School of Civil and Environmental Engineering

Ph.D. Thesis Defense Announcement

SEEDING NUTRIENT MANAGEMENT SUSTAINABILITY THROUGH DIGITAL AGRICULTURE

By Abigail R. Cohen

Advisor:

Dr. Yongsheng Chen

Committee Members:  Dr. Xing Xie (CEE), Dr. Katherine Graham (CEE), Dr. Ameet Pinto (CEE), Dr. Rhuanito Soranz Ferrarezi (University of Georgia)

Date and Time:  August 12, 2025. 2:00 – 4:00 PM EST

Location: ES&T 3229, or on Zoom

ABSTRACT
Human activity dominates earth’s biogeochemical cycles, with agriculture
prevailing as one of the most consequential aggregate of processes on the planet.
As the basis for supporting life and all human activity, agriculture provides a critical
opportunity to mitigate some of these consequences through nutrient management
technology. While improvements to sensing, imaging, and computation can help
precision fertilization, phenotyping bottlenecks and a lack of unified modeling
approaches prevent scaling of precision nutrient management approaches. This
work provides a comprehensive critical review of technologies for nutrient
management in controlled environment agriculture (CEA), synthesizing literature
into a modular “system of compartments” framework called dynamically controlled
environment agriculture (DCEA).Recognizing bottlenecks and the need for modular, non-destructive nutrient
management models, we conducted a pilot-scale multimodal greenhouse
experiment using hyperspectral imaging (HSI) and multispectral imaging (MSI) to
expand DCEA. The experiment tested the limits of nutrient removal efficiency,
demonstrating significantly lower nutrient concentrations (50% of recommended)
could yield lettuce with over 75% fresh weight compared to the control, with healthy
tissue nitrogen concentrations. Next, the study provided a proof-of-concept using
Vision Transformer (ViT) deep learning (DL) architecture for regression analysis of
raw HSI data, bypassing the lengthy preprocessing pipelines typical of HSI-DL
approaches and facilitating end-to-end analysis.
Finally, a novel application of autoencoders (AE), a lightweight, adaptable DL
model, was deployed for anomalous growth detection using the in vivo MSI data.
Additionally, the study included an innovative energy use comparison of models
spanning the efficiency-accuracy spectrum (random forest and ViT), directly
addressing the environmental impacts of nutrient management automation and the
application of automation in resource-constrained edge environments. Results
revealed a 2% reduction in nitrogen application could completely offset the energy
use of the most complex DL model evaluated. Together, the work provides a
roadmap for resource-aware nutrient management modeling for precision
agriculture to facilitate environmentally sustainable food production.