Food computing refers to the integration of digital technologies, such as artificial intelligence (AI), the Internet of Things (IoT), and data-driven approaches, to address various challenges in the food sector. It encompasses a wide range of technologies that improve the efficiency, safety, and sustainability of food systems, from production to consumption. It represents a transformative approach to addressing challenges in the food sector by integrating AI, the IoT, and data-driven methodologies. Unlike traditional food systems, which primarily focus on production and safety, food computing leverages AI for intelligent decision making and the IoT for real-time monitoring, enabling significant advancements in areas such as supply chain optimization, food safety, and personalized nutrition. This review highlights AI applications, including computer vision for food recognition and quality assessment, Natural Language Processing for recipe analysis, and predictive modeling for dietary recommendations. Simultaneously, the IoT enhances transparency and efficiency through real-time monitoring, data collection, and device connectivity. The convergence of these technologies relies on diverse data sources, such as images, nutritional databases, and user-generated logs, which are critical to enabling traceability and tailored solutions. Despite its potential, food computing faces challenges, including data heterogeneity, privacy concerns, scalability issues, and regulatory constraints. To address these, this paper explores solutions like federated learning for secure on-device data processing and blockchain for transparent traceability. Emerging trends, such as edge AI for real-time analytics and sustainable practices powered by AI–IoT integration, are also discussed. This review offers actionable insights to advance the food sector through innovative and ethical technological frameworks.

Federated Learning (FL), a key paradigm in privacy-preserving and distributed machine learning (ML), enables collaborative model training across decentralized data sources without requiring raw data exchange. FL enables collaborative model training across decentralized data sources while preserving privacy. However, selecting appropriate clients remains a major challenge, especially in heterogeneous environments with diverse battery levels, privacy needs, and learning capacities. In this work, a centralized reward-based payoff strategy (RBPS) with cooperative intent is proposed for client selection. In RBPS, each client evaluates participation based on locally measured battery level, privacy requirement, and the model’s accuracy in the current round computing a payoff from these factors and electing to participate if the payoff exceeds a predefined threshold. Participating clients then receive the updated global model. By jointly optimizing model accuracy, privacy preservation, and battery-level constraints, RBPS realizes a multi-objective selection mechanism. Under realistic simulations of client heterogeneity, RBPS yields more robust and efficient training compared to existing methods, confirming its suitability for deployment in resource-constrained FL settings. Experimental analysis demonstrates that RBPS offers significant advantages over state-of-the-art (SOA) client selection methods, particularly those relying on a single selection criterion such as accuracy, battery, or privacy alone. These one-dimensional approaches often lead to trade-offs where improvements in one aspect come at the cost of another. In contrast, RBPS leverages client heterogeneity not as a limitation, but as a strategic asset to maintain and balance all critical characteristics simultaneously. Rather than optimizing performance for a single device type or constraint, RBPS benefits from the diversity of heterogeneous clients, enabling improved accuracy, energy preservation, and privacy protection all at once. This is achieved by dynamically adapting the selection strategy to the strengths of different client profiles. Unlike homogeneous environments, where only one capability tends to dominate, RBPS ensures that no key property is sacrificed. RBPS thus aligns more closely with real-world FL deployments, where mixed-device participation is common and balanced optimization is essential.

Ochratoxin A (OTA) contamination presents significant risks in viticulture, affecting the safety and quality of wine and grape-derived products. This study introduces a groundbreaking method for early detection and management of OTA, leveraging environmental data such as temperature and humidity. A function derived from chemical analysis was developed to estimate OTA concentrations and used to label a synthetic dataset, establishing safe thresholds. Two AI models were trained: one for the detecting of OTA presence and the other for classifying the concentration range. These models were deployed on a M5Stick C+, a microcontroller designed for real-time data processing. The inference process is optimized for rapid response, requiring minimal time to deliver results. Additionally, the low power consumption of the M5Stick C+ ensures that the device can operate throughout the harvest period on a single charge. The system is able to transmit inference data via MQTT for real-time analysis. This comprehensive approach offers a scalable, cost-effective, on-site solution that is autonomous, eliminating the need for domain experts and extensive resources. The robustness of the system was demonstrated through its consistent performance across multiple test sets, providing an effective enhancement to food safety in grape and wine production. The study also details the system architecture, describes the function used for data labeling, outlines the training and deployment processes of the models, and finally, assesses the testing of the overall system.

This study presents a novel federated learning (FL) methodology implemented directly on STM32-based microcontrollers (MCUs) for energy-efficient smart irrigation. To the best of our knowledge, this is the first work to demonstrate end-to-end FL training and aggregation on real STM32 MCU clients (STM32F722ZE), under realistic energy and memory constraints. Unlike most prior studies that rely on simulated clients or high-power edge devices, our framework deploys lightweight neural networks trained locally on MCUs and synchronized via message queuing telemetry transport (MQTT) communication. Using a smart agriculture (SA) dataset partitioned by soil type, 7 clients collaboratively trained a model over 3 federated rounds. Experimental results show that MCU clients achieved competitive accuracy (70–82%) compared to PC clients (80–85%) while consuming orders of magnitude less energy. Specifically, MCU inference required only 0.95 mJ per sample versus 60–70 mJ on PCs, and training consumed ∼70 mJ per epoch versus nearly 20 J. Latency remained modest, with MCU inference averaging 3.2 ms per sample compared to sub-millisecond execution on PCs, a negligible overhead in irrigation scenarios. The evaluation also considers the payoff between accuracy, energy consumption, and latency through the Energy Latency Accuracy Index (ELAI). This integrated perspective highlights the trade-offs inherent in deploying FL on heterogeneous devices and demonstrates the efficiency advantages of MCU-based training in energy-constrained smart irrigation settings.