Road Extraction From High-Resolution Remote Sensing Images of Open-Pit Mine Using D-SegNeXt

High-precision 3-D road networks in open-pit mines (OPMs) play a crucial role in production planning, truck dispatching, and unmanned driving. Compared with urban road networks, the boundaries of OPM roads are indistinct, with varying widths. The colors of these roads blend with the surrounding environments and they undergo rapid changes. Thus, accurately, efficiently, and timely obtaining mining road networks still face many challenges. With the development and popularization of UAV technology, it is now possible to obtain real-time spatial data. We propose a hybrid dilated multiscale convolution attention (HDMSCA) unit and design the D-SegNeXt network. This method offers several advantages. First, it reduces the computational complexity and enlarges the receptive field through hybrid dilated convolution. Second, residual networks and multiscale convolutions can extract local, distant, long, and narrow features, thereby enhancing the network's ability to capture long-range dependencies. In addition, we construct an OPM road dataset and test the models on it. The experimental results demonstrate that our model outperforms several benchmark networks in both image classification and road extraction. Our D-SegNeXt model achieves a Top-1 acc score of 82.8% on ImageNet-1k, an intersection over union (IoU) score of 75.59% on the OPM road dataset, and an IoU score of 67.96% on the DeepGlobe Road Extraction Challenge dataset. Our dataset and code are available at https://github.com/orgs/D-SegNeXt/repositories.