| dc.description.abstract | Specialty crop production, spanning ornamental nurseries and controlled-environment agriculture, faces rising labor costs, a shrinking workforce, and a growing need for scalable, data-driven management. Traditional crop monitoring, inventory, and field navigation remain labor-intensive and error-prone at scale. This thesis advances automation in specialty crop production through intelligent robotic systems and AI-driven analytics, presented as three application-driven studies that integrate robotics, deep learning, and sensor technologies.
The first study addresses plant inventory management in nurseries through a generative AI-based digital twin pipeline. Stable Diffusion XL with custom LoRA models generated realistic synthetic nursery images across 36 bed combinations, while Grounded DINO and the Segment Anything Model (SAM) produced over 0.77 million annotated plant instances without manual labeling. A YOLO11x-seg model trained on this synthetic data achieved 96.5% counting accuracy (MAE 7.7, RMSE 8.31) on real nursery beds, confirming viable synthetic-to-real transfer.
The second study delivers a field-validated autonomous navigation framework. STRIDE-Net, a hybrid perception module, combines YOLOv8 instance segmentation with PCA and Hough Transform refinement for bed-edge detection and centerline tracking. Benchmarked against classical vision methods, the YOLOv8n-seg variant achieved the best deployment balance, with smooth steering (FAD1 0.011 rad/s², FAD2 0.019 rad/s³) at roughly 12.9 FPS. Across 36 beds, intra-bed navigation reached a mean lateral deviation of 6.47 cm (SD 1.89 cm), inter-bed transitions a 1.49 cm waypoint error, and integrated runs a 10.16 cm mean absolute deviation, consistently within the 20 cm safety threshold.
The third study introduces a low-cost autonomous overhead imaging platform for greenhouses. A ceiling-mounted galvanized-conduit serpentine rail carries a 3D-printed sensor enclosure housing a ZED 2 RGB-D camera, a Living Optics hyperspectral camera, and environmental sensors. Deployed at Auburn University with lettuce, its RGB-D pipeline (SAM3 pseudo-labeling, YOLO26l-seg, ortho-strip compositing) reached a mask mAP@50 of 0.941 and detected 227 of 240 plants (94.6%). Mean equivalent diameter was 80.9 mm, projected area 5.95 × 10³ mm², and volume proxy 3.79 × 10⁵ mm³, offering a scalable alternative to commercial gantry systems. | en_US |
