Abstract
Road extraction or segmentation from high-resolution remote sensing imagery has attracted considerable attention in various applications including urban planning, transportation planning, and road monitoring. Furthermore, road extraction from satellite images is of great significance to emergency management and risk assessment. However, it remains challenging due to various uncertainties such as the occlusions of vehicles, trees, and clouds. Over the last few years, many traditional machine learning approaches have been suggested for extracting road objects from remote sensing images. However, the traditional road extraction approaches are costly, labor-intensive, and unreliable. Moreover, the traditional techniques suffer from either insufficient training data or high costs of manual annotation. Compared with the traditional machine learning methods, with the high accuracy and reliability, several novel deep learning methods have achieved great successes in road extraction. In this paper, we evaluate the innovations and achievements in the U-Net architecture and provide significant reflections on recent trends. For road extraction from high-resolution satellite images, we present an overview of the four state-of-the-art deep learning methods: U-Net and its variants such as Attention U-Net, Residual U-Net, and U-Net++. Through the comparative experiments, we highlight the strength and identify gaps and limitations of each deep learning method using various evaluation metrics: accuracy, intersection over union, area under curve, and F1 score. According to the experimental results, the U-Net++ used to extract roads from high-resolution satellite imagery has the highest average performance measure. Finally, while discussing the experimental results, a crucial conclusion is presented with suggestions and future works.