Annotated images of dreissenid mussels (Dreissena spp.) in down-looking color images in Lake Michigan, USA, to support machine learning modeling

This dataset is comprised of lakebed photographs collected by an autonomous underwater vehicle (AUV) in Lake Michigan, USA. Photos were annotated to identify invasive dreissenid mussels (Zebra and Quagga mussels) for use in machine learning and algorithm development, specifically to train and validate a transformer-based segmentation model (MusselFinder) for automated mussel density estimation. The dataset includes imagery subsets from specific survey locations, including waters near Muskegon, MI, and Elk Rapids, MI, and a sample from a broader area of Lake Michigan. Data files consist of original down-looking color images in WebP format, pixel-wise binary segmentation masks in PNG format (where mussels are represented by designated pixel values and the background by black pixels), and point-label annotations stored in JSON format. Approximate georeference locations and other data associated with these images are accessible in the accompanying metadata spreadsheet.