Classification of plant growth‐promoting bacteria inoculation status and prediction of growth‐related traits in tropical maize using hyperspectral image and genomic data

Abstract

Recent technological advances in high-throughput phenotyping have created new opportunities for the prediction of complex traits. In particular, phenomic prediction using hyperspectral reflectance could capture various signals that affect phenotypes genomic prediction might not explain. A total of 360 inbred maize (Zea mays L.) lines with or without plant growth-promoting bacterial inoculation management under nitrogen stress were evaluated using 150 spectral wavelengths ranging from 386 to 1,021 nm and 13,826 single-nucleotide polymorphisms. Six prediction models were explored to assess the predictive ability of hyperspectral and genomic data for inoculation status and plant growth-related traits. The best models for hyperspectral prediction were partial least squares and automated machine learning. The Bayesian ridge regression and BayesB were the best performers for genomic prediction. Overall, hyperspectral prediction showed greater predictive ability for shoot dry mass and stalk diameter, whereas genomic prediction was better for plant height. The prediction models that simultaneously accommodated both hyperspectral and genomic data resulted in a predictive ability as high as that of phenomics or genomics alone. Our results highlight the usefulness of hyperspectral-based phenotyping for management and phenomic prediction studies.

Genome‐wide association analysis of hyperspectral reflectance data to dissect the genetic architecture of growth‐related traits in maize under plant growth‐promoting bacteria inoculation

“…A total of 86 SNPs were selected from the BayesC analysis using the posterior inclusion probability threshold of 0.10 for PH, SD, and SDM (Figure 3 and supporting information Tables S2–S6 ). This number was higher than a previous study (Yassue et al, 2022 ) that used a non‐Bayesian whole‐genome regression model likely because BayesC has a higher statistical power to detect genetic signals. There were eleven common SNPs between the current study and Yassue et al ( 2022 ) (Table S7 ).…”

“…This number was higher than a previous study (Yassue et al, 2022 ) that used a non‐Bayesian whole‐genome regression model likely because BayesC has a higher statistical power to detect genetic signals. There were eleven common SNPs between the current study and Yassue et al ( 2022 ) (Table S7 ). PH showed the highest number of selected markers (21 and 24 for B+ and B−, respectively), whereas SDM had the lowest (5 for B+); no SDM markers were detected in B−.…”

“…Image processing through the Spectral Python (Boggs, 2014 ) module was performed by applying a mask to remove the background from the image, and the mean reflectance of each pixel was used for further analysis. Hyperspectral imaging and processing details are described in Yassue et al ( 2022 ). In addition, 131 hyperspectral indices (mathematical band combinations) were calculated based on the mean reflectance value for each wavelength using the R package hsdar (Lehnert et al, 2019 ).…”

“…A recent study showed that hyperspectral image data can be used to perform phenomic prediction for PH, SD, SDM with reasonable accuracy (Yassue et al, 2022 ). In the current study, we investigated if these hyperspectral reflectance values are under genetic control and whether they are associated with PH, SD, and SDM.…”

From pixel to knowledge: how high-throughput phenotyping helps to dissect the genetic architecture and improves predictive ability in maize under inoculation with plant growth-promoting bacteria

“…A summary of hyperspectral imaging and processing can be found in Yassue et al (2022a). In addition, a total 131 hyperspectral indices were calculated based on the mean reflectance value for each wavelength using the R package hsdar (Lehnert et al, 2019).…”