Theses and Dissertations
ORCID
https://orcid.org/0000-0002-3143-1159
Advisor
Souza Martins, Vitor
Committee Member
Borges Ferreira, Lucas
Committee Member
Gharakhani, Hussein
Committee Member
Samiappan, Sathishkumar
Committee Member
Zhang, Xin
Date of Degree
5-15-2026
Original embargo terms
Embargo 1 year
Document Type
Dissertation - Open Access
Major
Engineering (Biosystems Engineering)
Degree Name
Doctor of Philosophy (Ph.D.)
College
James Worth Bagley College of Engineering
Department
Department of Agricultural and Biological Engineering
Abstract
Very high spatial resolution (VHSR; i.e., < 5-m) land cover maps provide precise spatial inventories of the Earth's surface, including natural and anthropogenic features that are critical for understanding geospatial patterns and processes. Historically, the complexity inherent to VHSR optical imagery has posed analytical challenges for traditional land cover classification approaches that are frequently applied to coarser-resolution data. These challenges have largely been addressed by the emergence of deep learning-driven classifiers, but the large volume of ground truth data required for robust training of these classifiers remains a significant barrier to operational mapping applications, particularly where ground truth data is scarce or acquiring commercial VHSR imagery is cost-prohibitive. This dissertation explores the state of the art in deep learning-based land cover classification at VHSR and identifies wide gaps in data availability and methodological approaches. An ensemble of deep land cover classification models was pre-trained with self-supervised learning by utilizing “Bootstrap Your Own Latent” (BYOL) as a pretext task with 377,921 patches of unlabeled color-infrared aerial imagery, leading to the creation of a new 1-m land cover product for the state of Mississippi, USA, using only 1,000 labeled image patches. Validation of the resultant 8-class bitemporal land cover map shows overall accuracies of 87.14% for 2023 and 84.78% for 2016 using an independent assessment dataset of 25,000 locations. Finally, a flow matching-based super-resolution approach is developed to super-resolve the 10-m visible and near-infrared bands of Sentinel-2 imagery to 2.5-m spatial resolution, enabling access to temporally dense VHSR imagery that is otherwise cost-prohibitive to acquire at large scales. This technique has the novel capability of navigating the perception-distortion trade-off at inference time, enabling a single model to produce both spectrally reliable and visually realistic super-resolved outputs without the need for retraining. This super-resolution model is used to create a national-scale 2.5-m imagery product over the conterminous United States for 2025, as well as an annual land cover product for the Chesapeake Bay watershed for 2020–2025 at 2.5-m spatial resolution, demonstrating the potential of super-resolution approaches to enhance the spatial resolution of widely available satellite imagery for VHSR land cover classification.
Sponsorship (Optional)
MAFES SRI, Mississippi Space Grant Consortium Graduate Fellowship
Recommended Citation
Hester, Dakota, "Addressing data scarcity for land cover mapping at very high spatial resolution using deep learning" (2026). Theses and Dissertations. 6912.
https://scholarsjunction.msstate.edu/td/6912
