Conference Agenda

This study explores the use of Global Navigation Satellite System Interferometric Reflectometry (GNSS-IR) to monitor vegetation growth and water status in a grass-covered agricultural field via proximal sensing. It investigates whether GNSS reflected signals, analyzed through signal-to-noise ratio (SNR) measurements, can reliably indicate canopy structure and dielectric properties.

Data were collected with a GNSS antenna in an experimental meadow in Mondovì (NW Italy), using GPS and GLONASS constellations. GNSS-IR exploits interference between direct and reflected signals to estimate the effective reflector height, corresponding to vegetation phase centre height above the soil. Tracking its temporal variations allows inference of crop growth dynamics.

Amplitude and phase changes in the interference pattern are sensitive to surface dielectric properties, influenced by soil and plant water content. Thus, GNSS-IR provides both vegetation height and moisture information without destructive sampling or invasive sensors.

Vegetation height and soil moisture derived from GNSS-IR were compared with ground measurements using rulers and buried soil sensors. Preliminary results show strong agreement for canopy height and potential for monitoring water status. These findings highlight GNSS-IR as a promising, low-complexity tool for precision agriculture, leveraging existing GNSS infrastructure to obtain key biophysical parameters efficiently.

Purpose of the work

Although precision agriculture and smart farming technologies have advanced rapidly, their large-scale adoption has been slower than expected. This study examines the prerequisites for the large-scale deployment of agricultural robotics in Finland and identifies the key factors influencing adoption.

Description of the methods used

The study is based on a feasibility analysis assessing technological readiness, operational requirements, and ecosystem conditions for implementing robotic solutions in agriculture. Structural trends in the sector, together with potential barriers and enabling factors, were analysed.

Results

The results indicate that major barriers include the lack of high-quality reference data, the need to better align farming practices with digital technologies, and unresolved economic and organizational issues. At the same time, labour shortages, an ageing farming population, rising production costs, and tightening environmental requirements are increasing pressure for automation, particularly in labour-intensive sectors such as dairy, horticulture, and specialty crops.

Conclusions

Adoption barriers are primarily socio-technical rather than purely technological. Reliable digital infrastructure, hardware suited to northern conditions, and Living Lab approaches integrating development, testing, and user engagement are key to accelerating the adoption of agricultural robotics.

The Digital Product Passport (DPP) is emerging as a key instrument in the European data ecosystems, intended to support transparency and traceability across value chains. This work presents a proof of concept for an animal level Digital Product Passport for cattle, designed to enable greenhouse gas (GHG) reporting by combining public and private data using existing digital data sources and data space standards.

The development followed a Design Science Research methodology, using iterative cycles of problem explication, artefact design, demonstration, and evaluation. Requirements were derived from GHG monitoring needs and an analysis of Finnish dairy farm data streams, and machine-readable semantic data models and digital twin templates for core cattle attributes were designed. Cross-organisational data sharing was demonstrated within a minimum viable data space.

The proof-of-concept demonstrates feasibility and identifies key technical and non-technical barriers to implementation. Digital twin technology provides a promising, extensible structure for animal level DPPs, while challenges remain in data standardisation, governance, and cross-system interoperability. Effective data space governance and access control mechanisms are essential to ensure farmer trust and regulatory acceptance.

Authors received financial support from the European Union’s Horizon Europe Coordination and Support Action under grant agreement no. 101134866 (project Digi4Live, https://digi4live.eu/).

This work addresses automatic yield monitoring in row-structured orchards where GNSS positioning is unreliable due to canopy cover. We present a GNSS-free autonomous fruit mapping system based on known-coordinate markers and a depth camera. The objective is to generate georeferenced measurements of fruit number and size over time, enabling automatic tree-level yield monitoring.

At time_zero, a reference 3D model of an apple orchard was generated using LiDAR-SLAM and constrained by topographic surveys of 15 markers installed on poles, all referenced in the EPSG:32632 coordinate system. Subsequent acquisitions (time_n) are performed using a depth camera mounted on a rover optimised for orchard mobility. The automatic processing pipeline integrates deep-learning–based apple detection with marker recognition. Image-detected apples are projected in 3D using stereo depth, while the camera trajectory is estimated through visual odometry. The markers act as control points to compensate for drift and georeference detections without onboard GNSS. Georeferenced apples from time_n are mapped onto the time_zero model within an interrogable GIS environment.

Validation on eight markers showed an RMSE of 0.08m. The system becomes operational from the first visible marker (with fixed heading direction), and can be integrated into robots for agricultural operations, enabling GNSS-free automatic mapping and monitoring.

This work presents the experimental integration of a GNSS-independent orchard mapping system on AgriMaRo.Q, a reconfigurable omnidirectional rover designed for precision agriculture. The robot features a triangular chassis supported by three swerve-drive units and a motorized track-adjustment mechanism, enabling continuous variation of track width while maintaining compactness and stability. This configuration ensures high omnidirectional maneuverability, reduced soil compaction, and adaptability to crop environments such as orchards.

The rover was equipped with a ZED 2 stereo camera and an onboard Jetson Orin NX for real-time acquisition and processing. The perception pipeline combines a deep-learning–based apple detector with stereo depth reconstruction to obtain 3D fruit positions in the camera frame, alongside ArUco marker detection for pose estimation. Because stereo visual odometry is affected by cumulative drift, a GNSS-free correction strategy was implemented using georeferenced ArUco tags as spatial anchors. When a marker is detected, its measured pose is compared with its surveyed global coordinates, and the resulting translation offset is applied to subsequent detections to maintain global consistency.

Field experiments conducted in a orchard demonstrate stable mobility, reliable temporal alignment between detections and poses, and effective drift mitigation, enabling coherent tree-level fruit mapping in GNSS-denied conditions.

This study presents a Small Language Model (SLM) developed to support local smart farming by enabling systematic and context‑aware agricultural knowledge acquisition. To achieve practical deployment in resource‑constrained local environments, an open‑weight 2B‑parameter base model with strong Japanese performance was selected. The model was adapted through Continual Pre‑training (CPT) using agricultural extension exam datasets, complemented by subsets of the base model’s pre‑training and pseudo‑data to mitigate catastrophic forgetting while achieving effective domain specialization. Retrieval‑Augmented Generation (RAG) was additionally integrated to incorporate external agricultural information beyond the model’s internal knowledge. Evaluation experiments assessed GPU memory usage, inference latency, domain adaptation, and preservation of general knowledge. CPT with mixed data improved domain performance while reducing forgetting compared with CPT using domain data alone. RAG provided further performance gains, and the combined CPT-RAG model achieved the best results, demonstrating complementary benefits. All models exhibited sub‑second inference latency and memory usage comparable to the base model, indicating feasibility for deployment directly within local farming environments. These findings demonstrate that combining CPT with mixed data and RAG enables compact SLMs to support knowledge‑intensive local smart farming while maintaining general language capability.