Cross-Market Non-Destructive Quality Assessment of Arumanis Mangoes Using Genetic Algorithm–Optimized Ensemble Models and Low-Cost AS7265x UV–Vis–NIR Spectroscopy

Research background. Arumanis mango is a high-value tropical fruit whose market acceptance depends primarily on internal quality attributes. However, reliable quality grading remains difficult because Arumanis is a green-mature cultivar that shows minimal external colour changes during ripening. Moreover, most existing non-destructive sensing studies rely on controlled experimental conditions, high-cost instruments, or homogeneous sample sources, with limited evaluation of model robustness across heterogeneous market environments. Experimental approach. This research evaluates a low-cost, portable, and non-destructive quality assessment framework combining an AS7265x UV-Vis-NIR multispectral sensor with a genetic algorithm-optimised committee machine (GA-CM) to predict physicochemical parameters of Arumanis mangoes. A total of 200 samples were collected from both modern and traditional markets to capture variability in maturity levels and postharvest handling practices. Non-destructive spectral reflectance data were acquired at 18 wavelengths (410–940 nm), followed by destructive reference measurements of total soluble solids, total sugar, vitamin C, pH, and water content. Three base learners, e.g. linear regression (LR), support vector regression (SVR), and artificial neural network (ANN), were developed and integrated using two ensemble strategies: simple averaging and GA-based weighted averaging. Model performance was assessed using the coefficient of determination (R²) and mean squared error (MSE). Results and conclusions. The results demonstrate that mangoes from different market typologies exhibit similar spectral trends but differ in reflectance magnitude, reflecting variations in maturity and postharvest conditions and yielding a robust, heterogeneous dataset. Among individual models, ANN and SVR significantly outperformed LR, with ANN showing superior capability in capturing non-linear relationships between spectral signatures and internal quality attributes. Ensemble modelling further enhanced prediction accuracy, with the GA-CM consistently outperforming both individual learners and simple averaging by optimally weighting base models. High predictive performance was achieved across all quality parameters, with an average R² of approximately 0.97. Novelty and scientific contribution. This study provides a novel contribution by introducing a GA-optimised ensemble framework that leverages complementary modelling strengths under low spectral resolution conditions. By validating the system across mixed supply chains rather than controlled environments, the proposed approach demonstrates strong generalisation capability and practical applicability for real-world, non-destructive quality assessment of Arumanis mango.