Université de Sherbrooke
Ph.D. candidate
Supervisor: Samuel Foucher
Mickaël Germain
Start: 0000-00-00
End: 0000-00-00

Project

A Federated Intelligent Digital Infrastructure for Decision Support in Integrated Crop Protection of Major Field Crops in Quebec: Integrating Multi-Source Data and Spatio-Temporal AI Models for Soybea
Quebec agriculture faces a critical challenge: sustaining the productivity of major field crops while reducing the environmental impacts associated with pesticides and other crop protection interventions. This challenge is particularly acute in Montérégie, a major agricultural region where maize and soybean production are central to the agroecosystem. It is further intensified by climate variability, the increasing pressure of key biotic stressors such as soybean Sclerotinia stem rot and common lambsquarters (Chenopodium album) in maize, and the fragmentation of information required for timely and reliable decision-making. A wide range of decision-relevant resources is currently available, including field observations, weather data, in situ sensor measurements, satellite imagery, institutional archives, predictive models, and agronomic knowledge bases. However, these resources remain distributed across heterogeneous systems and are difficult to integrate within a coherent, operational, and privacy-compliant decision-making framework. This limitation is especially important in Quebec, where data governance is shaped by a stringent regulatory environment, including Law 25 on the protection of personal information. As a result, existing decision support systems often struggle to generate crop protection alerts that are context-aware, traceable, explainable, and directly useful to agronomists and producers. To address this gap, this research proposes the design of a federated intelligent digital infrastructure to support a crop protection decision support system for major field crops in Quebec. The proposed framework is structured around three complementary research axes. The first focuses on the integration and orchestration of distributed agricultural resources through a multi-agent architecture. The second addresses context engineering, a central scientific challenge in this research, by combining structured memory, dynamic retrieval of task-relevant information, and continuous refinement of system responses over time. The third involves the development of a structured testbed to evaluate the system’s coherence, traceability, and agronomic usefulness after its technical feasibility has been established. The proposed system is built around a conversational interface powered by a large language model. This interface acts as an orchestration layer that coordinates specialized agents and mobilizes the most appropriate tools for each task, including partner predictive models, structured geospatial queries, and knowledge retrieval from a curated documentary knowledge base. Rather than replacing agronomic expertise, the objective is to augment expert judgment by delivering information that is integrated, contextualized, transparent, and actionable. By developing a realistic proof of concept grounded in the agronomic, institutional, and regulatory context of Quebec, this research aims to contribute both to applied crop protection decision-making and to the methodological advancement of intelligent, federated, and context-aware decision support systems in agriculture.

Keywords

Agriculture de précision, aide à la décision, Phytoprotection, télédétection, Systèmes multi-agents, Apprentissage fédéré