Conference Agenda

Accurate leaf-level 3D segmentation is essential for high-throughput plant phenotyping, yet existing methods often depend on depth sensors or extensive manual annotation, limiting scalability. This work presents PlantLiftSeg, a fully automated and annotation-free pipeline for 3D leaf instance segmentation using RGB-only imagery. Multi-view RGB images are captured by a custom vision-based autonomous UAV designed for GPS-denied greenhouse environments. Per-plant 3D Gaussian Splatting (3DGS) models are reconstructed from the captured images. The core pipeline, text-prompted 2D mask generation via SAM3, and contrastive instance feature learning on 3DGS, is applied sequentially: first with a plant-level prompt to isolate each plant from background clutter, then with a leaf-level prompt segments individual leaves. Feature coherence weighting is introduced for robust 2D-to-3D label assignment, while boundary scale loss and 3D spatial regularization improve instance separation. Evaluated on manually labeled point clouds of 110 leaves from five greenhouse-grown muskmelon plants, PlantLiftSeg achieves a mean IoU of 0.94, AP@50 and AR@50 of 1.00, and AP@50-95 and AR@50-95 of 0.94 and 0.99, respectively, demonstrating accurate and scalable 3D leaf segmentation in controlled greenhouse settings. The RGB-only and annotation-free framework reduces hardware cost and labeling burden, supporting practical deployment in controlled-environment agriculture.

Monitoring beehive foraging activity is essential for colony health and sustainable apiculture. Traditional methods, such as manual bee counting, are labor-intensive and susceptible to noise, colony size variation, and environmental fluctuations. This study proposes an IoT-enabled smart beehive system for continuous monitoring and short-term forecasting of foraging activity. Foraging behavior was characterized using two indicators: (1) incoming pollen-carrying bee count, representing harvest magnitude, and (2) pollen collection rate, defined as the proportion of incoming bees carrying visible pollen loads. The system was deployed on four hives and operated continuously for over one year, collecting hourly bee activity and environmental data that were transmitted to a cloud platform for storage. Both indicators were forecast 24 hours ahead using three time-series models: Gated Recurrent Units (GRU), Temporal Convolutional Networks (TCN), and a hybrid TCN-GRU model. Model performance was evaluated across different beehives and seasons. Results show that pollen collection rate prediction achieved an average Mean Absolute Error (MAE) below 0.06, while the incoming pollen-carrying bee count achieved a normalized MAE below 20%. The hybrid TCN-GRU model demonstrated more stable performance than individual models. The proposed monitoring and forecasting framework provides reliable short-term forecasting of colony foraging dynamics, supporting data-driven smart beekeeping management.

Insect pest monitoring using sticky traps is widely adopted in integrated pest management; however, many existing systems require dedicated hardware, limiting scalable deployment in diverse agricultural environments. This study proposes a deep learning framework for tiny insect recognition from mobile phone images captured under real field conditions. The framework consists of three stages. First, a pretrained CLIP encoder extracts visual embeddings, which are combined with an SVM classifier to form an anomaly detection module that distinguishes sticky trap images from irrelevant inputs. Second, image quality screening, perspective correction, and illumination normalization are applied to enhance robustness under varying field conditions. Third, insect detection is performed using a YOLOv11 model with sliding window inference using SAHI, achieving a mAP@0.5 of 0.93 for densely distributed small insects. Detected regions are subsequently classified using a lightweight ResNet18 model. To improve discrimination between visually similar categories, an additional dust class and object size features are incorporated, enhancing thrips recognition while maintaining stable performance for gnats and whiteflies. Field evaluations demonstrate F1 scores of 0.90, 0.96, and 0.94 for thrips, gnats, and whiteflies, respectively. The proposed framework integrates foundation models with task-specific design for scalable mobile pest recognition in the field.

