TL;DR
This paper addresses the challenge of accurately modeling sugarcane growth by calibrating the ORCHIDEE-STICS agro-land surface model. By identifying and optimizing the most influential parameters governing leaf area index (LAI) across multiple sites, the study demonstrates that the "figure of merit" scoring function provides the most robust calibration, improving LAI simulation accuracy from 67% to 79% agreement.
What Problem Does This Paper Address?
Accurately simulating crop growth, particularly for major bioenergy crops like sugarcane, is crucial for understanding their impact on the biosphere-atmosphere exchange of carbon, water, and energy, and for predicting regional yields. Agro-land surface models (agro-LSMs) aim to bridge the gap between detailed crop models and large-scale vegetation models, but they often require extensive parameterization. A key challenge is identifying which parameters are most critical for model performance and how to calibrate them effectively across diverse environmental conditions. This paper focuses on the ORCHIDEE-STICS model for sugarcane, specifically addressing the need to accurately simulate leaf area index (LAI), a fundamental driver of biomass production and carbon assimilation, by calibrating its phenology-governing parameters.
Methods
The study employed a two-step approach:
- Sensitivity Analysis (SA): The Morris method was used to identify the most influential parameters affecting sugarcane's maximum Leaf Area Index (LAI) across eight different field trial sites. This method is a computationally efficient "one-at-a-time" approach that helps rank parameters based on their impact and the presence of nonlinearities.
- Multisite Calibration: After identifying the top three most sensitive parameters, a factorial calibration was performed. This involved systematically varying these parameters within their defined ranges and evaluating the model's performance using three different scoring functions:
- Root Mean Squared Error (RMSE): A standard measure of the average magnitude of the errors.
- Figure of Merit (FM): A metric based on the ratio of overlapping observed and simulated LAI curves.
- Bayesian Quadratic Misfit Function (J): A function that includes prior information about parameter values.
The robustness of the calibration was then assessed using a leave-one-out cross-validation technique, where each site was sequentially removed from the calibration set to evaluate how well the optimized parameters performed on unseen data.
Datasets
The study utilized data from eight sugarcane field trial sites:
- Australia: Ayr (3 years), Ingham (2 years), Grafton (1 year)
- La Réunion: Colimaçons (1 year), Tirano (1 year)
These sites provided measurements of crop biomass and LAI at multiple points during the growing season. Meteorological forcing data (precipitation, air temperature, solar radiation) were obtained from the Australian Bureau of Meteorology and local stations in La Réunion, and supplemented with ERA-Interim reanalysis data.
Key Insights and Ideas
- The "Figure of Merit" Reigns Supreme: The study strongly suggests that the "Figure of Merit" (FM) scoring function is the most robust and reliable metric for calibrating agro-LSMs like ORCHIDEE-STICS for sugarcane, especially when aiming for multisite applicability.
- LAI is Driven by a Few Key Players: Despite the complexity of crop models, the sensitivity analysis revealed that only a handful of parameters significantly influence LAI. This is crucial for efficient model calibration and reduces the computational burden.
- Plant Density and Nitrogen Stress are Critical: The three most influential parameters identified were:
- The maximum predefined rate of LAI increase (related to plant density and competition).
- A parameter defining a plant density threshold below which individual plants don't compete for LAI.
-
A parameter defining a threshold for nitrogen stress on LAI. This highlights the interconnectedness of physical growth processes and nutrient availability in determining sugarcane's vegetative development.
-
Multisite Calibration is Essential for Robustness: Calibrating a model on a single site can lead to overfitting and poor generalization. The cross-validation approach demonstrated that multisite calibration, particularly with the FM function, leads to more stable and reliable parameter sets.
Key Results
- The sensitivity analysis identified dmax (maximum rate of LAI increase) as the most influential parameter for LAI simulation across all sites, followed by βdens (plant density threshold) and INNmin (minimum nitrogen nutrition index).
- The multisite calibration using the FM scoring function improved the average agreement of LAI simulations from 67% to 79%.
- While all three scoring functions (RMSE, FM, J) improved LAI simulations, the FM function demonstrated the most consistent performance across different sites during cross-validation, indicating greater robustness.
- The RMSE and FM scoring functions showed similar cross-validation behavior, but the FM function was deemed more robust due to its less pronounced sensitivity to the exclusion of specific sites during calibration.
Limitations and Future Work
- Residual Error: Despite calibration, a residual error in LAI simulation persists, particularly during the late development cycle. This is attributed to potential limitations in parameter values, forcing data accuracy, or model structure (e.g., factors not explicitly modeled).
- Generic Soil Parameterization: The use of a generic soil physics parameterization across all sites ignores local soil specificities, which can be crucial for accurately describing soil water content, water availability, and root stress.
- Variety and Ratooning: The study acknowledges that assuming a single variety and not accounting for ratooning (where sugarcane is cut and regrows from the stubble) can introduce discrepancies, especially when comparing real-world agricultural practices to model assumptions.
- Future Work: The authors suggest that a more detailed description of soil conditions and the inclusion of other crop development drivers (like biomass and net primary productivity) would be necessary for a more comprehensive model simulation. Further research could also explore the impact of specific sugarcane varieties and ratooning practices on model performance.
Summary updated