Improving SAM-Road Model for Occlusion Handling in Road Networks Extraction from Satellite Images with Gamma Correction and Modified A* Algorithm

Abstract

Background: Occlusion in satellite imagery often leads to disconnected road networks, reducing the quality and reliability of geospatial data, which in turn hampers infrastructure and transportation planning. Although models like SAM-Road show promising results, they still struggle in handling occluded areas, especially at intersections and curved roads. Some existing methods, such as the extended-line approach, have been proposed to address occlusion; however, they are typically limited to handling linear road segments and are less effective in complex structures. To overcome these limitations, this study enhances the SAM-Road framework by incorporating gamma correction and a modified A algorithm in the post-processing stage. This combined approach improves the visibility of partially hidden roads and successfully reconnects fragmented segments, even in non-linear and occlusion-heavy areas. Objective: This study aims to improve the accuracy of road network extraction by enhancing the SAM-Road model with gamma correction and a modified A* algorithm, specifically to address the problem of occlusion in satellite imagery. Methods: This research adopts a quantitative experimental approach. Gamma correction is applied to enhance the visual contrast of roads in occluded satellite images, while the A pathfinding algorithm is modified to reconnect disjointed road segments. The integrated method is then evaluated using accuracy metrics, specifically the TOPO and APLS (Average Path Length Similarity) scores. Results: The experimental findings indicate that each method—SAM-Road baseline, gamma correction, the modified A* algorithm, and their combination—delivers distinct performance improvements. Gamma correction alone achieves the best results at gamma 1.5 (TOPO 80.61%) and 1.25 (APLS 70.91%) on SpaceNet, and gamma 2.0 (TOPO 78.56%) and 1.25 (APLS 68.73%) on City-scale. The modified A* algorithm performs best at 16/8 (TOPO 80.22%) and 32/16 (APLS 70.71%) on SpaceNet, and 16/8 (TOPO 77.29%) and 64/32 (APLS 70.94%) on City-scale. When combined, the method yields results within the range of TOPO 75.41–80.59% and APLS 67.59–71.17% on SpaceNet, and TOPO 76.52–78.56% and APLS 66.39–71.19% on City-scale. Conclusion: This study concludes that the integration of gamma correction and a modified A* algorithm effectively addresses occlusion-related challenges in satellite imagery. While each technique contributes unique improvements, their combination significantly enhances the accuracy and continuity of extracted road networks—not only in straight road segments such as extended-line method, but also in more complex or occluded areas. The results confirm that this hybrid approach yields road extraction outputs that more closely align with ground truth in terms of topology and structure. Future research could explore integrating gamma correction and the modified A* algorithm directly into the training process, aiming to enhance model performance while maintaining high accuracy. Keywords: SAM-Road, modified A*, gamma correction, occlusion, road network extraction, satellite imagery, topology, geometry