Integrating Sensor Fusion, Machine Learning, and Physics-Based Modelling for Predictive Shelf-Life Assessment of Apples in Cold Storage

Authors: Dhritiman Saha; Thongam Sunita Devi; Puneet Kumar; Ranjeet Singh; Jaskaran Singh Brar; Priyabrata Kapri

Book: Transforming Food Analysis and Control: Non-Destructive Technologies and Innovations

Editors: K. Prasad, S. Shekhar, M. K. Paswan, and R. Ahuja

Publisher: SLIET Longowal

Publication year: 2026

Publication date: 2026-07-20

ISBN: 978-81-990304-1-1 (e-book); 978-81-990304-0-4 (hardback)

Chapter pages: 181–200

DOI: To be added after DOI registration.

Abstract

Accurate prediction of shelf-life in cold-stored apples remains a complex challenge due to the interplay of physiological, biochemical, and environmental factors. This study develops an integrated computational framework combining dense sensor data acquisition, machine learning (ML) algorithms, and physics-based modeling to predict core temperature dynamics and shelf-life trajectories of Red Delicious apples over six months. The high-resolution dataset includes continuous measurements of temperature, humidity, and critical quality parameters such as total soluble solids (TSS), titratable acidity (TA), moisture, and weight loss. Core temperature prediction was achieved using extreme gradient boosting (XGBoost) and long short-term memory (LSTM) networks, with XGBoost outperforming LSTM by attaining an R² of 0.88. An entropy-based multi-criteria weighting methodology objectively ascertained the relative importance of multiple quality indices, emphasizing TSS and titratable acidity. Incorporation of a temperature and humidity modified Arrhenius equation provided an explicit mechanistic model for TSS degradation kinetics, capturing nonlinear decay trajectories under realistic storage conditions. Critically, hybrid modeling, integrating autoregressive lagged TSS features with ML-predicted core temperature and physics-modeled degradation rates yielded superior TSS prediction (R² ≈ 0.99), underscoring the temporal autocorrelation and temperature dependencies essential for robust shelf-life estimation. However, inherent limitations include potential noise amplification in numerical differentiation of quality rates and assumptions of homogeneous storage conditions. Validation through time-series cross-validation protocols minimized information leakage and evaluated generalizability. This comprehensive approach demonstrates the feasibility of digital twin frameworks that fuse data-driven and mechanistic models for precision post-harvest management.

Keywords: apple cold storage; shelf-life prediction; core temperature; total soluble solids; machine learning; Arrhenius kinetics; digital twin; XGBoost; hybrid modeling

Full text: Download the chapter PDF