Long term wetland subclass extraction in Bangladesh using Google Earth Engine and sample migration

Wetlands are vital ecosystems threatened by climate change and human activities. However, their long-term changes in Bangladesh remain insufficiently understood, especially at the level of wetland subclasses. Here we leverage Landsat satellite imagery from 2000 to 2024 within the Google Earth Engine platform, combined with a sample migration strategy and a random forest machine model, to map wetland subclasses (flooded flats, mangroves, and marshes) and analyze their spatiotemporal dynamics. The method performed well in validation, achieving overall classification accuracies ranging from 82.49% to 89.26%, with low levels of disagreement between mapped and reference data, indicating stable and reliable classification performance. The results show that over the past two decades, wetland extent in Bangladesh has displayed considerable spatiotemporal variability and experienced an overall decline. This approach enables consistent long-term monitoring of wetland dynamics and provides a useful basis for wetland management and conservation in Bangladesh, with potential applicability to other regions. Wetlands in Bangladesh declined and shifted over two decades, revealed through long term subclass mapping using satellite images and machine learning classification, providing a basis for improved conservation.