Segmentation Is Not the End of Road Extraction: An All-Visible Denoising Autoencoder for Connected and Smooth Road Reconstruction

With a plethora of remote sensing (RS) images, deep neural network-based semantic segmentation model (SegModel) achieves commendable road extraction performance. However, the occlusions caused by vehicles, roadside objects, and shadows cannot be directly identified as road pixels, especially on high-resolution RS images. Therefore, relying only on a single SegModel to guarantee road connectivity and boundary smoothness in road extraction tasks is extremely difficult. To address this issue, this article puts forward a “segmentation-with-reconstruction” framework, which comprises a SegModel to generate the binary road labels from RS images and a reconstruction model to refine the road labels. Specifically, the former can be compatible with arbitrary existing SegModels, while the latter is built by our proposed model named all-visible denoising autoencoder (AV-DAE). The AV-DAE is designed to be an encoder–decoder architecture that takes topology-corruption road labels as inputs and true road labels as outputs. To better train the AV-DAE, we further present three noise-adding strategies to corrupt road labels for diverse patterns and train the AV-DAE to reconstruct them. Being RS-image-agnostic, the AV-DAE pays more attention to the spatial features rather than the spectral features, which enables it to recover the road topology by improving the connectivity and boundary smoothness. Finally, elaborate simulation results demonstrate that the proposed framework can significantly improve the connectivity and boundary smoothness of the extracted roads while achieving a competitive road extraction performance and high generalization ability compared to the benchmarks.