Conference Agenda

Purpose: GNSS spoofing and signal interference increasingly threaten agricultural operations, while standard GNSS without RTK cannot provide trustworthy positioning. At the same time, policy frameworks such as the CAP require verifiable proof of where, when and how field operations were performed. This work demonstrates how integrating Galileo’s High Accuracy Service (HAS) and Open Service Navigation Message Authentication (OSNMA) into agricultural workflows can deliver both reliable positioning and regulatory‑grade execution evidence.

Methods: We developed a demonstrator linking a task‑map provider, an FMIS (Farmmaps) and an ISOBUS terminal. OSNMA authenticates GNSS signals to detect spoofing, while HAS provides high‑accuracy positioning without local RTK. Signed task maps for spraying and fertilizing were generated in the FMIS and executed on the machine using authenticated HAS positioning.

Results: The system produced cryptographically signed as-applied maps that were returned to the FMIS, where every processing step was verifiable. This enabled end-to-end proof of execution for agricultural operations, combining authenticated positioning with high-accuracy GNSS.

Conclusions: The results demonstrate that Galileo HAS and OSNMA can provide a strong foundation for trustworthy positioning and for regulatory-grade evidence. In addition, the results indicate that this approach can be integrated into existing agricultural machinery and FMIS architectures with limited adaptations.

Representative Discrete Element Method (DEM) modelling of flaky particles for the design and optimization of handling equipment and machinery requires a structured geometric characterization that can capture wide size distributions and pronounced dimensional variability. This study presents an image-based framework for two-dimensional shape characterization and DEM particle dimensions generation of flaky particles. More than 15,000 woodchip particles were analyzed through contour extraction, convex hull processing, and Minimum Bounding Box computation. The bounding box defines the principal in-plane dimensions of each particle, providing consistent geometric descriptors for irregular shapes. Particles were systematically classified into primary groups using uniform 2 mm length-based bins according to their maximum in-plane length. Within each primary group, additional width-based subgrouping was applied using the same bin size to obtain statistically representative average geometric dimensions for each one. The extracted mean length and width values for each group were combined with experimentally measured thickness values to generate equivalent three-dimensional parallelepiped particles. This grouping strategy provides reduced-dimensional variability while preserving the population-level geometric characteristics required for DEM simulations. The proposed methodology establishes a structured pathway from image detection to statistically representative particle descriptors for DEM modelling of biomass particles with a large size distribution.

Vehicle overturning has been a persistent safety issue in agriculture worldwide. It not only threatens farmers’ safety but also hinders the advancement of farm automation. In Japan, the agricultural labor force has been declining significantly, making overturning prevention in agricultural off-road vehicles a critical challenge for improving farm safety and efficiency. During field operations, these vehicles often travel on undulating off-road terrain, such as unpaved roads, steep slopes, and inclined side paths. Such harsh conditions can induce dynamic instabilities, including jumping and slippage, which degrade vehicle performance and increase overturning risk. This study investigates the effect of jumping nonlinearity on vehicle instability through numerical analysis. MATLAB/Simulink® and the driving simulator software CarSim® are used as simulation platforms. Simulations of an agricultural vehicle model with jumping nonlinearity are conducted under various terrain conditions, including sinusoidal excitation, bumps, and slopes. Comparisons between the MATLAB/Simulink® model and CarSim® are performed to validate the appropriate use of the driving simulator. Dynamic instability is quantified using nonlinear time-series analysis, such as Poincaré sections and Lyapunov exponents, together with conventional Fourier spectral analysis. Results demonstrate that jumping nonlinearity induces severe dynamic instability, highlighting the necessity of appropriate control strategies to enhance vehicle stability.

Fires originating in cereal combine harvesters can lead to forest fires with severe consequences, both for the environment and due to the risk of loss of human life and infrastructure. The specific points on combine harvesters where fires typically start include the engine area, the cutter bar, electrical equipment, transmissions, and bearings. In many cases, the fire is triggered by contact between crop residues and high-temperature components of the machine. The ignition temperature of cereal residue is approximately 250 °C.

An in-cab monitoring system for combine harvesters has been developed and implemented. This system is capable of informing the operator in real-time about fire risks. The system features temperature probes located in critical areas of the machine, external temperature and relative humidity sensors, a real-time data storage and transmission system, an in-cab monitor, and a GPS antenna to geolocate all information.

This system was implemented in a cereal combine harvester during the 2025 harvest season, and its performance was validated under real working conditions throughout the entire campaign. Additionally, the relationships between the harvester's operating parameters, the yield map, and the data provided by the monitoring system were analyzed.

Weed management in woody perennial crops remains dependent on
broadcast herbicide applications, which pose environmental risks, cause crop
phytotoxicity, and increase production costs. Although robotic weed control has
advanced in open-field row crops, orchard environments introduce additional
challenges, including irregular geometries, trunk obstacles, canopy-induced illumination
heterogeneity, and degraded GNSS performance. This study presents
GreenBot, a modular autonomous robotic platform designed for selective undercanopy
weed spraying in orchards. The robotic platform integrates three modular
systems: perception, trunk avoidance, and precision spraying. The perception
system is based on stereo vision, LiDAR, GNSS-RTK, and IMU localization.
The perception system integrates an algorithm based on YOLOv8 for weed detection
and the extraction of weed relative coordinates. The trunk avoidance system
comprises a hybrid trunk-avoidance mechanism attached to a six-degree-offreedom
robotic arm, equipped with independently controlled flat-fan nozzles
that enable selective under-canopy weed spraying while avoiding trunk spray.
Field trials conducted in commercial almond orchards in southern Spain demonstrated
that the system avoided spraying the trunks. However, the detection model's
performance degraded substantially under heterogeneous canopy light conditions.
Results confirm the technical feasibility of modular systems for robotic
spraying in woody crops while highlighting the need for specific weed perception
training and datasets.

Reliable perception remains a critical bottleneck in autonomous robotic apple harvesting due to frequent occlusions from leaves, stems, branches, and neighboring fruit that directly affect grasp feasibility. While modern object detection models can localize apples accurately and assign occlusion classes, they rely primarily on geometric cues and exhibit non-negligible misclassification under complex orchard conditions. Full human verification ensures safety but severely limits throughput.

This study proposes a hybrid perception framework that integrates occlusion-aware apple detection with Vision–Language-Model reasoning for human-in-the-loop supervision during robotic apple harvesting. A transformer-based RF-DETR model performs occlusion-based apple detection, where a probabilistic calibration scheme identifies predictions with low reliability. These ambiguous cases are provided to VLM for semantic occlusion reasoning and selectively escalated to human supervision when uncertainty or disagreement between RF-DETR and VLM decision persists. Results demonstrated occlusion-aware detection mAP@50 of 86.2%, with calibration allowing 47% of detections to be directly accepted. VLM showed 74.2% classification accuracy on uncertain cases, highlighting promise for selective human-in-the-loop correction.