Assessment of parameter uncertainty in plant growth model identification

Yuting Chen, Paul-Henry Cournède2012 · DOI

TL;DR

This paper tackles the challenge of accurately estimating parameters and their uncertainties in complex, nonlinear plant growth models, especially when data is limited or irregular. They propose an improved Bayesian particle filtering method (ICPF) that offers more reliable uncertainty estimates than traditional GLS methods, demonstrating its effectiveness on sugar beet growth data.

What Problem Does This Paper Address?

Accurately modeling plant growth is crucial for predicting crop yields and understanding plant physiology. However, a major hurdle is "parameter identification" – figuring out the correct values for the many parameters that define these models. This is difficult because:

  • Complex Dynamics: Plant growth models are often highly nonlinear, meaning small changes in parameters can lead to large, unpredictable changes in output.
  • Limited Data: Experimental data for plant growth is frequently sparse, collected at irregular intervals, or subject to noise.
  • Uncertainty Quantification: Simply getting a single "best guess" for a parameter isn't enough. We need to know how confident we are in that guess (i.e., its uncertainty). Traditional methods often rely on linear approximations for this, which break down in the face of strong nonlinearities, leading to unreliable uncertainty estimates.

The authors aim to compare different methods for parameter estimation and, importantly, for quantifying the uncertainty associated with those estimates, particularly in the challenging context of limited and noisy data.

Methods

The paper compares three main approaches:

  • Generalized Least Squares (GLS): A standard, computationally efficient method for parameter estimation. It provides point estimates and an approximation of parameter uncertainty, but its reliance on linear approximations makes it less reliable for highly nonlinear models. The paper uses the "2-stage Aitken estimator" for GLS.
  • Convolution Particle Filtering (CPF): A Bayesian approach that treats parameter estimation as a probabilistic problem. Instead of a single point estimate, CPF provides a posterior distribution for each parameter, allowing for more robust uncertainty quantification (e.g., credible intervals). It estimates both parameters and hidden states of the model simultaneously.
  • Iterative Convolution Particle Filtering (ICPF): This is the paper's novel contribution. It's an iterative version of CPF designed to improve performance, especially with stochastic models and sparse/irregular data. ICPF refines parameter estimates by iteratively updating them, considering both parameter and noise model uncertainties. It also incorporates a bootstrap method for more reliable confidence interval estimation.

Datasets

The study uses a dataset of sugar beet growth measurements collected in 2008. This dataset includes measurements of green foliage mass ($Q_g$), senescent foliage mass ($Q_s$), and root mass ($Q_r$) for 20 individual plants over several time points. The data was averaged and scaled to represent a per-square-meter level.

Key Insights and Ideas

  • Nonlinearity is a Killer for Simple Uncertainty Estimates: The paper strongly suggests that traditional GLS methods, while fast, provide unreliable uncertainty estimates for complex, nonlinear models like the LNAS plant growth model.
  • Bayesian Methods Offer Better Uncertainty: CPF and ICPF, being Bayesian, provide a more principled way to capture parameter uncertainty through posterior distributions.
  • ICPF Outperforms CPF in Robustness: While CPF is good, the iterative nature of ICPF, especially when dealing with noise and limited data, leads to more robust and reliable parameter and uncertainty estimates.
  • Noise Matters: The ICPF's ability to explicitly model and estimate noise parameters (both process and measurement noise) is a key factor in its improved performance.
  • Computational Cost vs. Reliability Trade-off: ICPF is computationally more expensive than GLS and CPF, but this cost is justified by its superior reliability in uncertainty quantification.

Key Results

  • GLS Limitations: The GLS method showed good fitting for some biomass components ($Q_g$) but struggled significantly with the senescent foliage mass ($Q_s$), indicating unreliable parameter estimates for this aspect. Its uncertainty estimates were also deemed less reliable.
  • CPF and ICPF Performance: Both CPF and ICPF provided better overall fits and more reasonable parameter estimates than GLS.
  • ICPF's Superior Uncertainty: ICPF consistently provided the most reliable uncertainty estimates, often encompassing the estimates from other methods. While CPF offered narrower uncertainty intervals, ICPF's intervals were found to be more robust and trustworthy, especially in robustness tests.
  • Robustness Test Findings: When tested with virtual data generated from known parameters, ICPF demonstrated significantly better robustness and more accurate confidence intervals compared to CPF.
  • Computational Cost: ICPF was the most computationally intensive method, taking around 18 hours for a single test run with specific settings, whereas GLS and CPF were much more efficient.

Limitations and Future Work

  • Computational Cost of ICPF: The authors acknowledge that ICPF's high computational demand is a significant drawback.
  • Multiplicative Noise Assumption: The paper notes that the multiplicative assumption for noise in the LNAS model might be a simplification and could potentially twist estimations when biomass increases significantly.
  • Future Work: The authors suggest exploring combinations of methods, such as using GLS initially to narrow down the parameter search space for particle filters. They also propose further steps including uncertainty analysis and testing the predictive capacity of the models. A richer dataset could also be beneficial for re-evaluating conclusions.

Summary updated

Abstract

For the parametric identification of plant growth models, we generally face limited or uneven experimental data, and complex nonlinear dynamics. Both aspects make model parametrization and uncertainty analysis a difficult task. The Generalized Least Squares (GLS) estimator is often used since it can provide estimations rather rapidly with an appropriate goodness-of-fit. However, the confidence intervals are generally calculated based on linear approximations which make the uncertainty evaluation unreliable in the case of strong nonlinearity. A Bayesian approach, the Convolution Particle Filtering (CPF), can thus be applied to estimate the unknown parameters along with the hidden states. In this case, the posterior distribution obtained can be used to evaluate the uncertainty of the estimates. In order to improve its performance especially with stochastic models and in the case of rare or irregular experimental data, a conditional iterative version of the Convolution Particle Filtering (ICPF) is proposed. When applied to the Log Normal Allocation and Senescence model (LNAS) with sugar beet data, the two CPF related approaches showed better performance compared to the GLS method. The ICPF approach provided the most reliable estimations. Meanwhile, two sources of the estimation uncertainty were identified: the variance generated by the stochastic nature of the algorithm (relatively small for the ICPF approach) and the residual variance partly due to the noise models.
BibTeX
@misc{chen2012assessment,
  author = {Yuting Chen and Paul-Henry Cournède},
  title = {{Assessment of parameter uncertainty in plant growth model identification}},
  year = {2012},
  doi = {10.1109/pma.2012.6524817},
}
Figure