AI-Driven Machine Vision Frameworks for Ornamental Plant Nursery Inventory Management and Disease Phenotyping in Peach Orchards

Abstract

Precision agriculture increasingly depends on automated sensing and machine learning to overcome the labor intensity, subjectivity, and scale limitations of manual field measurement. This thesis develops AI-driven machine vision frameworks for two underserved domains: ornamental nursery inventory management and foliar disease phenotyping in commercial peach orchards. Across both, the central challenge is the same: deploying vision systems that are robust to severe inter-plant occlusion, scarce labeled data, and the gap between general-purpose benchmarks and production agricultural environments.

The first study presents KBTrack, a tracking and counting framework for ornamental nursery plants. It combines an ensemble of transfer learning-adapted, CutMix-augmented YOLOv11x-seg segmentation models with a ByteTrack-based identity inference module that recovers plant identities after prolonged occlusion by reasoning over a sliding observation queue. The work formalizes the distinction between trajectory continuity and identity stability, the property required for accurate counting, and introduces Corrected ID Precision to evaluate it. Within a georeferenced cloud architecture linked to UAV orthomosaics, KBTrack reached a detection mAP@50 of 0.982 and a counting accuracy of 0.987 (RMSE = 4.188), reducing identity switches by 53% compared with the strongest baseline.

The second study extends this approach to three-dimensional phenotyping by estimating per-plant volume from RGB video captured during a single bed traversal. The pipeline integrates monocular pose estimation, KBTrack tracking, and a training-free multi-view extension of SAM 3D Objects that fuses per-frame diffusion predictions through confidence-weighted multi-diffusion fusion. On 670 azalea plants across six commercial beds, it achieved a mean absolute error of 1,209 cm³ and an R² of 0.93, outperforming visual SLAM, 3D Gaussian splatting, stereo point cloud fusion, and visual odometry baselines.

The third study develops a neurosymbolic framework for predicting tree-level foliar bacterial spot severity from near-infrared hyperspectral imagery (440–900 nm) acquired by a ground robot. It couples a band-attention encoder with spectrally anchored concept neurons, attention-based multiple instance learning that eliminates the need for pixel-level labels, a temporal attention module spanning four acquisition dates, and five differentiable symbolic constraints with a semi-supervised consistency term. Under leave-one-tree-out cross-validation, the framework achieved an RMSE of 1.45, an R² of 0.32, and a Spearman correlation coefficient (ρ) of 0.76 on the 0–9 severity scale, providing performance comparable to standard baselines while remaining interpretable under severe label scarcity.

Together, these studies demonstrate that purpose-built, domain-aware computer vision frameworks can provide scalable and interpretable agronomic measurements in settings where automation has historically struggled. They further show that embedding domain knowledge into machine learning models is essential for achieving robust performance when labeled data are limited.