Machine learning offers the promise of rapidly increasing the ability of scientists and researchers to identify and predict phenomena. It is no secret that a key requirement of any machine learning training task is labeled data… lots and lots of labeled data. As the phenomena to be predicted become increasingly complex and discrete, the task requires not only labelling but also manually denoting bounding boxes that direct the machine to the portion of a larger image which depicts the labeled phenomenon. IMPACT's machine learning team has developed Image Labeler, a web-based tool that allows users to create tagged images that can be used to train different machine learning models for predicting Earth science phenomena.
IMPACT's machine learning team focused on Earth science phenomena as the scarcity of quality training data for Earth science is well known. Image Labeler allows you to create and manage labeled Earth science images for machine learning. With Image Labeler, you can label images to indicate the presence or absence of a target Earth science phenomenon. Additional features include: extracting images from satellite imagery, drawing bounding boxes, and assigning teams for labeling.
Prasanna Koirala, a member of the machine learning team, points out that:
"Image Labeler allows teams to efficiently create new training sets. Following that logic, the images don’t need to be of Earth science origin. A user can upload images of any type and create a team with members that are responsible for tagging the images such as 'yes-cat' or 'no-cat'."
In its current release, the Image Labeler supports GeoTIFFS and Shapefiles along with WMS. It can use shapefiles to extract images as well as look at the images as a shapefile in the map. Additionally the Image Labeler offers API endpoints and supports colored shapefiles that allow users to tag different events. Learn more about the latest release of the Image Labeler tool.
Check out Prasanna's LinkedIn profile.
