-to-year flux variability. This emphasizes the need to assimilate data over several years, including contrasted weather regimes, to improve the simulated flux inter-annual variability.
BibTeX
@misc{santaren2013ecosystem,
  author = {Diego Santaren and Philippe Peylin and Cédric Bacour and Philippe Ciais and Bernard Longdoz},
  title = {{Ecosystem model optimization using in-situ flux observations: benefit of monte-carlo vs. variational schemes and analyses of the year-to-year model performances}},
  year = {2013},
  doi = {10.5194/bgd-10-18009-2013},
}

Loading related content...

Figure