Contained here are jpegs containing coastal RGB satellite images along with a water vs. land mask. Each image is 256 pixels by 256 pixels. Geographic scope: Delmarva (USA), Virginia Beach (USA), New Jersey (USA), Long Island (USA), Duck, NC (USA), Northern Tuscany Littoral Cell (Italy), Torrey Pines, CA, (USA), Narrrabeen Beach (Australia), Truc Vert (France) Temporal range: 1984 to 2022 Satellites: Landsat 5, 7, 8 and Sentinel-2 All images were downloaded from Google Earth Engine using CoastSat download tools. The datasets are arranged into 'train', 'val', and 'test' folders. Within each of those folders are two folders 'a' and 'b'. 'a' contains the images (RGB), whereas 'b' contains the labels (land vs. water mask). All images were augmented with the four following augmentations: horizontal flip, vertical flip, 90 degree clockwise rotation, 90 degree counterclockwise rotation, and a horizontal+vertical flip. For training a new segmentation model, it is advised to not do any of these rotational or flip augmentations since they have already been performed. Instead, possibly experiment with other augmentations like introducing noise into the imagery. These images were used to train an image-to-image translation generative adversarial network. The code and model weights (generator and discriminator) are available at https://github.com/mlundine/Shoreline_Extraction_GAN. To get to the files locally, you can download the .zip from Zenodo and then unzip the .zip file.