Machine Learning-Based Prediction of Firmness in Coated Bananas Under Retail Conditions

This study explored the use of machine learning (ML) algorithms for the non-destructive prediction of banana firmness under retail conditions, facilitating real-time quality assessment, optimising supply chain decisions, and improving postharvest management. Bananas were coated with the optimal formulation of Opuntia ficus-indica mucilage (OF) and stored at 23 ± 2 ℃ for 10 days. A factorial experimental design was employed, with edible coating and storage duration as primary factors. Banana parameters, including respiration rate, ethylene production, weight loss, and colour, were measured alongside firmness. Collected data was used to develop predictive models using ML techniques, namely Partial Least Squares (PLS) regression, Ridge regression, and Elastic Net regression. The results showed that banana firmness could be predicted using non-invasive attributes, with respiration rate and weight loss being the most influential predictors. Among the models tested, PLS regression exhibited the highest predictive accuracy, with an R² of 0.978, RMSE of 0.097, MAE of 0.009, and R²-adjusted value of 0.940. Ridge regression followed closely (R² of 0.972, RMSE of 0.110, MAE of 0.012, and R²-adjusted of 0.922), while Elastic Net regression, though slightly less precise, still demonstrated strong predictive capability (R² = 0.956, RMSE = 0.142, MSE = 0.020, R²-adjusted = 0.801). This study also demonstrated that the application of an optimised Opuntia ficus-indica mucilage extended the shelf-life of bananas by four days. This approach allows real-time quality assessment, enhancing quality control, reducing postharvest losses, and improving inventory management in the fruit industry.