Remote Sensing Road Extraction by Refining Road Topology

Abstract

Remote sensing road extraction is one of the research hotspots in high-resolution remote sensing images. However, many road extraction methods cannot hold the edge interference, including shadows of sheltered trees and vehicles. In this paper, a novel remote sensing road extraction (RSRE) method based on deep learning is proposed, which considers the road topology information refinement in high-resolution image. Firstly, two parallel operations, which named dilation module (DM) and message module (MM) in this paper, are embedded in the center of semantic segmentation network to tackle the issue of incoherent edges. DM containing dilated convolutions is used to capture more context information in remote sensing images. MM consisting of slice-by-slice convolutions is used to learn the spatial relations and the continuous prior of the road efficiently. Secondly, a new loss function is designed by combining dice coefficient term and binary cross-entropy term, which can leverage the effects of different loss. Finally, extensive experimental results demonstrate that the RSRE outperforms the state-of-the-art methods in two public datasets.