Road Segmentation for Remote Sensing Images Using Adversarial Spatial Pyramid Networks

Road extraction in remote sensing images is of great importance for a wide range applications. Because the complex background, and high density, most existing methods fail to accurately extract road network that appears correct complete. Moreover, they suffer from either insufficient training data or costs manual annotation. To address these problems, we introduce new model apply structured domain adaption synthetic image generation segmentation. We incorporate feature pyramid (FP) into generative adversarial networks minimize difference between source target domains. A generator learned produce quality images, discriminator attempts distinguish them. also propose FP improves performance proposed by extracting effective features all layers describing different scales' objects. Indeed, novel scale-wise architecture introduced learn multilevel maps improve semantics features. For optimization, trained joint reconstruction loss function, which minimizes fake real ones. experiments on three sets prove superior approach terms accuracy efficiency. In particular, our achieves state-of-the-art 78.86 IOU Massachusetts set with 14.89M parameters 86.78B FLOPs, 4× fewer FLOPs but higher (+3.47% IOU) than top performer among approaches used evaluation.