Fondements et état de l’art

Nathalie Mitton, Ludovic Brossard, Tassadit Bouadi, Frédérick Garcia, Romain Gautron, Nadine Hilgert, Dino Ienco, Christine Largouët, Évelyne Lutton, Véronique Masson, Roger Martin‐Clouaire, Marie-Laure Mugnier, Pascal Neveu, Pierre‐Marie Preux, Hélène Raynal, Catherine Roussey, Alexandre Termier, Véronique Bellon MaurelAgritrop 2022

Abstract

Modern agriculture is navigating a "complexity crisis," where the transition to sustainable agroecology requires managing intricate interactions between genetics, soil health, weather, and socio-economic factors. This paper synthesizes the current state of the art in digital agriculture, outlining a roadmap for integrating the Internet of Things (IoT), Big Data, hybrid modeling, and robotics. By establishing a seamless data lifecycle and prioritizing FAIR (Findable, Accessible, Interoperable, and Reusable) data principles, digital tools can provide the precision necessary to optimize resource use and reduce chemical inputs.

Introduction

The shift toward agroecology demands a move away from coarse observations and intuition toward high-resolution, data-driven management. Traditional agricultural methods often lack the granularity required for precision interventions. To bridge this gap, researchers are developing "responsible digital agriculture" frameworks that leverage a suite of technologies to capture, process, and act upon environmental data at various scales—from individual genes to global climate patterns.

The Data Lifecycle and Knowledge Engineering

A core challenge in digital agriculture is managing the "Big Data of the Living." Agricultural data is uniquely heterogeneous and spans multiple dimensions.

  • Sensing and Connectivity: The use of Low Power Wide Area Networks (LPWAN) and LoRa protocols enables remote sensing in isolated areas. For supply chain transparency, blockchain technology is being explored to provide tamper-proof traceability.
  • Semantic Interoperability: To ensure that data is actionable across different platforms, the paper emphasizes Semantic Web technologies (OWL, RDF, SPARQL). Utilizing ontologies like the FAO’s AGROVOC thesaurus and platforms like AgroPortal ensures that a soil sensor's data in one region is interoperable with decision-support tools developed elsewhere.

Hybrid Modeling and Advanced AI

The field is moving away from purely empirical "black box" models toward hybrid approaches.

  • Mechanistic-Data Fusion: Hybrid models combine mechanistic biological descriptions (such as the STICS crop model) with data-driven machine learning. This approach preserves physical/biological consistency while leveraging the predictive power of AI.
  • Temporal and Sequential Analysis: Architectures like CNN-LSTM are employed to analyze temporal sequences, such as satellite imagery time-series or animal physiological data (e.g., heart rates).
  • Reinforcement Learning (RL): RL is increasingly used to optimize complex decision-making processes in uncertain agricultural environments.

Robotics and Autonomous Navigation

While automation has seen significant success in structured environments—such as milking robots in dairy farms—the "open field" presents unique challenges.

  • Traversability-Aware Navigation: Modern agricultural robots must navigate unstructured terrain. Using LiDAR and computer vision, these systems are evolving to perform "traversability-aware" navigation, distinguishing between obstacles that must be avoided and those (like tall weeds) that can be traversed.
  • Swarm Robotics: There is an emerging trend toward "flocks" of small, simple robots working in coordination. This approach helps mitigate soil compaction, a major issue caused by heavy traditional machinery.

Human-Centric Decision Support

Digital tools are designed to augment, not replace, human expertise.

  • Interpretable AI: For farmers to trust algorithmic recommendations, models must be transparent. The paper advocates for local interpretability tools like SHAP or LIME to explain specific predictions, such as disease outbreaks or livestock estrus.
  • Decision Support Systems (DSS): The success of these technologies relies on effective UI/UX design that respects the farmer's domain knowledge and provides actionable "restitutions" of complex data.

Key Results and Future Directions

Significant progress has been made in livestock monitoring and remote sensing. AI-driven analysis of sensors can now detect estrus in dairy cows with high precision, even in "silent" cases. Furthermore, satellite constellations like the European Space Agency’s Copernicus program (Sentinel-1 and Sentinel-2) provide high-resolution global data for monitoring nitrogen and water stress.

However, several hurdles remain:

  • Connectivity: Many agricultural areas are "white zones" with no signal, necessitating the development of "data wells"—mobile sinks like drones that collect data from isolated sensors.
  • Data Labeling: High-quality labeled data is scarce and expensive, driving the need for semi-supervised learning research.
  • Ethics and Sovereignty: The massive collection of farm data raises critical questions regarding data ownership and the potential for a digital divide between large-scale and small-scale operations.

Source: Agritrop Cirad (https://agritrop.cirad.fr/601192/)

Summary updated

Abstract

Source Agritrop Cirad (https://agritrop.cirad.fr/601192/)
BibTeX
@inproceedings{mitton2022fondements,
  author = {Nathalie Mitton and Ludovic Brossard and Tassadit Bouadi and Frédérick Garcia and Romain Gautron and Nadine Hilgert and Dino Ienco and Christine Largouët and Évelyne Lutton and Véronique Masson and Roger Martin‐Clouaire and Marie-Laure Mugnier and Pascal Neveu and Pierre‐Marie Preux and Hélène Raynal and Catherine Roussey and Alexandre Termier and Véronique Bellon Maurel},
  title = {{Fondements et état de l’art}},
  booktitle = {Agritrop},
  year = {2022},
}
Figure