Mapping coastal wetland changes from 1985 to 2022 in the US Atlantic and Gulf Coasts using Landsat time series and national wetland inventories

The areal extent of coastal wetlands is declining rapidly worldwide, and scientists and land managers need land cover maps that show the magnitude and severity of changes over time to assess impacts and develop effective conservation strategies. Within the United States (US), widely-used, continental-scale wetland land cover data products are either static in time (The National Wetlands Inventory) or have a course temporal resolution and do not distinguish between different types of change (the NOAA Coastal Change Analysis Program, C-CAP). This study presents a new coastal wetland geospatial data product that leverages the Landsat database and maps annual land cover across the US Atlantic and Gulf Coasts from 1985 to 2022. The algorithm was trained on the existing US wetland inventories to make the final maps compatible with products that are used in operational management. A multi-stage classification approach was designed that uses the Continuous Change Detection and Classification (CCDC) algorithm to characterize time series of remote sensing reflectance with fitted harmonic functions and identify when changes likely occurred. The fitted time series models are then input into a random forest classifier to make a class prediction. An annual-scale random forest classification is performed in parallel, and results from both algorithms are combined and analysed to detect both gradual and abrupt changes and to identify transitional time series segments. A time series smoothing procedure is subsequently applied to ensure class transitions are logical and consistent and extract a summative change characterization map that shows the severity and spatial density of change. The final maps distinguish between four homogenous classes and six mixed classes, representing areas that are transitioning between classes and where the boundaries between classes are unstable. The algorithm uses data and tools within the Google Earth Engine platform, making it accessible and scalable. The average overall accuracy is 93.7%, and the average class omission and commission errors are 6.7% and 6.4%, respectively. A variety of change detection comparisons were performed, using the existing wetland inventory that employed a fundamentally different change detection approach, and a more comparable annual-scale, Landsatderived product that estimated changes across the Northeastern Atlantic Coast. These comparisons show that the new products’ severe change magnitude matches that of the