Enhancing Road Network Extraction from Remote Sensing Images

Satellite image analysis for road network information extraction is crucial, particularly in remote sensing. This initiative is significant for urban planning, real-time road map updates, and autonomous driving technology development. However, challenges arise, including small-scale roads, complex paths, and potential obscuration in satellite images, complicating detection by deep Convolutional Neural Networks (DCNNs). This article evaluates three road network extraction models U-Net, LinkNet34, and LinkNet34++—emphasizing data augmentation benefits. LinkNet34, based on ResNet34, stands out for efficiency, using a pretrained network for accelerated learning, especially suitable for medium-resolution images. Our study assesses accuracy, recall, F1 score, and Jaccard Index (IoU) comprehensively with a diverse dataset. Metrics highlight LinkNet34++'s superior efficiency compared to U-Net and LinkNet34, particularly with applied data augmentation, showcasing its capability to handle road image variability and enhance overall model robustness in road network extraction.