Semester of Graduation

Spring 2026

Degree

Master of Science (MS)

Department

Department of Geography and Anthropology

Document Type

Thesis

Abstract

Soil salinity is a critical environmental issue in coastal regions, particularly in Louisiana's Coastal Belt, where saltwater intrusion, driven by rising sea levels and storm surges, disrupts freshwater ecosystems and threatens salt-sensitive vegetation. This research is designed in two main chapters (Chapter 2 and Chapter 3). In Chapter 2, we evaluated multiple linear, machine learning, deep learning models integrating Sentinel-2 images to select the best approach for assessing soil salinity dynamics across four seasons (Spring, Summer, Fall, and Winter) in 2023, during an extreme drought event. In Chapter 3, we evaluated the short-term trend of wetland seasonal salinity dynamics from 2020 to 2024 using the optimal model developed in Chapter 2. For chapter 2, initially six models: Support Vector Machine (SVM), Random Forest Classifier (RF), Artificial Neural Networks (ANN), Extreme Gradient Boosting (XGB), Partial least squares regression (PLSR), and Convolutional Neural Network (CNN) 1D, were employed to detect salinity and assess predictive accuracy. Each model was evaluated using R2 and RMSE on the training and test data. Across all seasons, initially ANN, XGB, and SVM performed consistently better, with Test R2 values ranging from 0.65 to 0.78, indicating that these models explained a portion of the variability in the data. To further improve the base model, we employed a weighted-average ensemble, and the best-performing model was the weighted average of ANN, XGB, and SVM. However, the most stable model, with consistently lower RMSE, was the weighted-average ensemble of ANN and SVM. So, for cleaner estimation of wetland salinity and map development, we used a weighted average ensemble of ANN and SVM. The ensemble model's accuracy Test R2 ranged from 0.68 to 0.80 across 4 seasons. The final predicted salinity maps across all seasons show that low- to moderate-salinity classes dominate inland regions (Salinity PPT 0 to 18 parts per thousand (PPT)), whereas higher salinity levels are concentrated in coastal and estuarine zones with 18 to 32 PPT. The results demonstrate the effectiveness of Sentinel-2–derived indices combined with robust machine-learning models for regional soil salinity mapping, while highlighting the importance of careful model selection and validation in data-limited environments.

After raster prediction, we further performed SHapley Additive exPlanations (SHAP) values to evaluate each variable's contribution to the model. SHAP explains how the prediction values change with respect to the high and low values of each variable. The SHAP results demonstrate that vegetation indices consistently remain the most influential predictors across seasons. Because lower (negative) values of vegetation indices helped the model interpret the stress signals in vegetation reflectance, and higher (positive) values of vegetation indices revealed non-saline zones. Besides, salinity-specific and reflectance-based indices exhibit season-dependent importance, confirming the necessity of integrating multiple indices for robust seasonal coastal wetland salinity prediction. The sensitivity results confirm that vegetation indices dominate model performance in spring, summer, and fall, indicating vegetation stress signals, and help distinguish higher-vegetated areas, where salinity values tend to be lower, from exposed wetlands. On the other hand, salinity-specific indices become relatively more influential during winter, emphasizing the importance of season-dependent predictor contributions in coastal wetland salinity modeling.

After model sensitivity evaluation, we used the most stable model, a weighted average ensemble of (ANN and SVM), for spatiotemporal (2020 to 2024) wetland salinity prediction. For short-term seasonal trend analysis, we first evaluated the characteristics of the CRMS data. If the data quality shows a larger difference between mean and median and greater variability in higher- and lower-salinity values, the model might fail and produce incorrect results. After evaluating the data characteristics, we selected 14 seasons from 2020 to 2024 for the spatiotemporal assessment of soil salinity. We found test accuracy results ranging from 0.60% to 0.80% (R2) across different seasons. The summer and fall seasons demonstrated higher accuracy due to higher, clearer spectral reflectance. During winter and spring, the data becomes noisier, and prediction accuracy drops. After predicting all rasters for all seasons, a parish-wise trend analysis showed that in summer, salinity intensity and variability are substantially higher than in other seasons. Across all parishes, medians increase progressively from 2020 to 2024, with the most pronounced expansion observed in Plaquemines, St. Bernard, and Terrebonne. The upper whiskers extend beyond 25–30 PPT in some years, reflecting extreme values and strong positive skewness. The interquartile ranges widen notably in 2023 and 2024, indicating greater heterogeneity.

Finally, how different land-use land-cover classes were affected by different salinity values was evaluated from 2020 to 2024 across seasons. This research has some limitations, such as Sentinel-2 images having higher spatial resolution (10); we could not find cloud-free images during summer and Fall 2021, despite having usable salinity data from the CRMS sites. In such cases, Landsat images (30m) could be more useful, but they would lack the precision of Sentinel-2. Another limitation is the data from CRMS sites. Storms and other natural events sometimes damage CRMS sites, and if there is not much variation in soil pore water salinity data, the model might fail to predict accurately. But in such cases, developing traditional interpolation maps could provide some insights. However, this research takes the first step toward developing high-resolution AI-based salinity maps for coastal Louisiana. A few modifications or statistical evaluation of the spectral indices could help better understand the characteristics of the entire training dataset, and SI7 could be a better alternative to BI2 or Int2 for understanding real salinity physics. Though this research sought to incorporate indices from all groups, including salinity, brightness, intensity, and vegetation, to ensure sufficient transferability across the whole coast of Louisiana and other study areas where exposed salt crusts or other exposed features on the ground may occur. But a few cases are also evident in which indices from the BI group or the Int group can sometimes mislead in accurate salinity estimation. Because built-up areas tend to appear brighter. Also, at the end of this research, the final pipeline demonstrates a simplified yet advanced approach for accurately and rapidly mapping wetland salinity dynamics. The model is useful for assessing situations where CRMS site salinity values show a significant increase and for analyzing the spatial patterns of soil salinity. It can support policymakers in natural resource management by identifying suitable locations for shrimp or crawfish farming and guiding other management decisions. The spatiotemporal trends from 2020 to 2024 would help planners, policymakers, and wetland management authorities understand salinity hotspots over the past five years and take the necessary steps.

Date

4-13-2026

Committee Chair

Meng, Xuelian

LSU Acknowledgement

1

LSU Accessibility Acknowledgment

1

Download

Available for download on Friday, April 13, 2029

Share

COinS