Transformer with Reinforcement and Actor- Critic for Edge-vectorization (TRACE)

The National Geospatial-Intelligence Agency (NGA) has a unique mission set that includes not only providing accurate and timely geospatial intelligence (GEOINT) data to the warfighter, but also generating and maintaining accurate “foundation data.” The generation of Foundation Feature Data (FFD) vector features such as roads and building outlines from aerial imagery is an incredibly labor-intensive manual process. Because of this, the data is not updated rapidly enough to maintain relevance. Current automated feature extraction methodologies often yield incomplete, inaccurate, and non-routable road networks, particularly in scenarios with visual obstructions like tree-lined roads. To address this problem, the collaboration between In-Depth Engineering and Miami University of Ohio proposes the “Transformer with Reinforcement and Actor-Critic for Edge-vectorization” (TRACE) solution – a refined approach to advance road extraction methodologies that trains the latest CNN-based segmentation models to be robust to occlusions and couples that with an actor-critic-based Reinforcement Learning (RL) technique for growing road networks. Together, these form a solution that is more robust than either one by itself. The team executed an aggressive Phase I effort and demonstrated the feasibility of the approach, developing training and testing pipelines that will be used in Phase II to complete the TRACE prototype.