Machine Learning to Understand the Importance of Factors Impacting Recent Marshland Loss in Coastal Louisiana

For the past three decades, the Landsat remote sensing data has been used to map changes in Louisiana’s coastal wetlands. In this project, machine learning techniques were employed to model region-wide datasets collected from Landsat imagery to accurately predict changes in the Normalized Difference Water Index (NDWI) from aerial (NAIP) imagery between the years 2019 and 2021, during which two major hurricanes hit the Barataria and Breton Sound basins. Python coding and the Scikit-Learn library were utilized to build a random forest (RF) model with the goal of achieving high prediction accuracy of land/water cover change in the Mississippi River Delta region. The RF model’s performance was assessed using the testing R^2 value, which resulted in a value of nearly 0.6. The variables that proved most significant in supporting the model’s accuracy, listed from most to least important, were the green vegetation index (NDVI), distance from the Hurricane Zeta and Ida tracks, and river sediment sources, and elevation. The importance values indicate which data set contributed the most to prediction accuracy, providing valuable insights for future data collection efforts to improve the RF predictions. To further improve the model in future iterations, incorporating additional maps for subsidence and hyperparameter tuning is being tested. Accurately predicting changes in wetlands in the Mississippi River Delta is crucial for understanding the impacts of both human activity and climate change on this vital ecosystem in the face of sea level rise and more intense tropical storms.