Multi-Scale Adaptive Road Extraction Network for High-Resolution Remote Sensing Images

Road extraction from high-resolution remote sensing images faces challenges such as discontinuities caused by occlusions from trees and buildings, as well as computational complexity arising from varying road orientations and widths. In order to solve these problems, this paper proposes a Multi-Scale Adaptive Road Extraction Network (MARENet), which takes ConvNeXt as the backbone and designs the Atrous Pyramid Convolution (APC) module, which can accurately extract the roads in the image, and the Frequency Adaptive Transformer (FAT) module, which can focus on the important features in the image and avoid the problem of incomplete extraction. The network can effectively solve the above two key problems and achieve accurate road extraction. EARENet was evaluated experimentally on a commonly used and challenging DeepGlobe dataset. The proposed method is superior to the existing methods in terms of accuracy, precision, recall, F1-score and IoU, and the results show that the proposed network has robustness and superiority in processing high-resolution road remote sensing images.