Coastal Image Labeler

Project Overview

Documentation Last Commit DOI Actions Status1 Actions Status2

Generally, the Coastal Image Labeler presents users with coastal images to be labeled using a given set of questions. We designed the labeler to be hosted on a virtual machine and exposed via a web address. Labels provided by the user are written to a database and exported later by a user or admin.

There are two general roles in this project:

  • An Admin — who uploads images to be labeled, developing questions for the labeler to ask, assigning sets of images to users, managing the VM that hosts the project, and exporting data from the database.
  • A Labeler — who navigates to the website, logs on to the server, labels images, and can download their individual labels.

The Coastal Image Labeler Documentation is focused on:

Project Goals

  • The Coastal Image Labeler is designed to collaboratively label coastal images and then provide these labeled images as open data (FAIR) for general community use.

Some FAQs:

  • Why labeling images?

    • Labeled images are important for supervised machine learning research. There are many well known labeled image databases (e.g., ImageNet), but these existing databases tend to focus on general features (e.g., cats, dogs, horses, etc.). Our goal with this project is to develop a discipline-specific database of labeled images that is relevant for coastal scientists.
  • Why not use an existing tool for labeling?

    • Many good labeling tools already exist, but our goal with this project is to create a tool for collaborative, asynchronous labeling. Additionally, we wanted a tool to easily accommodate multiple users labeling a single image (to ensure correct labeling via consensus).
  • Isn't this similar to iCOAST from the USGS?

    • Yes — it is definitely one of the inspirations for this project. The USGS iCOAST project is an example of a labeled coastal database for storm impacts that was labeled collaboratively. We are extending this idea in at least 2 ways: First, the Coastal Image Labeler does currently host NOAA post-storm images (we have released some data already), but any image can be loaded and any question set can be created (for example, we have already labeled wave-scarp interaction images, and images of beach state). Second, this project is very closely tied to machine learning — crowdsourcing labels for coastal images to advance ML applications to Coastal science. One clear example of this linkage is — for some image catalogs — the images a user labels are shown to them in a specific order to help the ML algorithm learn samples that are confusing (i.e., active learning).