Generalizability and transferability of machine learning models using hyperspectral reflectance data for maize...

Area of Science:

• Plant science
• Spectroscopy
• Machine learning

Background:

• Hyperspectral reflectance enables rapid, non-destructive plant phenotyping.
• Machine learning models are used to predict plant traits from spectral data.
• Challenges remain in model generalizability and transferability.

Purpose of the Study:

• To benchmark machine learning models for hyperspectral trait prediction.
• To compare Partial Least Squares Regression (PLSR) and Support Vector Regression (SVR) performance.
• To assess the impact of data aggregation and model generalizability.

Main Methods:

• Collected hyperspectral reflectance and 25 plant traits (anatomical, gas exchange, fluorescence) from 320 recombinant inbred lines over three seasons.
• Employed a nested cross-validation framework to evaluate model performance and generalizability.
• Investigated the influence of different data aggregation strategies on predictive accuracy.

Main Results:

• Single cross-validation with Mean Squared Error (MSE) metric performed comparably to more complex validation methods.
• Optimal trait prediction accuracy depended on the specific combination of machine learning model and data aggregation level.
• Structural and biochemical traits demonstrated superior generalizability and transferability compared to physiological traits.

Conclusions:

• This study provides a rigorous benchmark for machine learning models in hyperspectral trait prediction.
• Model performance and generalizability are influenced by trait type, model choice, and data aggregation.
• Limitations exist for achieving robust generalization of physiological trait predictions across diverse conditions.