Coastal Land-Cover Data Derived from Landsat Collection 2 Data, Northern Chandeleur Islands, Louisiana

For each image acquisition date, Landsat Collection 2 ARD top-of-atmosphere (TOA) reflectance (reflective bands) and TOA brightness temperature (BT; thermal infrared [TIR] bands) were downloaded from EarthExplorer (https://earthexplorer.usgs.gov/) and Level 2 surface reflectance-derived NDVI images were downloaded from the EROS ESPA On Demand Interface (https://espa.cr.usgs.gov/).

All images were batch-processed using Spatial Model Editor in ERDAS IMAGINE. The TOA and BT bands were stacked to create 9-band multispectral images and clipped to the study-area extent. From these composite images, two additional spectral indices, mNDWI (Xu, 2006) and NBLI (Li and others, 2017), were calculated.

Land-cover classification was performed using progressive, automatic thresholding of spectral indices. First, Otsu's method (Otsu, 1979) for automatic histogram thresholding was applied to mNDWI images to create a binary land-water raster for each image acquisition date. Second, water area was masked from NBLI and NDVI images, and Otsu's method was applied to the masked images to create binary bare earth-"unclassed" and vegetated-"unclassed" rasters, respectively. Because thresholding was applied to masked NBLI and NDVI simultaneously, some pixels along the sand-vegetation boundary were classed as both sand and vegetated in this step. These pixels were mapped as sand using the following rule: if NBLI = sand and NDVI = vegetated, then final = sand (based on visual analysis during workflow development) (Bernier and others, 2021). Next, water, bare earth (sand), and vegetated areas were masked from mNDWI images, and Otsu's method was applied to the masked mNDWI image to separate the "submerged" and "emergent" intertidal areas. Finally, the binary land-cover images were converted to thematic rasters, merged, and single-pixel "clumps" were removed using a 3x3 majority filter to create a final land-cover raster dataset. The resulting land-cover rasters use the naming convention YYYYMMDD_LC##_nchan_islp_lcr_ce.img, where YYYYMMDD denotes the image-acquisition date (4-digit year, 2-digit month, 2-digit day) and LC## denotes Landsat 8 (LC08) or Landsat 9 (LC09) image source. All steps were batch-processed using the Image Processing toolbox in MATLAB (Otsu thresholding and binary image creation) or Spatial Model Editor in ERDAS IMAGINE (spectral index masking and generation of classed land-cover rasters).

For 5 datasets, thresholding resulted in significant misclassification of back-barrier seagrass beds as vegetated. These were identified based on visual comparison with ancillary datasets and vegetation persistence maps, and misclassed extents were manually cleaned from the final datasets by selecting and re-classing the pixels in ERDAS IMAGINE (Bernier and others, 2021). The cleaned land-cover rasters use the naming convention YYYYMMDD_LC##_nchan_islp_lcr_ce_cl.img, where YYYYMMDD denotes the image-acquisition date (4-digit year, 2-digit month, 2-digit day) and LC## denotes Landsat 8 (LC08) or Landsat 9 (LC09) image source.

The histogram (pixel count) for each raster was exported to a text file using Spatial Model Editor in ERDAS IMAGINE, merged into a single file in MATLAB, and used to calculate the total area for each land-cover type based on the relationship 1 pixel = 30 meters x 30 meters.