Design and Deployment of a Digital Twin for the Optimization of an Intelligent Irrigation System in Precision Agriculture

Abstract

Agriculture is increasingly affected by water scarcity, climate variability, soil degradation, and the growing demand for food production. Irrigation management is one of the most critical factors in this context, because traditional irrigation practices are often based on fixed schedules, manual observation, or empirical experience. Such approaches can lead to water waste, plant stress, and reduced productivity. In response, this thesis presents the design and deployment of Agrio, a Digital Twinbased intelligent irrigation platform for precision agriculture. The proposed system integrates a physical smart farming prototype, Libelium Smart Agriculture sensing equipment, XBee radio-frequency communication, a Raspberry Pi gateway, HiveMQ/MQTT messaging, a FastAPI backend, a PostgreSQL database, a Random Forest-based machine learning service, a Digital Twin state and simulation layer, an explainable Multi-Agent System (MAS) decisionsupport service, a Next.js web dashboard, a Flutter mobile application, a WeMos actuator, and a water pump. The platform receives real sensor data from a controlled prototype environment, normalizes and stores the measurements, updates the virtual state of the irrigation zone, predicts irrigation need, simulates future moisture evolution, combines Digital Twin, ML, weather, agronomic-note, water-balance, and safety evidence, generates recommendations, supports human validation, schedules irrigation actions, and controls the physical pump through MQTT commands. The scientific contribution of this work is the study of Digital Twins in agriculture, with a focus on their maturity levels, enabling technologies, irrigation applications, and limitations. The engineering contribution is the implementation of an end-to-end cyber-physical prototype that connects sensing, communication, backend processing, artificial intelligence, simulation, an explainable MAS decision layer, user interfaces, deployment, and actuation. The MAS is not a computer-vision system; plant observations are provided as farmer text notes and combined with structured sensor, weather, ML, water-balance, and safety evidence. The results show that the system goes beyond a classical IoT monitoring dashboard by providing synchronization, prediction, what-if simulation, explainable recommendation, scheduling, and physical control. The current implementation remains a prototype tested in a controlled environment, and future work should extend it to long-term field validation, stronger agronomic calibration, cybersecurity, and advanced optimization techniques***

L’agriculture moderne fait face à des défis importants liés à la rareté de l’eau, aux variations climatiques et à la nécessité d’optimiser la production agricole. Dans ce contexte, l’irrigation intelligente constitue une application centrale de l’agriculture de précision. Ce mémoire présente la conception et le déploiement d’Agrio, une plateforme d’irrigation intelligente basée sur le concept de jumeau numérique. Le système proposé combine un prototype physique, un kit Libelium Smart Agriculture, une communication radiofréquence XBee, une passerelle Raspberry Pi, la messagerie MQTT/HiveMQ, un backend FastAPI, une base de données PostgreSQL, un service d’apprentissage automatique basé sur Random Forest, une couche de simulation du jumeau numérique, un tableau de bord web, une application mobile Flutter, un actionneur WeMos et une pompe d’eau. La contribution scientifique porte sur l’étude de l’état de l’art des jumeaux numériques en agriculture et leur positionnement par rapport aux systèmes d’irrigation intelligents existants. La contribution d’ingénierie réside dans la mise en oeuvre d’une chaîne complète reliant les capteurs, la communication IoT, le traitement backend, la prédiction, la simulation, un système multi-agent explicable, la recommandation, la validation humaine et le contrôle physique de l’irrigation. Le MAS ne repose pas sur une caméra; les observations sur les plantes sont saisies sous forme de notes textuelles du fermier. Les résultats montrent que le système dépasse la simple visualisation IoT en intégrant la synchronisation, la prédiction, la simulation, l’explication de la décision et l’action supervisée