Spectral regions informative for polyphenols and flavonoid content in spinach. Open field reflectance investigation through linear regression and machine learning

Reflectance spectroscopy in the visible to near-infrared range (vis-NIR) offers a non-destructive method to assess the biochemical quality of leafy vegetables. In this study, we aimed to evaluate the potential of vis-NIR spectroscopy to predict total polyphenol content (TPC) and total flavonoid content (TFC) in spinach (Spinacia oleracea L.) leaves, and to identify robust spectral quality indicators at the leaf level. Two spinach cultivars were grown under open-field conditions across two growing seasons (2020/2021 and 2021/2022), with six nitrogen fertilization levels (0–250 kg N ha−1). Spectral signatures were collected at the leaf level and matched with laboratory measurements of TPC, TFC, and dry matter content (DM). Partial least squares regression (PLSR) showed limited performance in predicting TPC on a fresh weight basis (TPC_FW: R2 = 0.47; RMSEσ = 0.72), while it achieved better results for TFC_FW (R2 = 0.68; RMSEσ = 0.55) and DM (R2 = 0.63; RMSEσ = 0.61). The random forest (RF) algorithm showed moderate performance for TFC_FW (R2 = 0.58; RMSEσ = 0.65) and DM (R2 = 0.60; RMSEσ = 0.65) but was ineffective for TPC_FW. Key wavelengths selected by both models showed overlapping regions related to chemical absorption features, revealing partial spectral redundancy between traits. However, the influence of DM on analyte prediction was limited. This work demonstrates the feasibility of developing rapid, leaf-level spectral proxies for nutraceutical traits in fresh spinach, offering a step toward in-field quality monitoring for leafy vegetables.