Optimised waveband selection for low-cost multispectral estimation of tomato lycopene concentration using machine learning

This paper presents an investigation into a cost-effective method for non-destructive lycopene quantification in tomatoes using multispectral imaging, while aiming for high precision and practical applicability across various phases of tomato harvesting, processing, and storage. Lycopene, a carotenoid with antioxidant properties, is known for its health benefits, with its consumption being linked with reduced risk of cardiovascular disease, cancer, and neurodegenerative disorders. Tomatoes are the primary dietary source of lycopene due to their high concentration levels and widespread consumption. This study adopts a multispectral imaging approach, strategically selecting wavebands to enhance sensitivity and accuracy beyond conventional RGB systems. It does this while limiting the number of wavebands to the minimum required to reduce hardware complexity and operational costs. A primary contribution of this work lies in the streamlined approach to waveband selection in optimised capture conditions, which iteratively adds wavebands and evaluates their individual contributions to the model's performance using the coefficient of determination of predictors (R²). The method is validated through repeated cross-validation. The study evaluates four machine learning methods—SVR (R² = 0.940), k-NN (0.920), CNN (0.932), and SNN (0.959), to assess their performance on low-cost hardware. Notably, a simplified two-waveband configuration using a fast SNN achieved an R² of 0.951 and RMSEP of 6.317mg/kg, offering substantial reductions in hardware cost and processing time while maintaining high predictive accuracy, making it a promising and inexpensive solution.