In recent years, agricultural digital twins have been increasingly demanded as a form of agricultural simulation in digital spaces based on real-world data. This technology requires a digital shadow that incorporates real plants into digital spaces as realistic three-dimensional models and analyzes biological information through 3D simulations. Here, the circadian clock is present in nearly all cells. It synchronizes its phase with neighboring cells and exhibits phase responses to changes in the external environment. Therefore, a three-dimensional circadian clock simulation for digital plants—whose appearance and geometry are reconstructed from real-world data and in which circadian clocks are implemented at the cellular level—is a fundamental technology for agricultural digital twins. In this study, we constructed a photorealistic autonomous point-cloud plant model by rendering realistic plant 3D point-cloud models using Unreal Engine and autonomously computing the temporal evolution of circadian clocks at each point. Using this model, we visualized local phase propagations and conducted local light stimulus simulations. Furthermore, by exporting time-series data of circadian clock phases in CSV format, we preserved the results of 3D simulations in digital space as numerical data. This study presents the first technology to visualize circadian phase dynamics along the realistic morphology of plants.

Plant factories utilizing artificial and solar light have become an established platform for advanced agriculture worldwide. Meanwhile, rapid advances in generative AI, digital twins, XR technologies, and humanoid robotics are transforming plant production research. In this study, we propose SPA-XR Robotics, a new research framework that integrates the Speaking Plant Approach (SPA) with photorealistic digital twins, XR technologies, and advanced robotics.

The proposed framework reconstructs SPA within a photorealistic digital twin environment built using Unreal Engine 5 (UE5). Plant and cultivation environments are represented using 3D Gaussian Splatting (3DGS), enabling realistic reproduction of crop structures and scenes in digital space. XR interfaces with hand-tracking interaction allow intuitive visualization of plant physiological states and human interaction within VR/AR/MR environments. Collaborative and humanoid robots are introduced to reproduce complex agricultural tasks such as harvesting and plant handling within the digital twin system.

By integrating plant physiological models including circadian clock dynamics with photorealistic digital twins and robotics, this research expands SPA and contributes to a new scientific framework for plant production systems.

The demographic structure of fish populations is fundamental for stock assessment and sustainable management, particularly for long-lived, high-value species such as southern bluefin tuna (Thunnus maccoyii). Reliable age determination is essential for constructing this profile. Conventionally, SBT ages are determined from otolith sections through expert interpretation, a process that is labor-intensive and subjective. This study proposes a multimodal deep learning framework for automated age estimation using otolith images and associated biological metadata. The dataset comprises 4,394 transverse otolith section images with corresponding otolith weight, fork length, and dressed weight. A pretrained ViT-Large backbone was employed for image feature extraction, while biological variables were encoded via a multilayer perceptron and fused at the feature level prior to classification. Model performance was evaluated against expert-guided readings using exact accuracy, ±1-year accuracy, and biological validation through von Bertalanffy growth modeling. The multimodal model achieved a ±1-year accuracy of 84.1% and a strong correlation with expert estimates (R² = 0.96). Growth curves derived from model-predicted ages were statistically indistinguishable from expert-derived curves (χ² = 1.448, p = 0.694), demonstrating its suitability for supporting large-scale fisheries monitoring and stock assessment.

This presentation reviews our research on resource regeneration technologies for circular food production in space and on Earth. A major challenge in closed and resource-limited food production systems is the recovery of plant-available nutrients from organic residues. To address this challenge, we have developed an integrated approach combining anaerobic digestion, biological oxidation, membrane filtration, and nutrient supplementation to generate regenerated liquid fertilizer for hydroponic cultivation. This technological framework links waste treatment and crop production by converting organic residues into reusable nutrient solutions. Our research has shown that digestate-derived liquids can be transformed into fertilizers compatible with hydroponic systems and that supplementation of deficient nutrients enables plant cultivation. These results support the feasibility of waste-derived fertilizer production as a strategy for closing material loops in food production systems. By introducing this research framework and its main outcomes, the presentation discusses its significance for bioregenerative life support systems in space and for sustainable circular agriculture on Earth.