Introduction
Functional-Structural Plant Models (FSPMs), such as the GreenLab model, are essential tools for simulating the complex interactions between plant genetics and environmental factors. These models track the allocation of biomass (carbon) to various organs—leaves, stems, and roots—over time. However, accurately estimating the parameters of these models is a significant challenge. Traditional methods often struggle with the non-linear nature of plant growth, the necessity of destructive sampling (which prevents longitudinal tracking of individuals), and the presence of "process noise" caused by environmental fluctuations.
Modeling Framework: From Deterministic to Stochastic
Most previous research treated plant growth as a deterministic process where only measurements were subject to noise. This study reframes the GreenLab model as a non-homogeneous hidden Markov model (HMM). By incorporating multiplicative model error, the framework accounts for environmental variability that affects growth across successive cycles. This approach allows for a more realistic representation of biological systems, where early-stage fluctuations can propagate through the plant's development, creating complex temporal correlations that simpler models ignore.
The SISRECM Algorithm
To perform Maximum Likelihood Estimation (MLE) in this complex HMM, the authors developed the SISRECM algorithm, which combines Sequential Importance Sampling with Resampling (SISR) and Expectation Conditional Maximization (ECM).
The E-Step: Particle Filtering
The Expectation step utilizes an improved particle filter (SISR). Unlike standard bootstrap filters, this "Improved Filter" incorporates current observation data to guide particles more effectively. This is crucial for approximating the hidden state—the true biomass—of the plant at each growth stage, especially when dealing with non-Gaussian noise and non-linear growth equations. The filter approximates the posterior distribution of the hidden states, allowing the algorithm to handle the stochastic nature of the growth process.
The M-Step: Conditional Maximization
The Maximization step employs the ECM algorithm to handle the high-dimensional parameter space. By partitioning parameters into subsets, the algorithm updates some via explicit analytical formulas (such as variance and leaf resistance) and others through numerical optimization using the BFGS algorithm. This "divide and conquer" strategy enhances computational stability and speed. Furthermore, a weighted averaging technique is applied to the estimates after a "burn-in" period to mitigate the variance inherent in Monte Carlo simulations.
Experimental Validation
The SISRECM approach was validated using both synthetic and real-world datasets.
- Synthetic Data: In simulations over 50 growth cycles, the algorithm successfully recovered the "true" parameters with high precision and low standard deviation across independent runs, confirming its theoretical robustness.
- Sugar-Beet (Beta vulgaris L.): Using experimental data from 2006 involving measurements of leaf blades and petioles, the HMM formulation was compared against standard deterministic models. Model selection criteria, including the Corrected Akaike Information Criterion (AICc) and the Bayesian Information Criterion (BIC), consistently favored the HMM approach.
A key finding from the sugar-beet data was the discovery of strong temporal correlations in organ development. While leaf and petiole masses showed little correlation at a single point in time, their growth trajectories over successive cycles were highly correlated ($r \approx 0.63$ to $0.66$), a phenomenon that deterministic models fail to capture.
Discussion and Future Directions
While the SISRECM algorithm provides a powerful framework for plant modeling, it is computationally intensive, requiring up to 10,000 trajectories for convergence in complex scenarios. Future research aims to address the challenges of small sample sizes common in destructive sampling and to extend the framework to plants with stochastic organogenesis, where the very number of organs produced is a random variable—a common feature in trees and more complex plant structures.
Summary updated