Hyperspectral Leaf Reflectance as Proxy for Photosynthetic Capacities: An Ensemble Approach Based on Multiple Machine...

Area of Science:

• Plant Physiology
• Agricultural Science
• Machine Learning in Biology

Background:

• Global agriculture faces production challenges due to climate change and population growth, necessitating genetically improved crop cultivars.
• Improving photosynthetic efficiency is key, but a phenotyping bottleneck hinders progress.
• Partial Least Squares Regression (PLSR) is used to link hyperspectral reflectance to photosynthetic capacities (Vcmax and Jmax), but its performance is inconsistent.

Purpose of the Study:

• To develop a novel, robust framework for estimating photosynthetic capacities (maximum carboxylation rate of Rubisco, Vcmax, and maximum electron transport rate, Jmax) using machine learning.
• To overcome the limitations of individual regression techniques and improve high-throughput phenotyping for crop improvement.

Main Methods:

• A framework combining six machine learning algorithms (ANN, SVM, LASSO, RF, GP, PLSR) was developed.
• Hyperspectral leaf reflectance (400-2500 nm) and gas-exchange data were collected from six tobacco genotypes.
• Regression stacking was employed to combine predictions from individual algorithms for enhanced accuracy.

Main Results:

• Individual regression techniques showed moderate predictive power for Vcmax (R²: 0.60-0.65) and Jmax (R²: 0.45-0.56).
• Regression stacking significantly improved predictions, increasing R² by 0.1 (0.08) and reducing RMSE by 4.1 (6.6) μmol m⁻² s⁻¹ for Vcmax (Jmax), representing an 8% (15%) RMSE reduction.
• The stacked model's superior performance is attributed to the combination of diverse algorithms and optimized weighting.

Conclusions:

• The developed stacked regression framework offers a more accurate and reliable method for estimating photosynthetic capacities from hyperspectral data.
• This approach effectively addresses the phenotyping bottleneck, paving the way for accelerated development of genetically improved crop cultivars.
• The stacked regression technique shows potential for application to other plant phenotypic traits, advancing high-throughput phenotyping in agriculture.