Conference Agenda

Obtaining high-resolution phenotypic data at the individual level is critical for precision breeding, yet unstructured fields and severe occlusion often compro-mise traditional monitoring methods. This study develops an autonomous robotic phenotyping system for robust spatial-temporal tracking of tomato growth. The platform utilizes a six-wheel drive system guided by a vision-based lane detection model. Optimized by a Boundary-aware Loss function, the model sharply defines boundaries and improves navigation. For reproducible spatial-temporal individual plants tracking, a localization algorithm utilizing lateral optical flow triggers Real-Time Kinematic (RTK) logging to establish a localized coordinate system. To ad-dress fruit occlusion, the system employs a semi-automated annotation pipeline to train an AISFormer model for amodal instance segmentation, inferring complete contours of occluded fruits. Experimental results demonstrate that the boundary-aware loss function improves mIoU by 4.32% and reduces the average lateral offset from 7.09 cm to 4.62 cm, while the RTK positioning achieved an average accuracy of 4.44 cm. Furthermore, the AISFormer model achieved a precision of 0.84 and a recall of 0.75 in highly occluded fruit detection. This study presents a robust and scalable solution for high-throughput phenotyping, improving both robotic navigational stability and the accuracy of individual-level yield estimation in complex agricultural settings.

Accurate plant counting during transplanting and early growth is critical for yield estimation, replanting decisions, and operation assessment, yet field videos are often degraded by camera shake, illumination changes, and partial occlusions. This study proposes an AI-based tracking-by-detection framework for automatic seedling and cabbage counting from transplanting-stage and growing-stage videos for reliable detection and counting. An improved YOLOv8 detector was developed for seedling and cabbage detection. To enhance multi-scale feature representation and suppress background noise, a convolutional block attention module (CBAM) was inserted after C2f blocks in the neck, using global max/average pooling to generate attention maps. For counting, detections are linked into consistent track IDs using BoT-SORT algorithm, and trajectories within a counting ROI are converted to final counts with automatic logging/visualization. Proposed detector achieved good performance (mAP 95.7%) and showed high correlation between measured and estimated counts (R² = 0.947). Tracker comparison showed a clear speed–accuracy across algorithms, with BoT-SORT reduced misses at lower FPS and remained robust to camera shake, illumination changes, and short occlusions. Error analysis on orthomosaic/field imagery revealed residual failures from rendering artifacts, false positives, and duplicate detections in repetitive patterns. Practical video-based counting enables tracker selection that balances real-time speed and counting accuracy.

The plot-level spatial heterogeneity in crop growth and phenology limits the accuracy of remote sensing-based yield prediction. To address this challenge, we proposed a quantitative framework to assess wheat growth uniformity and phenological uniformity (PU) and evaluate their contribution to predictive yield. A two-year field experiment was conducted with 210 wheat cultivars using UAV-based hyperspectral and RGB imagery. Growth uniformity indices (UIs) were derived from fractional vegetation cover, leaf area index, SPAD, and canopy height, while PU was assessed via deep learning-based phenological classification. Results showed that the Pielou index from LAI at flowering had stronger correlations with yield (r=-0.760) than traditional mean values. The EMS-GCN model achieved 86.2% accuracy for phenological stage classification. Integrating UIs and PU significantly improved prediction accuracy of yield (R²=0.708, RMSE=1.084 t/ha), outperforming models using only spectral features. PU was positively correlated with prediction accuracy, with optimal performance at PU=1. This framework provides a novel approach for crop monitoring and precision agriculture by incorporating uniformity metrics into remote sensing-based prediction models.

Individual, non-invasive animal monitoring in commercial dairy barns is essential for precision livestock farming and animal-level digital twins, but robust identity tracking remains challenging due to occlusions, high visual similarity among animals, and strong day–night illumination changes. This work presents a practical computer-vision pipeline that integrates real-time cow detection, identity-preserving multi-object tracking, and lightweight posture inference to generate continuous, per-cow state information under realistic barn conditions. The approach builds on a YOLOv11n detector fine-tuned on barn-specific imagery and a ByteTrack-based association stage to maintain stable identities through short-term occlusions. Our experimental dairy farm show that the system detects nearly nine out of ten cows in the scene while keeping false alarms low and localizing animals reliably. Also the pipeline tracked ten individual cows and produced time series of positions and postures , enabling welfare indicators such as lying time and activity patterns. Beyond welfare analytics, these individual-level trajectories provide a valuable linkage layer for multimodal sensing, where vision-derived animal location and behavior can be combined with near-floor gas measurements to support individual-animal emission tracking and analysis . This contribution outlines framework, key results, and next steps toward scalable, sensor-fusion-ready monitoring for climate and animal well fair research

Spatial distribution within the canopy, initialized at the seedling stage, dictates inter-plant competition and overall productivity. However, the impact of spatial uniformity on crop development remains under-explored due to a lack of precise phenotyping techniques. To address this, we developed a high-throughput UAV-based pipeline to quantify uniformity. The impact of such uniformity was evaluated on rice growth across a three-year field experiment of 168 plots with diverse uniformity gradients. Our approach integrates deep learning for individual plant identification with Voronoi tessellation to precisely localize plants and characterize spatial arrangement from UAV imagery after transplanting. Results confirmed a strong link between uniformity and cumulative canopy light interception (r = 0.71–0.82) following rapid canopy expansion. Consequently, higher uniformity significantly accelerated canopy closure and advanced flowering time. This drove significant increases in panicle density (r = 0.65–0.69), ultimately improving final yield (r = 0.39–0.44). Path analysis further revealed that these productivity gains are mediated by accelerated development and enhanced light interception. Here, we provide the first high-throughput quantification of uniformity at the seedling stage and a comprehensive evaluation of its impacts. This not only provides a useful tool but also establishes a mechanistic basis for optimizing rice yield through precise transplanting management.

Egg productivity is a key indicator of laying hen performance. In commercial cascade cage systems, cage-level egg production is currently counted manually, which is inefficient and labor-intensive. Existing computer vision methods rely on one camera per egg collection belt (ECB), leading to high equipment costs. This study developed a low-cost computer vision system to achieve accurate cage-level egg counting with fewer devices. A customized video capture scheme was designed to cover multiple ECBs, and a CV-based framework was proposed including the lightweight YOLO-Egg detector, an observation-centric tracking algorithm, and an egg counting and positioning method. Test results showed that YOLO-Egg achieved 96.1% detection accuracy at 132.1fps, and the overall counting accuracy reached 98.89%. Field validation on commercial farms yielded 93.13% average cage-level counting accuracy at 66.23fps, with processing time only 37.75% of the original video. This system enables automatic, real-time and precise cage-level egg monitoring, supporting intelligent precision poultry farming.