Robotic and AI-Based Systems for Inventory Management in Ornamental Nurseries and Biomass Estimation in Cover Crops

Abstract

Recent advancements in field robotics, machine vision, hyperspectral sensing, and deep learning have created new opportunities for automating agricultural monitoring tasks that have traditionally depended on manual labor. In this dissertation, four projects are presented to advance automated inventory management, plant quality assessment, and biomass prediction in specialty crop and sustainable agriculture systems. First, an automated plant counting and quality assessment system for container-grown ornamental nurseries is presented. Accurate inventory management is essential for monitoring sales, yield, space requirements, production timing, and crop quality; however, current practices rely heavily on manual counting, which is labor-intensive and often inaccurate. In this study, an unmanned ground vehicle, NurseryBot, was developed to collect downward-facing RGB video data from container nursery stock. A deep learning framework based on YOLO11 model variants was trained and tested using a panoramic image dataset for plant detection and instance segmentation. The YOLO11-m instance segmentation model achieved the highest performance for azalea detection and segmentation, with bounding-box precision of 0.971, recall of 0.970, and mAP50 of 0.941. For mask segmentation, the model achieved precision of 0.969, recall of 0.968, and mAP50 of 0.901. The proposed system achieved counting accuracies of 95.59% for azalea plants and 95.37% for sunshine plants. Image-derived morphological features, including area, perimeter, eccentricity, and circularity, were further used to support systematic plant quality assessment. These results demonstrate the potential of a robotic and deep learning-based system as a scalable alternative to manual nursery inventory practices. Second, an unsupervised domain adaptation approach for ornamental plant counting across species and growth stages is presented. Although deep learning-based segmentation methods have shown promise for automated plant counting, their dependence on repeated labeling for each plant class limits their scalability in commercial nurseries, where thousands of plants across wide varieties must be managed. To address this challenge, this study developed a domain adaptation framework that enables plant counting on previously unseen target classes with limited target-domain labeling. The model was designed to reduce the need for continuous annotation while improving generalization across plant growth variability, maturity stages, and species differences. The proposed approach reduced the miscount rate from more than 30% in non-adapted models to below 5% after domain adaptation. These findings show that unsupervised domain adaptation can substantially improve the scalability and adaptability of deep learning-based nursery inventory systems. Third, a hybrid robotic vision framework was developed to coordinate aerial and ground platforms for bed-level inventory assessment in ornamental nurseries, where manual row-by-row counting remains labor-intensive and inconsistent at commercial scale. Ground robots count accurately but must traverse every bed, while aerial imaging covers ground rapidly yet cannot resolve overlapping canopies. A DJI Mavic 3M UAV acquired ultra-low-altitude RGB video over azalea beds, and sampled frames were processed to estimate bed-level counts across 14 beds using three tracking-based approaches. KBTrack + Ensemble achieved the lowest bed-level error (mean absolute error 30.21 plants, mean absolute percentage error 11.76%, aggregate accuracy 98.36%, whereas BoT-SORT + Ensemble minimized the aggregate total but incurred larger bed-level error, revealing that block totals can mask errors that cancel across beds. An oversized-mask rule, flagging canopy masks exceeding 1.5 times the local median area, identified beds with likely canopy merging for targeted ground inspection. Operating both platforms concurrently reduced total time from 971 to 800 seconds, a 17.6% saving over a sequential ground-only baseline. This coordinated aerial-ground approach offers an efficient, scalable pathway for nursery inventory assessment and demonstrates the broader potential of hybrid robotics for agricultural automation. Finally, a hyperspectral imaging and machine learning framework was developed for non-destructive cover crop biomass prediction. Accurate aboveground biomass estimation is central to Kura clover (Trifolium ambiguum M. Bieb.) evaluation, where biomass governs forage value, living-mulch performance, and genotype selection in breeding programs. Conventional clipping and drying are direct but laborious, destructive, and hard to scale. Ground-based snapshot hyperspectral imaging was used to estimate dry biomass from Spring and Fall acquisitions in spring and fall 2025. Records from both dates were pooled and evaluated on a representative 30% held-out partition. Strict Spring-to-Fall transfer was retained as a secondary test of domain shift. Calibrated 440.9 to 898.5 nm cubes were converted into plot-level spectral and canopy features. These included standard normal variate correction, Savitzky-Golay first-derivative spectra, canopy descriptors, vegetation indices, and canopy-integrated terms. A Canopy-Integrated Stacked Ensemble gave the strongest result, with measured and predicted biomass agreeing at R² = 0.68, RMSE = 2.64 g, and Spearman rho = 0.82. Integrated near-infrared reflectance, plant-pixel counts, canopy cover, and plant area were the dominant predictors. These results show that Kura clover biomass can be estimated non-destructively when both seasonal canopy states are represented in calibration.