Accurate mapping of road networks is essential for navigation, disaster response, and urban planning. Traditional methods relying on remote sensing images, Lidar data, and GPS trajectories are hampered by time-consuming manual updates and incomplete extraction due to obstructions from trees, buildings, and shadows. However, recent advancements in deep learning models have shown promise in improving road mapping and automating the detection of missing roads.

This study uses a transfer learning approach, utilizing the pre-trained "Road-Extraction – North America" model from the ArcGIS Living Atlas of the World. Based on the Multi-Task Road Extractor architecture, this pre-trained model excels at identifying road pixels and effectively capturing connectivity between road segments. Using the GeoAI tools available in the ArcGIS system, we developed an enhanced deep learning model capable of identifying road networks with a Mean Intersection over Union (mIoU) score of 0.80. Despite its high performance, the model encounters challenges in areas with tall buildings and tree canopy cover, where occlusion-induced segmentation errors can occur.

To address these limitations and further improve the model accuracy, we have adopted a hybrid approach that integrates another deep learning model that detects only missing roads and incorporates data from various sources (Federal, Municipal, Overture Maps, Community Maps of Canada, etc.) to achieve robust missing road detection. The implications of this improved road network extraction extend beyond mapping, showing the potential to enhance the growth of smart cities, optimize traffic management, and improve emergency response times. This research aligns with Esri Canada's objective to advance road network mapping in Canadian cities and contribute to more comprehensive and accurate geographical information systems.