Image segmentation and classification labelling for geospatial imagery
Information from Azavea
Azavea, a company that applies AI to earth observation imagery for geospatial imagery challenges for civic, social, and environmental applications, has developed its own imagery segmentation and classification tools. GroundWork is a tool that creates training data for machine learning models run on geospatial imagery.
The company has been using it internally for six months already, and is now opening access to everyone looking to label geospatial imagery, for free.
The tool is designed specifically for geospatial imagery, and enables users to retain any spatial context encoded within the image. Other tools limit labels to squares, circles or lines. GroundWork’s image segmentation however allows users to draw whatever they need to (e.g. buildings, fields, trees, shores) following natural or man-made shapes.
To create a project, determine the project type (object detection, semantic segmentation or image/chip classification, then upload or drop in the image URL for the image. The application converts the tiff or GeoJSON image to a cloud-optimised GeoTIFF (COG) and breaks it up into tasks, which are 512×512 pixels each. Each task can be labelled individually according to classes you choose. On export the tool produces a STAC-compliant export. The company has also partnered with CloudFactory to provide users with high quality labels.
The free version of GroundWork offers ten projects and 10 GB of storage per user. The Pro version, which is in the works, will add further functionality such as validating already labelled tasks; visualising additional map layers to aid in labelling (including analyses like NDVI or NDWI); creating labelling campaigns; and managing teams.