Hyperspectral image-based measurement of total flavonoid content of leaf-use Ginkgo biloba L.

Abstract Ginkgo biloba leaf is a kind of food raw material with high concentration of total flavonoid content (TFC). Since conventional measurements of TFC are time-consuming and destructive, the approach of hyperspectral imaging-based prediction was investigated in this study. Hyperspectral images were collected in visual near-infrared (VIS-NIR) and short-wave near-infrared (SW-NIR). Initial comparative analysis on the full spans of the two wavelengths using partial least squares regression (PLSR) and support vector regression (SVR) showed that the TFC of the leaves could be barely captured only when PLSR working on short-wave infrared, with R p 2 and RMSE p on testing set being respectively 0.5496 and 0.6384 mg/g. To further improve prediction performance meeting industrial requirements, another comparative study on feature selection was conducted to fine-tune the PLSR on SW-NIR using a genetic algorithm (GA) and successive projections algorithm (SPA). Results showed that GA-PLSR with 50 characteristic wavelengths mostly from the ranges of 1100-1200 nm and 1400-1500 nm had a significant improvement in performance, giving 0.8482 and 0.2967 mg/g for R p 2 and RMSE p, respectively. Therefore, an approach, GA-PLSR modeling on SW-NIR hyperspectral images, was established for the ginkgo leaves industry to rapidly and non-invasively predict the leaf concentration from the total flavonoid of ginkgo biloba seedling leaf.