Conference Agenda

Accurate and scalable estimation of broccoli head size across diverse varieties and field conditions is essential for precision agriculture and harvest management, yet existing approaches are largely cultivar-specific and sensitive to leaf occlusion and imaging illumination variability. This study presents a generalizable UAV-based pipeline for multi-variety broccoli phenotyping under varying illumination and occlusion scenarios. UAV RGB imagery is preprocessed into georeferenced tiles to ensure consistent model input across different flight conditions. Head detection and segmentation are performed end-to-end using a YOLOv26 instance segmentation model, trained with a stepwise annotation strategy to handle varying growth stages. Temporal information across multiple observation dates is incorporated to improve robustness against occlusion-induced appearance changes. Where leaf occlusion partially conceals the head, a circle fitting method is applied to the visible contour to recover reliable diameter estimates. The overall pipeline adopts a modular, configuration-driven design, allowing adaptation to new varieties, flight parameters, and field layouts with minimal effort. This work establishes a scalable and reproducible foundation for UAV-based broccoli head monitoring deployable across diverse agronomic and environmental conditions.

Asynchronous broccoli maturation within the same field often requires multiple manual inspections for multistage harvest scheduling. To reduce labor and improve planning accuracy, this study proposes a drone- and deep learning–based harvest planning model. First, a multi-stage broccoli head image dataset was constructed from fields at different harvesting rounds. A fused algorithm pipeline was proposed. YOLO11n was used to localize broccoli heads in drone images, providing bounding boxes that served as prompts for SAM2 to extract accurate masks. Circle fitting using least squares and minimum enclosing circle methods was then applied to estimate head diameters. Field maturity density was calculated as the number of broccoli heads ≥12 cm in diameter per square meter to determine harvesting eligibility. For harvestable fields, the harvesting sequence was further optimized by evaluating the unit-time harvest profit. Experimental results showed that the proposed pipeline achieved a broccoli head detection accuracy of 0.979, a mask segmentation IoU of 0.846, and an average diameter estimation error of 4.35%, meeting practical application requirements. The case study further confirmed that the key indicators generated by the proposed model effectively supported broccoli harvest scheduling. This harvest planning model may serve as a reference for developing harvest planning methods for other crops.

Weeds are a major constraint to stable soybean (Glycine max) production. Dense planting has been proposed to suppress weeds by reducing cumulative light; however, quantitative evaluation of sub-canopy light under field conditions remains challenging. We developed a framework integrating unmanned aerial vehicle remote sensing (UAV-RS) and ray-tracing simulation to estimate photosynthetically active radiation (PAR) available to weeds at high spatio-temporal resolution.

A field experiment was conducted in Tokyo, Japan, from June to August 2025 using seven cultivars at two planting densities. RGB images were acquired by UAV-RS and processed into 3D point clouds via Structure-from-Motion. The point clouds were converted into simulation-ready models using a voxel representation with surface meshes, and compared with a grid-based method. Solar position was computed from time and geographic coordinates, irradiance was measured in the field, and sub-canopy radiation was simulated using the Helios ray-tracing software. Instantaneous shading and daily accumulated PAR were evaluated and validated with quantum sensors and photosensitive films.

The voxel-based approach outperformed the grid-based method. Daily accumulated PAR was estimated more stable than instantaneous shading, with relative error below 15%. These results demonstrate the potential of high-resolution PAR estimation for optimizing planting density and supporting low-input weed management.

While real-time monitoring of crop growth is essential for optimizing nitrogen and density management in potatoes, traditional methods often fail to capture the complex G×E×M interactions driving biomass accumulation. Accurate, high-throughput forecasting remains a challenge for precision agrotechnology. In this study, multi-year field experiments encompassing various N application rates and planting densities were conducted to acquire high-fidelity canopy spectral data and key agronomic parameters. First, we developed MST-Net, a physics-informed dual-stream architecture integrating Transformer and ResNet modules, which achieved high-fidelity LAI inversion by leveraging synthetic datasets from the PROSAIL model to mitigate canopy background interference and sample scarcity. Integrating the inverted LAI time-series, we subsequently constructed a Light Interception model coupled with a dynamic biomass allocation framework to elucidate the mechanics of DM production. Crucially, this integrated approach successfully quantified the impacts of field management regimes on source-sink dynamics. We mathematically characterized the continuous biomass partitioning among individual organs, revealing how specific N and D interactions modulate canopy light capture efficiency and alter organ-specific accumulation patterns. By bridging remote sensing inversion with dynamic physiological modeling, it offers a transformative tool for the multi-objective optimization of nitrogen and density regimes, paving the way for autonomous, data-driven decision-making in sustainable potato intensification.

Artificial supports are necessary in modern orchards but cause severe perception errors for robotic systems. Additionally, a lack of high-quality 3D simulation data hinders agricultural Embodied AI development. This study addresses these challenges by developing a high-fidelity digital twin generation pipeline that integrates 3D Gaussian Splatting with a novel unsupervised 3D segmentation approach. Data were collected using a handheld 3D scanner in Japanese peach and US cherry orchards, capturing diverse canopy geometries and support structures. To disentangle complex scenes, we propose a hierarchical segmentation strategy. First, the ground is removed by cloth simulation filter. To distinguish artificial structures from biological elements, a coarse-to-fine framework explicitly addresses the trade-off between sensitivity and specificity. The pipeline progresses from high-recall region extraction to strict multi-modal feature fusion that isolates high-confidence structural cores. Subsequently, a model-guided optimization leverages these geometric anchors to reconstruct support poles and trellis wires. Individual trees are then reconstructed using a bottom-up assembly strategy based on Euclidean clustering and direction-aware graph clustering. Evaluation demonstrates high accuracy, with F₁-scores of 99.58% for ground, 95.78% for poles, 90.78% for wires, and 90.74% for trees. This work establishes a scalable foundation for simulation-driven training and evaluation of robotic systems in complex agricultural environments.

Seed potato physiological age is important for crop establishment, yet it is of-ten assessed using destructive end-point traits such as sprout mass, which may not reflect sprouting dynamics or structural differences. We propose a non-destructive, time-resolved 3D pipeline that extracts sprout volume curves and eye-level traits, and applies dynamic growth modeling to quantify sprouting patterns. Tubers with contrasting physiological ages from different storage-temperatures were repeatedly imaged during sprouting. 3D point clouds were reconstructed by structure-from-motion. Sprouts were segmented on 2D imag-es using a deep-learning-based model, and mapped to 3D to obtain sprout point clouds. Sprouts were separated using density-based clustering. Sprout volume was estimated by PCA-based slicing, alpha-shape area, and integration. For each tuber, we calculated sprout volume over time, sprouted-eye counts, and sprouts per eye, then fitted growth curves to estimate growth rates. Mean sprouted eyes per tuber were 3.9 (Low), 4.1 (High-Short), and 4.0 (High-Long), and mean sprout tips per eye were 2.7, 3.0, and 3.4, respectively, with F1 of 0.901. Dynamic models differed among treatments, and were most sensitive to 8-16 day windows, with weaker effects for longer windows. This approach en-ables discrimination of sprouting patterns with physiological age and supports storage optimization and physiological-age assessment.