Filtrage par noyaux de convolution itératif

Yuting Chen, Benoît Bayol, Cédric Loi, Samis Trevezas, Paul-Henry CournèdeSPIRE 2012

Abstract: International audience

Introduction

Modeling biological systems, such as the seasonal growth of a sugar beet, presents significant challenges due to their highly non-linear nature and the presence of "hidden states"—variables like root biomass that are difficult or destructive to measure frequently. Researchers must often perform joint estimation, simultaneously determining the internal state of the plant and underlying biological parameters, such as resource use efficiency. Traditional methods, including the standard Convolution Particle Filter (CPF), frequently falter when data is sparse, initial parameter guesses (priors) are weak, or the distinction between biological variability and measurement error is unclear.

Methodology: The Iterative Convolution Particle Filter (ICPF)

To address these limitations, this paper introduces the Iterative Convolution Particle Filter (ICPF). Building upon the CPF, which utilizes Parzen-Rosenblatt kernels for density estimation, the ICPF incorporates two primary innovations:

  1. The Iterative Approach: Rather than a single pass through the time-series data, the algorithm cycles through the data multiple times. The particle distribution at the end of one pass initializes the next, allowing the filter to converge on true parameters even from uninformative starting points.
  2. Conditional Estimation: Parameters are bifurcated into functional parameters ($\Theta_1$) and noise parameters ($\Theta_2$). The algorithm first estimates the plant state and $\Theta_1$ while holding noise constant, then calculates the actual noise levels ($\Theta_2$) based on those estimates before restarting the cycle. This "ping-pong" stabilization continues until convergence is reached.
Convergence of the estimated parameter μ_a over successive iterations of the filter.

Experimental Framework

The study utilizes a stochastic sugar beet growth model based on the Beer-Lambert law. To validate the ICPF, the authors generated synthetic data simulating a 140-day growth cycle with observations every 5 days. Significant observation noise was injected to test the filter's robustness in scenarios mimicking actual agricultural field trials where frequent destructive measurements are impossible.

Key Results and Insights

The ICPF demonstrates that iteration is often more valuable than raw particle count. By "recycling" information, the filter achieves higher accuracy with fewer resources. Specifically, an Iterative Filter using only 1,000 particles outperformed a Standard Filter using 100,000 particles, providing tighter confidence intervals and more accurate state tracking.

Evolution of biomass over time, comparing the estimated hidden state, real state, and observations.

Furthermore, the conditional version of the algorithm successfully distinguished between biological signals and measurement error, accurately recovering noise standard deviations (e.g., estimating a true noise of 0.05 as 0.0537). The use of a parametric bootstrap also allows for the generation of reliable confidence intervals, turning previously "black box" dynamic models into quantifiable biological tools.

Limitations and Future Directions

While highly effective, the ICPF's primary drawback is the increased computational cost associated with multiple passes through the data. Future work will focus on optimizing the balance between particle count and iteration frequency, developing formal convergence theories, and applying the ICPF to real-world field data across diverse crop species to verify its robustness against environmental unpredictability.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{chen2012filtrage,
  author = {Yuting Chen and Benoît Bayol and Cédric Loi and Samis Trevezas and Paul-Henry Cournède},
  title = {{Filtrage par noyaux de convolution itératif}},
  booktitle = {SPIRE},
  year = {2012},
}
Figure