8th International Conference on Duckweed Research and Applications

High-Throughput Non-Destructive Prediction of Duckweed Protein Content Using Hyperspectral Imaging and Deep Learning

01 ott 2026, 10:00

15m

Sala Cinese (Department of Agricultural Sciences of the University of Napoli Federico II, Portici, Italy)

Oral Applications I a — food, feed & nutrition

Description

Duckweed is a promising plant-based protein and biomass resource, but efficient breeding and large-scale cultivation require rapid, high-throughput, and non-destructive evaluation of key traits. Conventional methods for determining crude protein and chlorophyll contents are labor-intensive, chemically demanding, and destructive, limiting their application in large-scale germplasm screening and real-time cultivation management. In this study, we developed an integrated high-throughput phenotyping platform combining hyperspectral imaging (HSI) and deep learning for simultaneous non-destructive prediction of crude protein content, leaf area-based relative growth rate (RGR), and chlorophyll content in duckweed. More than 200 duckweed samples cultivated in standardized 12-well microplates were scanned using an HSI system covering 400–2000 nm with a spectral resolution of 3 nm. A DeepLab v3+ semantic segmentation model was used to delineate duckweed fronds and quantify leaf area, enabling automated extraction of growth-related phenotypes. A Transformer-based spectral-spatial feature fusion model was further developed to integrate RGB-derived spatial features with hyperspectral signatures for biochemical and growth trait prediction. The platform achieved strong predictive performance for crude protein content (R² = 0.938, RMSE = 1.030, prediction accuracy = 95.8%), chlorophyll content (R² = 0.893, RMSE = 1.514, prediction accuracy = 94.7%), and leaf area-based RGR (prediction accuracy = 98.5%). With a throughput of up to 120 samples per hour, this HSI- and deep learning-based platform provides an efficient tool for duckweed germplasm evaluation, high-protein strain selection, and precision cultivation management.

Keywords

Protein; Hyperspectral imaging; Deep learning

References

[1] Gao, R., Li, Z., Ma, Z., et al. (2021). "Research on detection of crude protein content in pasture based on hyperspectral imaging." Spectroscopy and Spectral Analysis 41(10): 3158-3163.
[2] Aulia, R., Kim, Y., Amanah, H.Z., et al. (2022). "Non-destructive prediction of protein contents of soybean seeds using near-infrared hyperspectral imaging." Infrared Physics & Technology 127: 104365.
[3] Xuan, G., Jia, H., Shao, Y., & Shi, C. (2024). "Protein content prediction of rice grains based on hyperspectral imaging." Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 320: 124589.
[4] Wang, P., et al. (2022). "An improved DeepLab v3+ deep learning network applied to the segmentation of grape leaf black rot spots." Frontiers in Plant Science 13: 795410.