Prediction and visualization of sugar content distribution in sugar beet using hyperspectral imaging

Extended Abstract
Introduction
Sugar beet (Beta vulgaris) is the second-largest source of sugar globally and plays a critical role in food security. The sucrose content (sugar content) of sugar beet is a key indicator that determines its economic value for growers, the quality of sugar production in factories, and the performance of cultivars in research. Traditionally, polarimetry has been used to measure sugar content; however, it is time-consuming, relies on hazardous chemicals like lead acetate, and requires skilled personnel. These limitations necessitate faster, cost-effective, and reliable alternatives. This study explores hyperspectral imaging to predict and visualize sugar content in sugar beet paste, addressing the research question: can hyperspectral imaging accurately estimate and map sugar content at the pixel level?. The hypothesis is that hyperspectral imaging, combined with advanced preprocessing and regression models, can provide a robust alternative to polarimetry.
Materials and Methods
A total of 150 sugar beet samples, each containing 40–50 roots, were randomly collected from consignments delivered to the Torbat-Heydarieh Sugar Company. The samples were washed, pulped using an industrial 8-blade pulper, and homogenized for approximately 3 minutes to produce 400–500 grams of uniform pulp per sample. Each pulp was then divided into two portions: one for polarimetric sugar content measurement at the Sugar Beet Seed Institute (Karaj, Iran), using a Betalyzer device at a wavelength of 589 nm (ICUMSA method), and the other for hyperspectral imaging at the Automation and Computer Vision Laboratory, Ferdowsi University of Mashhad. A desktop hyperspectral imaging system (Parto Afzar Sanat, Zanjan, Iran) operating in the 400–950 nm range with a spectral resolution of 2 nm was used. The system featured a 200 mm scanning length, a spatial resolution of 0.05 mm, and four halogen light sources. Images were periodically calibrated using both white and dark reference panels. Two preprocessing methods-Standard Normal Variate (SNV) and Savitzky-Golay (SG) with a 5-point window and third-degree polynomial-were applied. Two wavelength selection techniques were evaluated: Competitive Adaptive Reweighted Sampling (CARS) and Successive Projections Algorithm (SPA). Three regression models were used: Multiple Linear Regression (MLR), Partial Least Squares (PLS), and Support Vector Regression (SVR). Outliers were removed using Monte Carlo Partial Least Squares (MCPLS) with 5000 iterations. Data analysis was performed using MATLAB (2022b) on an Acer Aspire system equipped with 20 GB RAM and an Intel Core i7 processor.
Results and Discussion
Among the 18 regression models developed, the PLS model with SNV preprocessing and SPA wavelength selection outperformed the others, achieving a coefficient of determination (R²) of 0.89 and a root mean square error (RMSE) of 0.28 for calibration, and an R² of 0.91 and RMSE of 0.24 for validation. This model effectively identified key wavelengths associated with sucrose absorption in the 400–950 nm range, with regression coefficients indic ating strong correlations (e.g., peaks at 620 nm and 750 nm, linked to chlorophyll and a water overtone, respectively). For the first time, this study visualized the spatial distribution of sugar content at the pixel level in sugar beet pulp, yielding an R² of 0.71 and RMSE of 2.89. The lower pixel-level accuracy compared with mean predictions may stem from localized sucrose variations, spectral noise, or model limitations, as the model was trained on average spectra. This finding aligns with previous studies, which reported superior PLS performance in sugar content prediction. The SNV preprocessing significantly reduced the effects of light scattering, enhancing model accuracy. The SPA algorithm efficiently selected optimal wavelengths, reducing model complexity. This visualization capability enables heterogeneity analysis, offering applications in quality control and process optimization in sugar factories. However, pixel-level accuracy requires further improvement through advanced algorithms and noise reduction techniques. Future studies should explore environmental factors (e.g., humidity, temperature) and integrate hyperspectral imaging with polarimetry to enhance precision..
Conclusion
Hyperspectral imaging, combined with SNV preprocessing, SPA wavelength selection, and PLS regression, provides a rapid and reliable approach for predicting sugar content in sugar beet pulp, achieving high accuracy (R² = 0.91, RMSE = 0.24). The novel visualization of sugar content distribution at the pixel level offers valuable insights into sample heterogeneity, with promising applications in quality monitoring within the sugar industry. Further research is required to improve pixel-level accuracy and address susceptibility to environmental factors.