Abstract
This study investigates the application of advanced sequential Monte Carlo (SMC) methods to bridge the gap between theoretical plant growth models and stochastic agricultural data. By comparing the Unscented Kalman Filter (UKF), Ensemble Kalman Filter (EnKF), and Convolution Particle Filter (CPF), we demonstrate that data assimilation (DA) can significantly enhance the predictive accuracy of crop yield models, reducing errors by over 40%.
Introduction
Predicting crop growth is a complex task due to parameter uncertainty and the stochastic nature of environmental factors. While mathematical models like the LNAS (Log-Normal Allocation and Senescence) model provide a theoretical framework, they often diverge from real-world observations. Data Assimilation (DA) offers a solution by using sparse, irregular field measurements to "nudge" model trajectories back toward reality. This paper evaluates which mathematical filters best handle the nonlinear and non-Gaussian characteristics of biological growth.
Methodology
Sensitivity Analysis and Parameter Estimation
To streamline the assimilation process, we first performed a sensitivity analysis using Sobol’s indices. This identified the most influential parameters—such as radiation use efficiency and leaf senescence timing—allowing us to focus estimation efforts where they impact the model most. We then utilized iterative versions of the filters (IEnKF, IUKF, and ICPF) to calibrate these parameters using historical datasets from the French Institute for Sugar Beet Research (ITB).
Sequential Monte Carlo Techniques
We compared three primary filtering techniques:
- Unscented Kalman Filter (UKF): Employs a set of deterministically chosen "sigma-points" to track the system's mean and covariance.
- Ensemble Kalman Filter (EnKF): A Monte Carlo approach suitable for nonlinear systems, though it retains some Gaussian assumptions.
- Convolution Particle Filter (CPF): An advanced particle filter that utilizes a mixture of Gaussian kernels. This approach prevents "sample impoverishment," a common issue where simulations collapse into a single state, making it particularly robust for highly nonlinear systems.
Results and Discussion
The Nonlinearity Advantage
The Convolution Particle Filter (ICPF) consistently outperformed Kalman-based methods. Plant growth, particularly the allocation of biomass between leaves ($Q_g$) and roots ($Q_r$), is inherently nonlinear. Unlike the EnKF, the ICPF does not assume underlying probability distributions are normal, allowing it to better capture the skewness resulting from environmental shifts.
Impact of Data Assimilation
A critical finding of this research is the superiority of Data Assimilation (DA) over simple Uncertainty Analysis (UA). While UA quantifies model variance, DA actively incorporates new information. Our results show that assimilating even a few early-season measurements can reduce the relative error of final yield predictions by up to 42.4%.
Statistical Performance
The ICPF achieved the highest log-likelihood and the lowest Akaike Information Criterion (AIC) scores, indicating a superior fit to the experimental data. Furthermore, the standard error of predictions decreased significantly across all methods once early-growth data was integrated, confirming the stability provided by DA.
Conclusion
Sequential Monte Carlo techniques, specifically the Convolution Particle Filter, provide a robust framework for improving plant growth models. While computationally more intensive than the Ensemble Kalman Filter, the CPF's ability to handle nonlinear, non-Gaussian dynamics makes it a powerful tool for precision agriculture. Future work will explore the scalability of these methods and their sensitivity to observation noise in diverse environmental conditions.
Summary updated