Integrating Topology and Geospatial Knowledge for Mapping Road Network Layers from High-Resolution Remote Sensing Images

The active research in artificial intelligence has inspired a radical transformation in problem-solving capabilities across various domains. Among all, it has proved its effectiveness in the remote sensing research community wherein researchers are utilizing different deep learning techniques for remote sensing scene understanding. Meanwhile, with high-resolution Earth observation data availability, automatically extracting spatial information has become a vital challenge. Additionally, mapping spatial entities such as road features is crucial for decision-making and improving transportation systems, urban planning, disaster rescue systems, and many more. This study proposes a novel deep learning-based convolutional neural network pipeline for road extraction and vectorization. The novelty of the proposed work lies in eradicating the occlusion problem faced due to the shadow of trees, tall buildings, or vehicles for road extraction tasks. The proposed model achieves 85.14% mIoU and 83.56% MCC on unseen test datasets compared to other state-of-the-art architectures.