Iterative convolution particle filtering for nonlinear parameter estimation and data assimilation with application to crop yield prediction

Yuting Chen, Samis Trevezas, Paul-Henry CournèdeSIAM Conf. on Control and its Applications 2013 · DOI

Abstract

The complexity of plant growth models and the scarcity of experimental data make the application of conventional data assimilation techniques rather difficult. In this paper, we use the Convolution Particle Filter (CPF) and an iterative adaptation, the Iterative Convolution Particle Filter (ICPF) for nonlinear parameter estimation. Both methods provide prior distributions in the Bayesian framework for data assimilation. CPF is sequentially used to update state and parameter estimates in order to improve model prediction and to assess the predictive uncertainty. The predictive performances of the two methods are evaluated by an application to the LNAS sugar beet growth model with three sets of real measurements, one used for parameter estimation and the two others used to test the model predictive capacity, both with and without data assimilation. Despite the low accuracy and the scarcity of the early data used for assimilation, the CPF-based data assimilation approach with the prior distribution based on ICPF estimations showed promising predictive capacities and provided robust confidence intervals. The method can therefore be considered as a potential candidate for yield prediction applications in agriculture.

Introduction

Predicting crop yield is a critical yet challenging task due to the high nonlinearity of plant growth models and the complex interactions between numerous parameters. Effective yield forecasting requires data assimilation—the continuous integration of field observations (e.g., biomass measurements) to update a model's state and parameters. However, traditional methods like the Extended Kalman Filter (EKF) often fail in highly nonlinear systems, while standard Particle Filters (PF) frequently suffer from "sample degeneracy," where the filter collapses as most particles lose weight. Furthermore, agricultural data is often sparse, irregular, and noisy, complicating the calibration process.

Methodology and Data

The authors propose a robust framework centered on the Convolution Particle Filter (CPF) and its iterative variant, ICPF, to address these challenges.

Sensitivity Analysis

To manage model complexity, the researchers first performed a sensitivity analysis using Sobol’s indices. This "screening" process identifies the most influential parameters driving model output. By fixing less significant parameters to their mean values, the dimensionality of the estimation problem is reduced, making the subsequent filtering more efficient.

Sobol's Indices for LNAS model parameters showing the most influential functional parameters.

Iterative Convolution Particle Filter (ICPF)

For initial off-line model calibration, the ICPF is employed. Unlike standard particle filters that rely on resampling, CPF uses a Parzen-Rosenblatt kernel to estimate probability density, circumventing sample impoverishment. The iterative version (ICPF) repeats this estimation, using the posterior distribution of one iteration as the prior for the next until convergence is achieved.

Sequential Data Assimilation

Once calibrated, the standard CPF is used for real-time data assimilation. This allows the model to "correct" its trajectory as early-season data becomes available, leading to more accurate end-of-season forecasts. The study utilized real-world sugar beet data from the French Institute for Sugar Beet Research (ITB), including a 2010 calibration set and 2006/2008 test sets.

Results and Discussion

The CPF-based approach demonstrated significant improvements over traditional Uncertainty Analysis (UA), which runs the model without data-driven updates.

Accuracy and Reliability

For the 2006 dataset, the relative error for green leaf mass ($Q_g$) prediction was slashed from 42.96% using UA to just 2.01% with ICPF-based assimilation. Beyond point accuracy, the method provided robust 95% confidence intervals. While UA often produced intervals that failed to capture actual measurements, the CPF approach generated narrower, more reliable intervals that consistently contained the real-world data points.

Comparison of ICPF and CPF predictions against Uncertainty Analysis (UA) for 2006 biomass evolution, demonstrating improved accuracy and tighter confidence intervals.

Key Insights

The success of the method hinges on two factors: the use of kernel smoothing instead of resampling to maintain particle diversity, and the generation of a high-quality "prior" distribution via ICPF. The results suggest that even with limited early-season measurements, the model can provide the predictive capacity necessary for agricultural decision-support tools.

Conclusion

While promising, the method's performance is sensitive to the characterization of observation noise. Additionally, the computational demand of processing hundreds of thousands of particles remains a consideration for high-dimensional applications. Future work may focus on generalizing noise models across different environmental conditions. Nonetheless, the ICPF/CPF framework represents a significant step forward in robust crop yield prediction under conditions of data scarcity and model nonlinearity.

Summary updated

BibTeX
@inproceedings{chen2013iterative,
  author = {Yuting Chen and Samis Trevezas and Paul-Henry Cournède},
  title = {{Iterative convolution particle filtering for nonlinear parameter estimation and data assimilation with application to crop yield prediction}},
  booktitle = {SIAM Conf. on Control and its Applications},
  year = {2013},
  doi = {10.1137/1.9781611973273.10},
}
Figure