Inverse Modeling of Hydrologic Parameters in CLM4 via Generalized Polynomial Chaos in the Bayesian Framework

Guang Lin, Georgios Karagiannis, Zhangshuan Hou, Maoyi Huang, Guang Lin 0001Comput 2022 · DOI

TL;DR

This paper introduces a novel Bayesian framework for estimating crucial hydrological parameters in the Community Land Model (CLM4). By combining generalized polynomial chaos (gPC) expansions with Bayesian variable selection, the method efficiently quantifies parameter uncertainty and identifies the most influential parameters for simulating land surface processes.

What Problem Does This Paper Address?

Land surface models (LSMs) like CLM4 are essential tools for simulating Earth's climate and water cycles. However, these models rely on numerous parameters that are often poorly constrained by observational data, leading to significant uncertainty in their predictions. Accurately estimating these parameters is crucial for improving model reliability. Traditional methods for parameter estimation and uncertainty quantification can be computationally prohibitive due to the complexity and high dimensionality of LSMs. This paper tackles the challenge of efficiently and robustly estimating these uncertain hydrological parameters within a Bayesian framework, specifically addressing the computational burden.

Methods

The core of this work lies in a two-step Bayesian methodology:

  1. Surrogate Model Construction: Instead of directly using the computationally expensive CLM4 model, the authors build a "surrogate model." This surrogate is a fast approximation of the CLM4's input-output relationship (specifically, latent heat flux as a function of hydrological parameters). They employ Generalized Polynomial Chaos (gPC) expansion for this purpose. A key innovation is the use of a Bayesian procedure for variable selection within the gPC framework. This allows them to not only build an accurate surrogate but also to simultaneously quantify the importance of different gPC terms (and thus, the input parameters) by estimating their inclusion posterior probabilities. This approach effectively handles bases selection uncertainty.
  2. Bayesian Inverse Modeling: Once the surrogate model is built, it's integrated into a Bayesian inverse modeling framework. This involves:
  3. Likelihood Function: Defining how well the model output (from the surrogate) matches observed data (latent heat flux).
  4. Prior Distributions: Specifying initial beliefs about the parameter values.
  5. Posterior Distribution: Using Bayes' theorem to update these beliefs based on the observed data.
  6. Markov Chain Monte Carlo (MCMC): Employing MCMC methods (specifically a Gibbs sampler) to sample from the (often intractable) posterior distribution of the parameters.

The paper highlights that their Bayesian gPC approach is "fully automatic," meaning it doesn't require manual tuning of algorithmic parameters.

Datasets

The study utilizes observational data from the US-ARM (ARM Southern Great Plains site, Lamont, Oklahoma), part of the FLUXNET database. This dataset provides half-hourly observations of ecosystem CO2, heat fluxes, and meteorological data. For parameter estimation, they used processed and gap-filled daily and monthly averaged data for observed latent heat fluxes and runoff measurements.

Key Insights and Ideas

  • gPC as a Powerful Surrogate: gPC expansions prove to be an effective and computationally efficient way to approximate the complex input-output relationships of land surface models.
  • Bayesian Variable Selection for gPC: Integrating Bayesian variable selection directly into the gPC construction is a significant contribution. It allows for automatic identification of important parameters and gPC terms, mitigating issues like overfitting and dimensionality reduction.
  • Quantifying Parameter Importance Probabilistically: The method provides probabilistic measures (inclusion posterior probabilities) for the significance of each input parameter, offering a more robust understanding of parameter influence than deterministic methods.
  • Seasonal Variability in Parameter Importance: The study reveals that the importance of hydrological parameters for simulating latent heat flux varies significantly across seasons, with parameters showing higher influence during the warmer months (May-August). This is intuitively linked to higher latent heat flux and variability during these periods.
  • Consistent Parameter Estimates: The posterior distributions obtained for the parameters are generally narrower than their prior distributions, indicating successful calibration and a reduction in uncertainty.

Key Results

  • The methodology successfully identified a subset of five significant hydrological parameters (Fmax, Cs, Fdrai, Qdm, and B) that are most influential for simulating latent heat flux in the CLM4 model for the US-ARM site.
  • The Bayesian inverse modeling step successfully shortened the ranges of possible values for most input parameters compared to their prior distributions, demonstrating effective calibration.
  • The predictive distributions of latent heat flux, generated using the derived surrogate model and posterior parameter estimates, showed good agreement with observed values, validating the methodology.

Limitations and Future Work

  • Model Structural Uncertainty: The paper focuses on parameter uncertainty and does not explicitly address model structural uncertainty (i.e., potential issues with the underlying equations or assumptions within CLM4 itself).
  • Generalizability: While the method is applied to the US-ARM site, further validation across different geographical locations and climate regimes would be beneficial.
  • Comparison with Other Methods: The authors mention that future work will involve comparing their proposed method with other surrogate-based approaches, such as neural networks and generalized linear models.
  • Extension to Other Outputs: The current study focuses on latent heat flux. Extending the methodology to other output variables (e.g., soil moisture, runoff) would be a valuable next step.

Summary updated

BibTeX
@article{lin2022inverse,
  author = {Guang Lin and Georgios Karagiannis and Zhangshuan Hou and Maoyi Huang and Guang Lin},
  title = {{Inverse Modeling of Hydrologic Parameters in CLM4 via Generalized Polynomial Chaos in the Bayesian Framework}},
  journal = {comput.},
  year = {2022},
  doi = {10.3390/computation10050072},
}
Figure