Advanced Machine Learning Techniques for Tidal Marsh Classification: A Random Forest Approach using Sentinel-2A

Abstract

Tidal marshes play a vital role in coastal ecosystems, functioning in climate change mitigation, water filtration, and protection from coastal erosion. However, mapping and monitoring of these ecosystems is often hampered by difficult accessibility and dynamic environmental conditions. This study aims to improve tidal marsh classification accuracy by applying a Random Forest (RF) algorithm supported by Sentinel-2A satellite imagery. This image provides various spectral parameters and vegetation indices, including B1, GNDVI, BSI, and NDWI. Three RF models with varying parameters were tested to determine their effectiveness in tidal marsh classification. The model with 26 parameters (Model 3) performed best, with the lowest RMSE value of 0.22, the highest AUC of 0.87, and the highest overall accuracy of 95%. These results show that combining critical spectral parameters in the RF model can significantly improve the classification accuracy and biomass estimation in tidal marshes. This study also confirmed the effectiveness of Random Forest in addressing the challenges of high-accuracy mapping and monitoring. These findings provide a solid foundation for tidal marsh ecosystem conservation and management applications and support the application of machine learning in coastal ecosystem mapping for better and more accurate results.