A Framework for Enhancing Shelf Life Prediction of Fresh Produce using Machine Learning

Thesis Abstract

Effective prediction of shelf life for fresh produce remains a critical challenge in food supply chains, impacting quality assurance, waste reduction, and consumer satisfaction. Current methods largely rely on empirical observations, basic statistical models, or heuristic guidelines, which often lack accuracy across diverse produce types and varied storage conditions. This study aims to develop a robust, flexible framework for enhancing shelf life prediction of fresh produce through the application of machine learning techniques, thereby providing a scientific basis for more precise and adaptable shelf life estimations. The specific objectives include (1) identifying key physicochemical and environmental factors influencing produce spoilage; (2) designing a comprehensive dataset encompassing these factors for multiple produce categories; (3) evaluating the performance of various machine learning algorithms, including support vector machines, random forests, and neural networks, in predicting shelf life; and (4) formulating an integrated predictive framework with practical usability for stakeholders in the food supply chain. The research adopts a mixed-methods approach, combining quantitative data collection and analysis with theoretical validation. The study population comprises selected fresh produce items—such as strawberries, lettuce, and avocados—sourced from several local suppliers, with a total sample size of 300 units (100 per produce category). Data collection involves both laboratory measurements and field observations physicochemical parameters including pH, moisture content, color change, firmness, microbial load, and volatile compound profiles are measured using standard analytical techniques such as high-performance liquid chromatography (HPLC), texture analyzers, and microbial assays. Environmental factors, such as temperature, relative humidity, and ethylene levels, are continuously monitored using data loggers. In addition, visual spoilage scores are recorded periodically. Data preprocessing includes normalization, feature selection via principal component analysis (PCA), and dataset partitioning into training and testing sets, with a 7030 split. The machine learning models are trained and validated through cross-validation methods, and their performance is evaluated using metrics including accuracy, precision, recall, F1-score, and root mean squared error (RMSE). The analytical framework incorporates supervised learning algorithms, with hyperparameter tuning conducted through grid search. Furthermore, the study applies the theory of predictive analytics and models based on the Theory of Constraints to optimize the prediction process, addressing bottlenecks in data acquisition and model deployment. Results are anticipated to demonstrate that some machine learning algorithms outperform traditional statistical models in accurately forecasting shelf life across different produce types and storage conditions. For instance, neural networks are expected to offer high predictive accuracy due to their ability to model nonlinear relationships among multiple variables. The study's primary contribution to knowledge lies in proposing a validated, adaptable predictive framework that integrates various physicochemical and environmental parameters through machine learning techniques for estimating the shelf life of diverse fresh produce. This framework can serve as a decision-support tool for producers, distributors, and retailers, ultimately reducing waste and enhancing consumer safety. The findings are expected to highlight the importance of incorporating multidimensional data and advanced algorithms to improve shelf life estimation accuracy, challenging conventional models that rely solely on superficial or unidimensional measures. The main conclusion underscores the superior performance of machine learning approaches, particularly neural networks and random forests, in predicting produce spoilage, thus advocating for their integration into standard food quality management practices. Recommendations include further refinement of the model with larger datasets, integration of real-time sensor data, and development of user-friendly software applications for industry stakeholders. The study advocates ongoing research into the application of emerging machine learning techniques, such as deep learning and ensemble methods, to further enhance prediction reliability. Overall, this research contributes substantively to the field of Food Science and Technology by demonstrating the efficacy of data-driven, AI-enabled solutions for preserving produce quality and reducing wastage in supply chains.