Improved Road Extraction Method Using SBD-Linknet and a Postprocessing Method for High Resolution Satellite Imagery

D-LinkNet has been a popular convolutional neural network. However, the scale of the model would be huge. Focused on the problem, this article made some improvements:(1) Replace the initial block with a stem block. (2) Add bottleneck layers before and after DBlock to form a new module called DBlockPlus. (3) Rebuild the entire network based on ResNet. The article also proposed a method named ESIPs to eliminate small independent patches and adopted a composite loss function called DF loss to better address the impact of extreme class imbalance in binary classification. And another composite loss function was introduced for comparative experiments. The experimental results show that the SBD-LinkNet not only has less parameters but also has a higher accuracy and the road extraction results presented a better visual effect after processing by ESIPs. And the model that uses DF loss also performs better in road extraction from complicated scene.