Spatiotemporal Analysis of Phragmites australis in the Mississippi River Delta Through Deep Learning Techniques and Empirical Orthogonal Function Based on Time-Series Sentinel-2 Imagery

Author

Manisha K C, Louisiana State University and Agricultural and Mechanical CollegeFollow

Semester of Graduation

Fall 2025

Degree

Master of Science (MS)

Department

Geography & Anthropology

Document Type

Thesis

Abstract

Phragmites australis is the most dominant emergent plant in the freshwater and brackish marshes in the lower MRD, significantly contributing to the stability of the marsh ecosystem. However, the widespread dieback of Phragmites has caused serious concerns about wetland stability and resilience. Traditional mapping methods relying on single-date imagery and concurrent ground samples are costly, time-consuming, and often do not generalize across different years. This research addressed the challenge of robust mapping and understanding Phragmites dynamics using time-series Sentinel-2 imagery from 2019 to 2024. The study explored an integrated approach across two objectives: firstly, assessing the robustness and temporal generalization of deep learning models for Phragmites classification when trained using multi-year past training datasets without the use of concurrent ground-truth data, and secondly, understanding the Phragmites health dynamics using the NDVI indices and quantify the dominant spatial-temporal patterns of NDVI variability using Empirical Orthogonal Function (EOF) analysis and correlate its time series with different environmental drivers. A U-Net deep learning model classified the wetland into Phragmites, non-Phragmites, and water. The results for the first objective demonstrated that while single-year models performed well with concurrent training samples (90% to 98% overall accuracy), cross-year transferability model accuracy relied on the choice of training sample time and similarity for vegetation condition, achieving an overall accuracy above 80% for most of the input scenarios. However, integrating training datasets from multiple past years resulted in accuracies comparable to those of models trained on the concurrent samples. The overall accuracy obtained was above 90%, and Phragmites accuracy was above 80% for most of the years, showing the potential for using historical past datasets without the need for concurrent data. This deep learning classification approach facilitates model training based on historical data, minimizing dependence on expensive concurrent field surveys and enhancing long-term wetland monitoring efficiency. Furthermore, the results for the second objective revealed three distinct spatial patterns: the primary seasonal cycle controlled by temperature and precipitation (EOF1), interannual variations linked to hydrological factors (EOF2), and localized, short-term disturbance-recovery cycles linked with hydro-climatic factors (EOF3). By linking classification accuracy to disturbance history and decomposing complex variability into dominant modes, this research collectively established a transferable framework for long-term monitoring in other coastal areas that experience significant phenological variation and rapid change.

Date

11-3-2025

Recommended Citation

K C, Manisha, "Spatiotemporal Analysis of Phragmites australis in the Mississippi River Delta Through Deep Learning Techniques and Empirical Orthogonal Function Based on Time-Series Sentinel-2 Imagery" (2025). LSU Master's Theses. 6258.
https://repository.lsu.edu/gradschool_theses/6258

Committee Chair

Meng, Xuelian