Conference Agenda

This study presents a holistic cyber-physical framework for precision agriculture, designed to overcome operational bottlenecks in high-value crop management. In response to climate variability and skilled labor shortages, it proposes a real-time Decision Support System (DSS) that transforms traditional offline geomatics into an on-device spatial computing ecosystem based on passthrough Augmented Reality (AR).

To overcome rural connectivity constraints, the system employs ultra-lightweight computer vision models executed directly on-device, including NMS-free object detection and zero-shot instance segmentation. Visual outputs are fused with IoT microclimatic time-series data through multimodal neural networks, enabling accurate crop analysis—from stress and phytopathology detection to yield estimation and structural assessment.

A key innovation addresses the spatial “Reality Gap” under dense canopies. An Agricultural Visual Positioning System (A-VPS) anchors the AR headset’s local 3D mesh to a pre-acquired Digital Twin using advanced Point Cloud Registration. Combined with GNSS-RTK data and processed through an Extended Kalman Filter (EKF) on the SE2(3) Lie group, this ensures sub-centimeter holographic stability.

Finally, through ISOBUS (ISO 11783) integration, the AR system acts as a Task Controller, converting augmented perception into automated mechatronic actions, bridging spatial AI and agricultural robotics within a scalable Agriculture 4.0 paradigm.

Quantitative analysis of operational loads is essential for improving the energy efficiency and performance of agricultural machinery. In a 103 kW-class round baler, the power take-off (PTO) is the primary energy source, and accurate prediction of PTO power requirements is critical for drivetrain and system optimization. However, empirical field testing is constrained by narrow harvesting windows and variable environmental conditions, limiting consistent data acquisition. Simulation-based modeling has thus emerged as an effective alternative for predicting dynamic loads under diverse operating conditions. Although previous studies using the Discrete Element Method (DEM) and Multi-Body Dynamics (MBD) have focused on localized phenomena such as crop flow and internal compression behavior, integrated prediction of total PTO power demand remains limited.
This study develops a high-fidelity DEM–MBD co-simulation framework to estimate PTO power requirements. The approach includes: (1) acquisition of real-time field load data using a torque sensor and DAQ system; (2) constitutive modeling of rice straw in DEM; (3) development of a dynamic mechanical model including pickup and plunger mechanisms in MBD; and (4) coupled simulation to evaluate crop–machine interactions. The model is validated by comparing simulated and measured torque profiles, establishing a reliable virtual platform for load prediction and digital twin implementation.

In Japan, contract farming of crops such as cabbage requires frequent monitoring for growth assessment, harvest timing, and yield forecasting; however, labor shortages make repeated manual monitoring increasingly difficult. Although drone surveys provide high-resolution spatial data, they require skilled operators, disrupt routine field operations, and involve hours of offline photogrammetric processing.

To overcome these challenges, this study presents an in-vehicle edge-computing sensing system that generates georeferenced field images in real time. The system integrates dual cameras with RTK-GNSS and an IMU to estimate vehicle pose and generate GeoTIFF raster images comparable to drone-derived orthorectified imagery without offline processing. In a prototype implementation, sensor streams at 24 Hz (camera), 10 Hz (RTK-GNSS), and 50 Hz (IMU) are processed concurrently, enabling georeferenced image generation at 2–5 Hz on a compact edge‑computing platform without GPU acceleration, executed in a parallel thread alongside data logging.

Field experiments were conducted in an actual cabbage plot using multiple ground markers distributed across the field, with marker heights adjusted to approximate canopy height. The system achieved an average positional error of 9.6 cm and an area estimation error of 7.9%, demonstrating sufficient accuracy for practical implementation.

Seed germination is a key physiological process that strongly influences crop establishment, uniformity, and yield potential. Accurate characterization of germination kinetics is therefore essential in seed science, plant breeding, and crop management. Conventional germination tests rely on manual counting and visual inspection, which are labor-intensive, time-consuming, and prone to observer variability. Recent advances in imaging technologies provide new opportunities for automated and non-destructive monitoring of germination dynamics.

This study presents the development and validation of a low-cost imaging system designed to quantify seedling germination kinetics through time-cumulative image analysis. The system integrates RGBs cameras, controlled illumination, and automated image acquisition to capture sequential images during the germination process. Image processing algorithms were implemented to detect and quantify germination events over time, enabling the construction of cumulative germination curves.

The proposed approach allows continuous, objective monitoring of germination while significantly reducing manual effort. Results demonstrate that the system provides reliable and reproducible measurements of germination dynamics, highlighting its potential as an accessible tool for high-throughput phenotyping, seed quality assessment, and plant breeding applications.

Plant classification plays an important role in many areas, including agriculture, environmental monitoring, and biodiversity conservation. Being able to correctly identify different plant species is especially valuable in precision farming, where accurate crop management can improve productivity and sustainability. With the rapid growth of deep learning, computer vision technologies have made remarkable progress, offering powerful tools for solving plant classification challenges.

In this study, we compared two advanced deep learning models a custom Convolutional Neural Network (CNN) and InceptionV3 both compared for plant classification tasks. Using a horticultural plant dataset, we trained and tested the two models under the same conditions to ensure a fair comparison. To evaluate their performance, we used key metrics such as accuracy, precision, recall, and F1-score. These measures helped us understand how well each model could correctly identify and categorize different plant species.

Our analysis highlights the strengths and weaknesses of both approaches, giving a clearer picture of how they perform in practical agricultural settings. The findings provide useful insights into which model is more suitable for building reliable, automated plant identification systems. Overall, this research contributes to improving plant recognition technologies and supports the broader goal of promoting more efficient and sustainable agricultural practices.

Autonomous navigation in agricultural inter-row environments remains challenging under sparse or absent vegetation conditions, where vineyard posts may represent the only reliable structural landmarks. Previous work demonstrated the feasibility of Global Navigation Satellite System (GNSS)-free Light Detection and Ranging (LiDAR)-based estimation of lateral offset and yaw using geometric regression on vineyard post returns. However, frame-by-frame geometry-based yaw estimation may exhibit short-term fluctuations caused by sparse LiDAR returns, partial occlusions, metallic reflections, and fitting noise. This paper presents an experimental comparison between geometry-derived yaw rate and Inertial Measurement Unit (IMU) angular velocity acquired from a co-located Livox Mid-360 sensor during GNSS-free vineyard inter-row navigation. The analysis is based on 18 acquisitions collected in a canopy-free vineyard across six 30 m inter-row paths, each repeated three times, for a total travelled distance of approximately 540 m. Results show that the geometry-derived yaw rate exhibits higher short-term variability than the IMU signal, with standard deviations of 0.003 rad s⁻¹ and 0.002 rad s⁻¹, respectively, and peak values of 0.007 rad s⁻¹ and 0.005 rad s⁻¹. These results demonstrate the complementarity of the LiDAR and IMU sensing and motivate future fusion-based approaches for more robust GNSS-free autonomous navigation in structured agricultural environments.