Reinforced Block-matching-based Approach for Extraction of Cartographic Road Networks from Satellite Imagery (ReCarto)

The National Geospatial-Intelligence Agency (NGA) stands at the forefront of providing timely, relevant, and accurate geospatial intelligence in support of national security. Integral to their mission is the accurate interpretation of data gathered from remote sensing satellites and generate foundation feature data vector (FFD) features like roads and building outlines. With the increasing availability of high-resolution satellite imagery and an influx of geospatial datasets, NGA faces the challenge of combining geographic information from overlapping sources. Deep Learning (DL)-based vision algorithms, particularly Convolutional Neural Networks (CNNs), have emerged over the past decade as transformative tools for feature extraction from these datasets. NGA’s ability to discern intricate details from multi-source remote sensing activities has revolutionized geospatial analysis.  Current DL-based models rely mostly on subjective visual interpretations that could lead to gaps or inaccuracies, especially in regions where obstructions or varying terrain exist. Such challenges will compromise the reliability of extracted data, making some road networks non-routable and potentially hindering NGA's mission-critical operations.  Recognizing these challenges, Knowledge Based Systems, Inc. (KBSI) proposes an innovative Reinforced Block-matching-based Approach for Extraction of Cartographic Road Networks from Satellite Imagery (ReCarto). The proposed methods use advanced information fusion methods for reinforcement learning and graph-theoretic reward mechanisms that will enhance road network extraction from high resolution satellite imagery. The proposed graph-theoretic approach facilitates a structured interpretation of road connectivity, ensuring the extraction of coherent and navigable pathways. Block matching enhanced reinforcement learning, on the other hand, allows the system to iteratively refine its predictions, learning optimal sequences of actions for precise feature delineation.