Intel® AI Lab, MILA, and CrowdAI have conducted a joint data-science collaboration to develop deep learning image-segmentation models that can be utilized to detect bridges in remote areas using satellite imagery. Working with the American Red Cross, Intel identified that it would be valuable to base this work on imagery of territory in Uganda, given the country’s historical suffering from both community epidemics and seasonal weather events.
To develop models for this region, Intel created a custom training dataset with CrowdAI, utilizing four band (RGB plus near-infrared) high-resolution satellite imagery as our inputs. Multiple deep-learning models were trained to segment images into road, waterway, bridge, and background classes.
Utilizing a custom evaluation method that is appropriate for the sparse bridge detection problem, the project selected a top-performing model with which to run inference and identify bridges across locations in Southern Uganda previously unseen to the model. Through our pipeline, the project was able to identify 70 new bridges.
Read the white paper: Intel & American Red Cross Southern Uganda Bridge Identification ›