A quality decay model with multinomial logistic regression and image-based deep learning to predict the firmness of ‘Conference’ pears in the downstream supply chains

Traditional quality-decay models (e.g., multinomial logistic regression) for fruit quality classification deals with tabular data which focus mainly on the storage parameters such as storage duration and conditions (D&C). Those parameters have the effects on quality decay at an aggregate scale across different experimental levels; they are not good at capturing the variations within each experimental level. This may restrict the predictive power of the traditional model. On the contrary, image-based deep learning models are dealing with individual products and can extract the deep features of each fruit to provide individual-based quality information but lack information regarding the post-harvest conditions (time of harvest, storage conditions etc.). In this research, we investigate the combined performance of the multinomial logistic regression with the image-based convolutional neural network (CNN) for quality prediction of ‘Conference’ pears (Pyrus communis L.) (measured by firmness) where the extracted deep features are used as the explanatory variables for the logistic regression model. The results show that combining deep features with D&C parameters are likely to improve the predictive power of the logistic regression model to predict the firmness of the conference pears. The managerial implications as well as future research directions are also discussed. • Traditional models focus on storage conditions, limiting predictive accuracy for fruit quality. • Image-based deep learning models provide individual fruit insights but lack post-harvest data. • Combining CNN deep features with logistic regression improves firmness prediction accuracy. • Deep features serve as explanatory variables in the multinomial logistic regression model. • Results suggest enhanced predictive power by integrating deep features with storage parameters.