Open-Pit Mine Road Extraction From High-Resolution Remote Sensing Images Using RATT-UNet

Abstract

With the further development of the construction of “smart mine,” the technology of mine unmanned truck has developed rapidly. Road data are an important prerequisite for the mature application of mining and transportation dispatching system and unmanned driving, so it is particularly important to extract the road in open-pit mines timely and accurately. Different from urban roads, mine roads have no clear road edge and more background interference. In view of the above problems, a convolutional neural network called RATT-UNet (R: residual connection; ATT: attention), which combines residual connection, attention mechanism, and U-Net, is proposed to extract mine road from high-resolution remote sensing images. The advantages of the proposed method are listed as follows: first, the well-designed RATT unit combining residual connection and attention mechanism is used to construct the proposed neural network, which improves the network’s perception ability of detailed features. Second, the abundant skip connections and residual connections can improve the information flow, allowing a better network to be designed with fewer parameters. Third, a composite loss function based on structural similarity is proposed, which effectively alleviates the noise and edge blurring phenomenon and improves segmentation quality. Finally, we perform postprocessing optimization operations on the road extraction results. Experimental results demonstrate that the proposed RATT-UNet outperforms all the comparing network models in terms of the quality of extraction results and evaluation indicators in the mining road extraction task.