Semester of Graduation

Spring 2026

Degree

Master of Science (MS)

Department

Geography and Anthropology

Document Type

Thesis

Abstract

Suspended sediment concentration (SSC) in the Lower Mississippi River is a major control on sediment delivery to coastal Louisiana, where approximately 4,833 km² of land has been lost since 1932. Effective sediment monitoring is therefore essential for understanding riverine sediment dynamics and informing coastal restoration planning. However, spatially continuous SSC estimation from satellite imagery remains challenging because coincident satellite and in-situ SSC observations are limited. To address this constraint, this thesis develops a two-stage pipeline for estimating SSC from 10 m Sentinel-2 imagery, using turbidity as an intermediate variable between surface reflectance and sediment concentration.

In the first stage, ACOLITE-corrected surface reflectance was related to in-situ turbidity using eight machine learning models trained on 70 satellite-turbidity matchups collected at Belle Chasse, Louisiana. In the second stage, predicted turbidity was converted to SSC using a site-specific power-law relationship calibrated from 102 paired field measurements. This design provides a physically interpretable approach to SSC retrieval while overcoming the scarcity of direct satellite-SSC matchups. A systematic comparison of atmospheric correction methods further showed that ACOLITE consistently outperformed Sen2Cor, with differences in reaching 0.39 for ensemble tree models.

Model performance was evaluated across multiple independent contexts. For end-to-end SSC estimation on 15 independent ACOLITE-processed scenes matched with field SSC observations within a ±2-day window, support vector regression (SVR) produced the highest performance, with an SSC of 0.881 and an RMSE of 14.19 mg/L. Transfer evaluation at Baton Rouge, approximately 200 km upstream, showed that the Belle Chasse-trained pipeline retained meaningful but reduced predictive ability at a new location, with a best SSC of 0.641 using ElasticNet, indicating partial geographic transferability but also the need for site-specific recalibration. When Belle Chasse and Baton Rouge data were combined for training , Ridge yielded the best performance across 31 independent scenes  of 0.799 and RMSE of 20.84 mg/L.

Overall, the results demonstrate that the proposed two-stage framework is a viable, reproducible, and operationally relevant method for estimating SSC from Sentinel-2 imagery, while also showing that model performance depends on atmospheric correction method, geographic setting, and evaluation context.

Date

3-18-2026

Committee Chair

Lei, Wang

LSU Acknowledgement

1

LSU Accessibility Acknowledgment

1

Download

Available for download on Saturday, March 18, 2028

Share

COinS