Image-Based Precision Phenotyping of Plant and Ecosystem Productivity Using Deep Learning and Vision Foundation Models
Year: 2026
Type: dissertation
Abstract: The growing need for sustainable agriculture and reliable assessment of plant and ecosystem productivity requires scalable phenotyping approaches to address global challenges in food security, environmental management, and climate change mitigation. Traditional methods based on manual field surveys and visual assessments are labor-intensive, subjective, and often unable to provide consistent quantitative information. This dissertation advances image-based precision phenotyping by integrating machine learning (ML), deep learning (DL), and vision foundation models (VFM) across four complementary investigations spanning vegetation composition assessment to ecosystem-scale carbon flux prediction. First, deep semantic segmentation-based DL models were evaluated for roadside vegetation composition and compared with UAV-based multispectral vegetation index methods, demonstrating superior species-level performance using DL-based RGB imagery. Second, a smartphone-based application was developed for rapid blueberry detection and yield estimation, integrating DL and ML strategies to provide an accessible field-deployable solution for plant breeders and growers. Third, DL models and VFM derived feature embeddings were integrated to improve blueberry yield prediction across diverse genotypes and growing conditions. By capturing detailed plant structural and morphological characteristics, VFM embeddings enabled a deeper understanding of plant architecture and visual traits associated with yield potential, ultimately leading to more reliable yield predictions. Finally, VFM were applied to ecosystem-scale productivity assessment by predicting net ecosystem exchange and gross primary production in a tidal salt marsh using image-derived representations and meteorological data. Collectively, this research demonstrates the potential of VFMs to improve image-based phenotyping by enhancing prediction accuracy, reducing dependence on extensive annotated datasets, and enabling scalable applications across vegetation assessment, crop yield prediction, plant phenotyping, and ecosystem carbon flux estimation.
Author: Singh, Puranjit
Institution: University of Delaware
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: green
OpenAlex ID: W7223248408
Cites:
Cited by:
Related to: