Our Work - Maxar Technologies SecureWatch
Computer Vision Aircraft Detection Models Add Deeper View into Airport Activity
SecureWatch is an advanced cloud-based subscription service for obtaining timely and secure access to Earth Intelligence. Organizations leverage SecureWatch for a wide range of geospatial projects – from detecting changes and monitoring assets, to assisting in humanitarian relief efforts.
Maxar turned to Wovenware to enhance the insights delivered through SecureWatch, with the addition of computer vision models that would help it to automatically detect aircraft in airports around the world. AI-powered geospatial insights are generated on a daily basis as new satellite imagery becomes available.
What We Faced
- Build a fast and efficient model to detect aircraft in 5,000 airports around the world in satellite imagery
- Build a training data-set and an independent validation set with a wide range of geographies and diverse satellite image resolutions and angles
- Build a model that is fast and cost-effective to maintain
- Integrate the machine learning model with Maxar’s Secure Watch Product
- Operationalize the retraining and deployment of new models to Maxar’s SecureWatch
How We Helped
Wovenware leveraged its private crowd to create high quality training and validation datasets and used its Innovation Sprint methodology to experiment and iterate through various models.
- A private crowd annotated 27K aircraft segmentations over 180 images from 42 airports around the world using Maxar’s proprietary toolkit DeepCore Workbench.
- The data science team built a YoloV5 deep learning model to detect aircraft, experimenting with several architectures, hyperparameter tuning, and balancing image sets
- The aircraft model was retrained every week with new imagery and labels and deployed to Maxar’s SecureWatch product
- Each version of the model was scored against an independent validation set to benchmark, measure and track performance improvements
Maxar has deployed an updated aircraft detection model which is 10X faster to train than previous models and significantly more accurate in performance. Not only can Maxar extract insights from airport activity everywhere in the world, but it is also doing it cost effectively. The high accuracy of the model allows image analysts to work more efficiently in extracting focal points in SecureWatch. The following are the metrics for the models:
- Recall: 80%
- Precision: 79%