Temperature plays a crucial role in preserving the quality and safety of food products, exerting a direct influence on the shelf life of most perishable items (Chen & Shaw, 2011; Karacan et al., 2023; Mercier & Uysal, 2018). Detailed monitoring of temperature, with maximum spatial resolution possible, allows for the description of temperature distribution at different points within a reefer container or a fruit pallet itself (Maruthi et al., 2023; Mercier & Uysal, 2018; Montanari, 2008; Zou et al., 2023). Such information enables the cold chain operators to take remedial actions to maintain product quality and prevent food loss (Gillespie et al., 2023; Ndraha et al., 2018; Swain & Jenamani, 2022).

Owing to the economical and operational limitations of physical sensors (Cheng et al., 2021; Guzmán et al., 2019; Mercier et al., 2017; Zou et al., 2023), prior studies have resorted to temperature estimation methods to enhance the spatial resolution of measurements using data from a limited number of physical sensors (Badia-Melis et al., 2018; Loisel et al., 2021). Such estimator is commonly referred to as a Virtual Temperature Sensor (VTS) or Soft Temperature Sensor (Brunello et al., 2021; Stavropoulos et al., 2023). Virtual sensors are often developed using data-driven approaches, as they can capture complex factors influencing temperature changes without the need for a comprehensive mathematical model (Hoang et al., 2021; Jung et al., 2020; Lim et al., 2022; Perera et al., 2023; Xu et al., 2022).

The development process of a data-driven virtual sensor includes: data acquisition, input variable selection and reduction, model choice and identification, and model validation (Curreri et al., 2020; Souza et al., 2016). Upon performing a comprehensive literature review, the following gaps were identified during the process: the lack of non-subjective methodologies for Input Variable Selection (IVS) and how to reduce the required number of the source-points, the limited investigations over hyper-parameter tuning and optimisation, and the insufficient validation of models utilising real-world conditions. These gaps highlight the need for further research and development to enhance the reliability and generalisability of data-driven virtual sensors.

In this study, all identified gaps were examined, and appropriate solutions were suggested and evaluated to resolve the issues. By addressing the stated shortcomings, this study aims to contribute to the advancement of fruit-pallet VTS development through a fully systematic framework utilising real-world data. The key innovation lies in the practical adaptation of established IVS techniques into an ensemble-like procedure, where multiple iterations are aggregated to generate robust feature rankings. This systematic approach, combined with the statistical analysis of key factors influencing the VTS performance as well as out-of-sample generalisation assessment, strengthens the applicability of VTS design for cold-chain operations.