Biomass prediction based on hyperspectral images of the Arabidopsis canopy

Authors: SONG, DI - University Of Illinois; DE SILVA, KITHMEE - University Of Illinois; Brooks, Matthew; KAMRUZZAMAN, MOHAMMED - University Of Illinois

Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 5/16/2023
Publication Date: 7/1/2023
Citation: Song, D., De Silva, K., Brooks, M.D., Kamruzzaman, M. 2023. Biomass prediction based on hyperspectral images of the Arabidopsis canopy. Computers and Electronics in Agriculture. 210: Article 107939. https://doi.org/10.1016/j.compag.2023.107939.
DOI: https://doi.org/10.1016/j.compag.2023.107939

Interpretive Summary: Accurately tracking growth and biomass is important to determine the health of plant and potential yield, however most methods are time-intensive and typically involve destroying the tissue. Hyperspectral imaging, which measures light coming off of a leaf in the visible and infrared wavelengths, is a promising method for non-invasively measuring plant traits such as biomass but there is no standard methods yet for processing and analyzing images. This work uses hyperspectral imaging of plants grown under a matrix of nitrogen and light conditions to test which imaging analysis approaches are best able to predict measured root and shoot biomass. The best method for isolating the plant material from the background was found to be the Normalized Difference Vegetation Index (NDVI). For isolating the most informative wavelengths, the BOotstrapping Soft Shrinkage (BOSS) method performed the best. This approach suggest that hyperspectral imaging combined with proper image analysis methods can be effective for accurately predicting shoot and root biomass and could help track these traits over time in a non-destructive manner.

Technical Abstract: Hyperspectral images provide detailed crop canopy spectral information to evaluate crop biomass. However, a large amount of information exists in hyperspectral images, which can make it difficult to accurately predict crop biomass. Therefore, this study aimed to eliminate irrelevant information from hyperspectral data to accurately predict shoot and root biomass of Arabidopsis. First, hyperspectral images of Arabidopsis were acquired in the spectral range of 400-1000 nm, and the background was removed using different techniques. Comparing the results of image processing based on image and spectral information methods, the average reflectance spectrum obtained by the spectral information based normalized difference vegetation Index (NDVI) segmentation resulted better shoot and root biomass prediction than the excess green index (ExG), CIELAB (Lab), and soil-adjusted vegetation index (SAVI) methods. Three wavelength optimization methods such as competitive adaptive reweighted sampling (CARS), bootstrapping soft shrinkage (BOSS), and non-dominated sorting genetic algorithm-II (NSGA) were used to extract useful information from hyperspectral images. BOSS selected the least number of wavelengths and the partial least squares regression (PLSR) models developed using BOSS produced better results for both shoot and root biomass. Only 19 informative wavelengths were selected and PLSR models accurately predicted shoot biomass with RC2 of 0.91, RV2 of 0.85, RMSEC, RMSEV of 0.013 g and 0.017 g, respectively, and root biomass with RC2 of 0.94, RV2 of 0.88, RMSEC and RMSEV of 0.03 g and 0.04 g, respectively. The results indicated that hyperspectral imaging is very effective for accurately predicting shoot and root biomass of Arabidopsis.