<p>Transportation is a mandatory and fundamental element to keep as a top priority for civilized and industrialized countries, in which roads play an important role. This paper proposes a novel deep learning pipeline for automatic road extraction from high-resolution remote sensing imagery, framed as a semantic segmentation task. Road extraction from such imagery is a challenging task due to complex backgrounds, shadow occlusions, and the intricate topology of road networks. The task is further complicated by issues such as low spectral variation between roads and non-road objects like building rooftops, leading to high intra-class variance. Most previous techniques failed to maintain edges and boundaries properly, which is an important criterion in accurate road extraction. To address these challenges, this paper presents a modified-UNet (M-UNet) architecture combined with a Directional Conditional Random Fields (DCRF) post-processing technique. The proposed M-UNet with DCRF is rigorously evaluated on two benchmark datasets: The Massachusetts Roads Dataset and the DeepGlobe Roads Dataset. Our method demonstrates statistically significant improvements, achieving a final Road Intersection over Union (IoU) of 96% on the DeepGlobe dataset and an accuracy of 97.90%, outperforming state-of-the-art methods.</p>