A Leaf-Level Spectral Library to Support High-Throughput Plant Phenotyping: Predictive Accuracy and Model Transfer
Authors:
Nuwan K. Wijewardane1, Huichun Zhang2, Jinliang Yang1, James C. Schnable1, Daniel P. Schachtman1, and Yufeng Ge1
Institutions:
1Mississippi State University; and 2Nanjing Forestry University, China
Abstract
Leaf-level hyperspectral reflectance data has become an effective tool for high-throughput phenotyping of plant leaf traits due to its rapid, low-cost, multisensing, and nondestructive nature. However, model calibration is often expensive regarding the number of samples, time, and labor; and models show poor transferability among different datasets. Building large spectral datasets across multiple species enables accurate model calibration and improves model transferability. The team pursued three specific objectives: (1) assemble a large library of leaf hyperspectral data (n=2,460) constructed from maize and sorghum, (2) evaluate two machine-learning approaches to estimate nine leaf properties, and (3) investigate the utility of this spectral library for predicting external datasets (n=445) including maize, sorghum, soybean, and camelina using extra-weighted spiking. Partial least squares regression models exhibited higher predictive performance than deep neural network models.
Models calibrated solely using the spectral library showed poor performance when applied to external datasets (R2<0.3 for N, P, and Ca with camelina samples). However, model transferability improved significantly when extra-weighted spiking with a small dataset (n=20) was employed (R2>0.6 for N, P, and Ca with camelina samples), indicating that it can improve the effectiveness and utility of spectral libraries in high- throughput phenotyping contexts.