Detecting Mowed Tidal Wetlands Using Time-Series NDVI and LSTM-Based Machine Learning

This study presents the first application of machine learning (ML) to detect and map mowed tidal wetlands in the Chesapeake Bay region of Maryland and Virginia, focusing on emergent estuarine intertidal (E2EM) wetlands. Monitoring human disturbances like mowing is essential because repeated mowing stresses wetland vegetation, reducing habitat quality and diminishing other ecological services wetlands provide, including shoreline stabilization and water filtration. Traditional field-based monitoring is labor-intensive and impractical for large-scale assessments. To address these challenges, this study utilized 2021 and 2022 Sentinel-2 satellite imagery and a time-series analysis of the Normalized Difference Vegetation Index (NDVI) to distinguish between mowed and unmowed (control) wetlands. A bidirectional Long Short-Term Memory (BiLSTM) neural network was created to predict NDVI patterns associated with mowing events, such as rapid decreases followed by slow vegetation regeneration. The training dataset comprised 204 field-verified and desktop-identified samples, accounting for under 0.002% of the research area’s herbaceous E2EM wetlands. The model obtained 97.5% accuracy on an internal test set and was verified at eight separate Chesapeake Bay locations, indicating its promising generality. This work demonstrates the potential of remote sensing and machine learning for scalable, automated monitoring of tidal wetland disturbances to aid in conservation, restoration, and resource management.