Semantic Segmentation of Sorghum Using Hyperspectral Data Identifies Genetic Associations

Area of Science:

• Plant Science
• Computational Biology
• Agricultural Technology

Background:

• Current plant image segmentation often distinguishes only plant vs. background.
• Organ-specific segmentation is crucial for detailed plant trait analysis.
• Hyperspectral imaging offers rich spectral information for plant analysis.

Purpose of the Study:

• To evaluate semantic segmentation approaches for hyperspectral images of sorghum plants.
• To classify pixels into nonplant, leaf, stalk, or panicle categories.
• To assess the potential for organ-level trait extraction and gene discovery.

Main Methods:

• Supervised classification models were trained using manually annotated hyperspectral images of sorghum.
• A range of semantic segmentation algorithms were evaluated for accuracy.
• Model performance was tested on both sorghum and maize datasets.

Main Results:

• Several algorithms achieved acceptable accuracy for sorghum organ segmentation.
• Models trained on sorghum generalized to classify maize leaves and stalks but not reproductive organs.
• Extracted trait measurements revealed genetic signals for known and novel phenotypes.

Conclusions:

• Organ-level semantic segmentation of hyperspectral data is effective for sorghum.
• This approach facilitates the identification of genes controlling morphological variation in sorghum and maize.
• It offers a powerful tool for quantitative trait discovery in grain crops.