Semester of Graduation

Spring 2026

Degree

Master of Science (MS)

Department

Department of Geography & Anthropology

Document Type

Thesis

Abstract

Coastal wetlands provide essential ecosystem services such as storm surge protection, flood control, carbon storage, and biodiversity habitat, yet face growing threats from sea-level rise, reduced sediment supply, land subsidence, and human impacts. This thesis introduces an integrated framework that combines remote sensing and time-series modeling to analyze, reconstruct, and forecast coastal land area changes over four decades (1984-2024) across four regions of coastal Louisiana: the lower Mississippi River Delta (MRD), the Atchafalaya River Basin, Vermilion Bay, and the Rockefeller Wildlife Refuge.

Landsat imagery was classified using a Random Forest model to generate annual and monthly land-area time series. To analyze long-term trends and seasonal patterns, we used Generalized Additive Models (GAMs), the Mann-Kendall test, the Theil-Sen slope estimates, and the breakpoint analysis. Monthly land-area anomalies were then evaluated in relation to hydro-environmental drivers, including river discharge, water level, precipitation, temperature, pore-water salinity, suspended sediment concentration, and storm activity. Results showed notable spatial differences: the Atchafalaya River Basin consistently gained land due to sediment input, whereas Vermilion Bay and Rockefeller Wildlife Refuge continued to lose land. The lower MRD demonstrated intricate, nonlinear patterns with periods of both gain and loss. Variations in land-area anomalies were more often linked to storm effects lagging by 0-3 months than to individual storm events.

To handle missing data caused by cloud cover and sensor limitations, we compared Kalman-based structural state-space modeling with Gaussian Process Regression (GPR) through cross-validation. The Kalman approach reduced reconstruction errors by up to 36% as compared to GPR and better maintained seasonal variability. Using the reconstructed time series, we evaluated seven forecasting models: Naïve, ARIMA, Prophet, Auto Neural Network Autoregression (Auto-NNAR), Tuned NNAR, XGBoost, and Ensemble. Machine learning approaches, particularly Auto-NNAR, consistently outperformed traditional statistical models in capturing nonlinear dynamics and structural breaks. Short-term (5-year) and exploratory long-term (10-year) forecasts show different trends: the Atchafalaya Basin continues to expand, Vermilion Bay gradually loses land, the lower MRD has highly variable changes, and Rockefeller presents a more complex pattern with marked short-term fluctuations. Uncertainty increases significantly over longer time horizons, underscoring the need for adaptive management that focuses on near-term forecasts.

This framework provides a scalable, transferable approach for monitoring and forecasting coastal dynamics in environments with limited satellite data, enabling better decisions for wetland restoration and coastal resilience planning.

Date

4-13-2026

Committee Chair

Meng, Xuelian

LSU Acknowledgement

1

LSU Accessibility Acknowledgment

1

Download

Available for download on Sunday, March 25, 2029

Share

COinS