At a Glance
National Geospatial-Intelligence Agency
Develop a trainable deep learning feature identification system capable of detecting and localizing important security and combat-related features in 2D and 3D satellite & drone imagery.
Created pre-processing, training & inference system trained on a combination of open and secret datasets. Object labels have hundreds of classes and often 200k building examples per image. This system is an on-going project with dedicated UI, Flask/Docker App, batch processing, and introduction of new multi-class training approaches. The system is currently being deployed via AWS GovCloud for all terrains and training locations.
Expertise & Technology
For the NGA-sponsored project, over a one year effort, working through Alion Science & Technology, we created a state of the art (SOTA) system to detect critical features in large cityscape satellite images by developing novel segmentation algorithms coupled to an image processing architecture to support GPU-accelerated capabilities in image ingestion, processing, and real-time change detection with tactically relevant accuracy and time.
SFL Scientific Solution
SFL Scientific trained the SOTA neural networks on a combination of the satellite image, masked data, and unlabeled data and developed special datasets for use by Alion and agency stakeholders by tuning and identifying labeling requirements, errors, and accelerators for millions of structures (farms, ports, skyscrapers, roads, etc.). Working with stakeholders, we identified algorithmic strategies to offset program technical constraints, such as off-angle images, cloud cover, smoke, and other common and uncommon outliers.
Storage of data structures, masks, building, and other feature statistics are secured in the AWS environment, with algorithms strengthened through training regimes and low-level adversarial attacks; We examined such effectiveness of several possible defense methods for image attacks, including training through image transformation and augmentation and GANs training, including compression.
The new capability provided for a modular environment, scalable with on-premise GPUs or AWS cloud infrastructure to support large-scale, algorithmic data labeling and segmentation of targets to augment human reviewers with sub-minute processing speeds. Further, SFL Scientific engineers developed UI and automation workflows for detailed display of density information, pixel-level masks, and structure prediction statistics for difficult environments (cloud cover, weather, off-angle) and showed human labeling deficiencies to improve the labeling accuracy of detection. By design, analysts could now just review and focus on hundreds of target structures in an image instead of hundreds of thousands in each quadrant.
Detection Ability: The developed deep learning system currently is able to process images of any building density and dimensions in under a minute using these modern detection methods with GPU acceleration, at over 95% accuracy. It is capable of recognizing the change of structures in various conditions but also can be used on trucks, cars, trains, roads, storage containers, and mask other desired features. The deep learning system currently is able to process full-resolution images (20k x 20k) of any building density and dimensions in under a minute.