On machine learning methods to estimate cannabidiolic acid content of Cannabis sativa L. from near-infrared hyperspectral imaging

This paper investigates the use of near-infrared hyperspectral imaging to estimate the cannabidiolic acid (CBDA) concentration of Cannabis sativa L. plant components such as buds and leaves. A total of 9 regression techniques were assessed with Gaussian Process Regression producing the best result. The least angle regression technique was the best performer of the interpretable models trialled. These models relate the measured pixel reflectance to cannabinoid content as determined by liquid chromatography-mass spectrometry. The regression models were applied to the complete hyperspectral image of the plant to produce CBDA estimation and distribution maps. The ultimate goal of this work is to develop a non-destructive method for Cannabis sativa L. assessment to improve both crop yield and the quality of the plant.