Developing a novel prediction method for dual-parameter passion fruit quality prediction based on transmission spectroscopy

This study aimed to develop a non-destructive method for assessing passion fruit quality by simultaneously predicting soluble solid content (SSC) and edible rate (ER) using hyperspectral transmittance imaging. To overcome spatial heterogeneity due to fruit morphology, genetic algorithm (GA) optimized region-of-interest (ROI) selection and non-uniform rational B-spline curve (NURBS) fitting were introduced. Key wavelengths were selected by comparing multiple algorithms, and machine learning models were evaluated for prediction performance. Results demonstrated that GA-optimized ROIs tailored to SSC and ER significantly enhanced accuracy over full-region analysis. Wavelength selection strategies diverged: SPA prioritized NIR bands for SSC, while SiPLS emphasized visible bands for ER. The proposed approach significantly improved prediction accuracy, with the best-performing random forest regression (RFR) model achieving superior performance (SSC: R² = 0.922, RMSE = 0.295°Brix; ER: R² = 0.913, RMSE = 1.166 %). This study provides an effective framework for non-destructive quality evaluation of thick-skinned fruits, with direct implications for automated sorting systems in the fruit industry.