An improved method for road extraction from high-resolution remote-sensing images that enhances boundary information

Abstract

At present, deep-learning methods have been widely used in road extraction from remote-sensing images and have effectively improved the accuracy of road extraction. However, these methods are still affected by the loss of spatial features and the lack of global context information. To solve these problems, we propose a new network for road extraction, the coord-dense-global (CDG) model, built on three parts: a coordconv module by putting coordinate information into feature maps aimed at reducing the loss of spatial information and strengthening road boundaries, an improved dense convolutional network (DenseNet) that could make full use of multiple features through own dense blocks, and a global attention module designed to highlight high-level information and improve category classification by using pooling operation to introduce global information. When tested on a complex road dataset from Massachusetts, USA, CDG achieved clearly superior performance to contemporary networks such as DeepLabV3+, U-net, and D-LinkNet. For example, its mean IoU (intersection of the prediction and ground truth regions over their union) and mean F1 score (evaluation metric for the harmonic mean of the precision and recall metrics) were 61.90% and 76.10%, respectively, which were 1.19% and 0.95% higher than the results of D-LinkNet (the winner of a road-extraction contest). In addition, CDG was also superior to the other three models in solving the problem of tree occlusion. Finally, in universality research with the Gaofen-2 satellite dataset, the CDG model also performed well at extracting the road network in the test maps of Hefei and Tianjin, China.