Prediction of lignan content in sesame seeds using near-infrared hyperspectral imaging and regression modeling

Abstract Sesame ( Sesamum indicum L.) seeds contain lignans, including sesamin and sesamolin, which possess important health-promoting properties. However, the simultaneous quantitative prediction of individual lignans, such as sesamin and sesamolin, using near-infrared hyperspectral imaging remains limited. In this study, a rapid, non-destructive method was developed to predict lignan content in sesame seeds using near-infrared hyperspectral imaging combined with regression models. A total of 180 F 5 lines derived from a cross between Milyang 78 and Milyang 82 were analyzed. Hyperspectral images were acquired in the 1,050–1,600 nm range, and five spectral preprocessing methods were applied: raw spectra (Raw), Savitzky–Golay smoothing (SG), standard normal variate (SNV), multiplicative scatter correction (MSC), and 1st derivative. Effective wavelengths were selected using genetic algorithm (GA) and competitive adaptive reweighted sampling (CARS). Four regression models–partial least squares regression (PLSR), multiple linear regression (MLR), support vector regression (SVR) and eXtreme Gradient Boosting (XGBoost)–were evaluated. Sesamin prediction was mainly associated with the 1,400–1,550 nm region, whereas sesamolin was associated with the 1,100–1,250 nm region. Effective wavelength selection significantly improved prediction performance for all lignans. The optimal models achieved R 2 = 0.834, RMSE = 0.937 mg/g, and NRMSE = 8.692% for sesamin (SG-GA-PLSR); R 2 = 0.618, RMSE = 0.380 mg/g, and NRMSE = 7.591% for sesamolin (MSC-CARS-MLR); and R 2 = 0.793, RMSE = 1.196 mg/g, and NRMSE = 8.521% for total lignan (Raw-CARS-PLSR). This study demonstrates the potential of hyperspectral imaging as an efficient tool for high-throughput screening of lignan content in sesame breeding programs.