Thèse Vers une Compréhension Contextualisée de la Tranistion Agroécologique dans le Context des Changements Mondiaux H/F Doctorat.Gouv.Fr

Établissement : Université Toulouse II Jean Jaurès École doctorale : EDMITT - Ecole Doctorale Mathématiques, Informatique et Télécommunications de Toulouse Laboratoire de recherche : IRIT : Institut de Recherche en Informatique de Toulouse Direction de la thèse : Nathalie HERNANDEZ ORCID 0000000338453243 Début de la thèse : 2027-09-01 Date limite de candidature : 2026-11-23T23:59:59 Cette thèse propose de développer une méthodologie, fondée sur la connaissance, visant à mesurer et à analyser les impacts sociétaux de la transition agroécologique dans le contexte du changement global, en reliant diverses sources de données. Les technologies de représentation des connaissances et du Web sémantique seront utilisées pour décrire les systèmes agricoles, les acteurs, les pratiques, les technologies, les événements, les indicateurs et la provenance. Les technologies d'adaptation au domaine, GraphRAG (Peng et al., 2026) et les modèles de fondation sur graphes (Graph Foundation Models, Wang et al., 2025) seront étudiés afin d'adapter les grands modèles de langue généralistes pour la découverte de données liées mais hétérogènes dans le domaine agricole. L'expertise agronomique permettra d'interpréter les conséquences sur les cultures au stress climatique, tandis que le cadre d'intégration mettra en relation ces réactions avec les pratiques agricoles et leurs conséquences sociétales. Agricultural and food systems are being reshaped by several dimensions of global change. Among them, climate change is recognized as one of the major challenges facing these systems, affecting temperature and precipitation patterns and increasing exposure to extreme events, such as droughts, floods, hailstorms, and heat waves (Jones, 2016; IPCC, 2021; Bhardwaj & Mishra, 2026). Between 1980 and 2012, Europe experienced an approximately 60% increase in the occurrence of loss-relevant natural extreme events (Hov et al., 2013). These phenomena impose additional environmental stresses on water resources, soil quality, biodiversity, agricultural productivity, and food security (Montoroi, 2018; Viglizzo & Ricard, 2021; Qiu et al., 2023).
In parallel, artificial intelligence (AI), connected equipment, agricultural robotics, and Agriculture~3.0 are transforming how farms are observed and managed. Through technologies such as sensors, drones, satellite imagery, machine learning algorithms, computer vision, and autonomous machinery, large volumes of heterogeneous data can be collected within short periods of time. These information sources support more adaptive and precise decision-making in agricultural systems. As a result, AI-based approaches are increasingly being applied to irrigation management, harvest timing, yield forecasting, stress detection, nutrient monitoring, soil assessment, and the early detection of diseases, pests, and nutritional deficiencies through image analysis (Peng et al., 2022; Chen et al., 2023; Lyu et al., 2023; Wilson et al., 2024; Oliveira & Silva, 2023). Such approaches are also being used in post-harvest processes, including grape quality assessment and wine production (Fuller et al., 2021). These developments interact with agroecological transitions that seek to reduce environmental pressures while maintaining viable, resilient, and socially equitable food systems.
The societal impacts of agroecological transitions cannot be interpreted independently of context. The same technology or agricultural practice may have different, or even opposing, effects depending on the territory, climatic conditions, soil, production systems, infrastructure, historical trajectory, and objectives of the actors concerned. Similarly, context-dependent responses are widely observed in agriculture, where crop performance results from interactions between genotype, environment, and management. Consequently, understanding agricultural transitions requires considering multiple interacting environmental, agronomic, technological, and socio-economic factors. Such interactions are often complex and difficult to analyze because they operate across multiple spatial and temporal scales.
Large volumes of heterogeneous data are now available to support such analyses. These data originate from multiple sources, including environmental monitoring systems, remote sensing platforms, agricultural robotics, periodic agronomic bulletins, and observations on social media (Jiang et al., 2020; Jiang et al., 2022). Although they provide complementary evidence about the evolution of agricultural systems, they differ in format, modality, spatial and temporal scale, quality, completeness, and provenance. Moreover, relationships between heterogeneous data are often implicit, fragmented, uncertain, or only partially observable.
Additionally, agricultural datasets are frequently incomplete due to missing observations, inconsistent measurements, differences in data collection protocols, and uneven spatial or temporal coverage. Such limitations hinder comprehensive and context-aware analyses of agroecological transitions. Although large amounts of information are now available, establishing relationships among environmental conditions, agricultural practices, technological interventions, and social outcomes remains a challenging task. Existing approaches are often restricted to specific datasets, applications, or data modalities, making it difficult to integrate complementary sources of information into a unified representation of agricultural systems. Therefore, there is a need for approaches capable of organizing knowledge and establishing links between heterogeneous data while preserving contextual information and supporting reasoning over incomplete data.
In this sense, this research proposes a smart knowledge-driven framework based on multimodal knowledge graphs, link prediction, and GraphRAG (Tual et al., 2026) techniques to integrate heterogeneous agricultural data and support the analysis of agroecological transitions under global change. By combining environmental, agronomic, technological, and social information within a unified representation, the proposed framework is expected to improve the integration of heterogeneous datasets and facilitate the discovery of previously unobserved relationships. This may contribute to more comprehensive analyses of agricultural systems and to a better understanding of the factors influencing agroecological transitions across different contexts. Therefore, the outcomes of this research have the potential to support researchers, decision-makers, and stakeholders involved in agricultural development by providing methods and tools for more informed, context-aware, and data-driven analyses of agricultural systems undergoing transformation.
(1)Design a multi-modal knowledge graph, linking environmental, agronomic and social data while preserving their uncertainty, incompleteness and spatio-temporal context
(2)Develop and compare link prediction methods, assess their generalizability across agricultural contexts and robustness to missing data
(3)Construct and validate new benchmarks and contextualized evaluation metrics to assess both link prediction soundness and its contribution to multi-factor analysis of the social impact of agroecological transition under global changes.
The research will follow an iterative and experimental methodology. Each iteration will combine computational experiments with validation by domain experts. The methodology will pay particular attention to the provenance, uncertainty, spatial and temporal context, and potential biases of the data. The first year will be dedicated to formalizing modality availability patterns in data lakes and designing a modality-aware GraphRAG system, evaluated against existing data discovery benchmarks. The second year will focus on the implementation and comparative evaluation of several generative imputation paradigms, involving the construction of a benchmark for incomplete data lakes created through controlled modality ablation. The third year will integrate MoE encoding, generative imputation, and KG representations into a unified pipeline, evaluated using the constructed benchmark and real-world open data lakes.