A leaf-level spectral library to support high-throughput plant phenotyping: predictive accuracy and model transfer

Area of Science:

• Plant science
• Spectroscopy
• Machine learning

Background:

• Leaf-level hyperspectral reflectance is valuable for high-throughput plant phenotyping.
• Model calibration is costly, and transferability across datasets is limited.

Purpose of the Study:

• Assemble a large leaf hyperspectral library (n=2460) for maize and sorghum.
• Evaluate machine learning for estimating nine leaf properties.
• Test spectral library transferability to external datasets (n=445) using extra-weighted spiking.

Main Methods:

• Collected hyperspectral data from maize and sorghum.
• Applied partial least-squares regression and deep neural networks.
• Utilized extra-weighted spiking to integrate external data.

Main Results:

• Internal cross-validation showed good performance for estimating nine leaf traits (mean R2=0.688).
• Models showed poor transferability to external soybean and camelina datasets (mean R2=0.159-0.337).
• Extra-weighted spiking significantly improved external dataset prediction (mean R2=0.536-0.574).

Conclusions:

• The spectral library enhances plant physiological and biochemical phenotyping.
• Extra-weighted spiking is crucial for improving model transferability and utility.