Earth observationWe annotate satellite images to train computer vision models at monitoring event and tracking traffic in remote places
Creating an advantage for extremely time-consuming annotation tasks
Earthcube develop monitoring solutions based on the automated analysis of geospatial information. Part of the data they process is satellite imagery, which requires very accurate annotation to train computer vision models at detecting objects such as airplanes, ships, and cars.
The geointelligence company struggled at speeding up their annotation pipeline with internal teams. What’s more, satellite images are so large and detailed that their annotation requires very specific attention and knowledge. Keeping annotation internally could not be an option anymore.
Ingedata assigned a team of photo interpretation specialists for the segmentation and classification of objects in satellite images. A rigorous feedback loop was installed and ensured further progress.
The last Client visit to Ingedata’s production center also bonded the team, which is now almost part of EarthCube’s employee count!
A specific workflow was designed to ensure that, despite the large size of the images, every pixel was analyzed. The quality control process also embeds specific tools to ensure that every part of the image is correct.
In addition, learning curve and further training allowed to refine the level of classification. The team has one technical referee for each main class recognition, and objects are now labeled down to an extremely precise level.
How do you make the difference between the roof of a truck and the roof a of prefab building?
By comparing identique locations across various points in time. The use of a Geographical Information System to produce the annotations was key to the success of the project. By being able to geolocate the satellite images, Ingedata were able to view the same geolocation across several timepoints. This allowed to compare object changes over time and confirm their identification.