Abstract: The road extraction of the remote sensing image plays an important role in the intelligent understanding of the ground. According to the structural characteristics of the road features, a road extraction method with structure similarity loss function and structural descriptor was proposed. Firstly, the proportion of the road is usually small in the remote sensing image, a shallow encoder-decoder based segment network with high resolution was proposed. Secondly, the structural similarity(SSIM) was introduced to the loss function and as the existing methods of road extraction network are mostly based on the comparison of the prediction and ground truth of each pixel value, the structural descriptor joined the task of road extraction as an optimization step which improves the ability of the network to make use of the structural information. Lastly, experiments on Massachusetts road dataset show that the proposed network gets the precision and F1-score up to 85.3% and 84.6%.

Key words: deep learning, remote sensing , road extraction, structural descriptor, semantic segmentation