RoadCT: A Hybrid CNN-Transformer Network for Road Extraction From Satellite Imagery

Electronic road map is essential to support many intelligent transportation applications, and extracting roads from satellite images is a promising approach for map service providers to update their road networks efficiently. Hence, this letter proposes a hybrid deep neural network called RoadCT to improve the performance of road extraction. RoadCT not only integrates the strengths of both convolution and transformer neural networks, but also adopts a relational fusion block to merge the road features with different receptive fields. Extensive evaluations based on two public datasets have illustrated that RoadCT outperforms other state-of-art algorithms by 1.1% - 3.9% on F1 Score and 1.6% - 6.0% on intersection over union.