An AI-Driven Context-Aware Framework For Visibility-Adaptive Road Network Reconstruction and Shortest-Path Healing
Keywords:
artificial intelligence, machine learning, satellite imagery, road network reconstruction, occlusion detection, context-aware graph healing, shortest-path optimization, road connectivity.Abstract
Satellite images can offer an essential input for road network mapping and applications including navigation, city planning, disaster management, and geographic analysis. Roads can be obstructed from view due to various obstacles like trees, shadows, buildings, clouds, and other objects that cause fragmented roads and disconnected road networks. Road-healing techniques have been using geometrical considerations like distance, direction, and cost of path up until now, but there may be mistakes because reasons for the gaps in the roads are ignored.
In order to solve this problem, this study proposes a context-aware framework utilizing AI for robust reconstruction of the road network as well as repairing of shortest paths in the presence of occlusions. This framework takes into account the context information provided by the images along with graph-theoretic inference for determining whether certain gaps are due to occlusions or they are actually road breaks. An AI or ML-based occlusion context analysis module is included as an extension to the current Context-Aware Reconnection Score (CARS), which allows context features around the gap candidates to be taken in conjunction with geometric features such as endpoint distance, alignment, and road width consistency.
The framework is assessed on the Deep Globe Road Extraction benchmark which consists of 6,226 satellite image tiles and 74,712 annotated gap positions. The use of controlled synthetic occlusions is required to provide ground truth labels for reconnection and quantitative assessment. The best existing CARS guided shortest path configuration managed to obtain an F1 score of 0.448 on the entire test set, beating the existing distance and angle based, geometric shortest path approach and CARS guided configuration alone. The proposed AI extension framework opens up the possibility of learning representations while utilizing the shortest path framework.
