Quantifying physiological trait variation with automated hyperspectral imaging in rice.

Area of Science:

• Plant Physiology
• Agricultural Science
• Remote Sensing

Background:

• High-throughput plant phenotyping facilities increasingly utilize hyperspectral imaging (HSI) to infer plant traits.
• HSI offers a non-destructive method for assessing physiological parameters in crops like rice (Oryza sativa).

Purpose of the Study:

• To evaluate the efficacy of HSI-derived canopy data for predicting physiological traits in diverse rice accessions.
• To assess the accuracy of HSI in classifying treatment groups and predicting leaf-level nitrogen (N) and carbon-to-nitrogen ratio (C:N).

Main Methods:

• Collected HSI canopy data from 23 genetically diverse rice accessions under contrasting nitrogen conditions.
• Employed Support Vector Machines for treatment group classification and Partial Least Squares Regression (PLSR) with RReliefF wavelength selection for N and C:N prediction.
• Optimized PLSR-RReliefF hyperparameters and evaluated model performance based on calibration set design.

Main Results:

• HSI successfully classified treatment groups with ≥ 83% accuracy.
• PLSR models accurately predicted leaf N (R² = 0.797) and C:N (R² = 0.592).
• Models trained on one rice subpopulation predicted traits in another, but not across different nitrogen treatments. Optimal performance was achieved with 300-400 wavelengths and a minimum of 62 calibration samples.

Conclusions:

• Canopy-level HSI data are valuable for estimating leaf-level N and C:N in diverse rice varieties.
• Calibration set design and hyperparameter optimization are critical for developing robust HSI-based plant phenotyping models.