Deep Learning-Based Methods for Road Extraction From Remote Sensing Images: A vision, survey, and future directions

A rising number of high-resolution photographs obtained using remote sensing are currently available. However, extracting the necessary information from such a massive volume of data in an accurate and timely manner remains a significant challenge. As crucial geographical information, road information is widely used in multiple scenarios, such as map updates, traffic management, and disaster management. However, due to the variation of road appearance and occlusion, automatically extracting the road from remote sensing images remains one of the most complex subjects in remote sensing. In recent years, the application of deep neural networks has resulted in substantial advancements in road extraction algorithms from satellite images. This paper provided a new vision and an overview of the most relevant datasets commonly used in evaluation methods for extracting roads from remote sensing images. Next, a comprehensive review of road extraction techniques was presented, specifically highlighting the model development process and their performance evaluation. Finally, the major challenges and future directions of road extraction from remote sensing images were discussed.