The extraction of information from remote sensing data is essential for numerous fields, including urban planning, transportation and traffic management, disaster response, and monitoring environmental changes. Automatic extraction of road networks from satellite images remains a significant challenge due to their diverse structures and scales. Deep learning models that are based on convolutional neural networks have demonstrated exceptional performance in this semantic segmentation task. In this work, a U-Net model has been developed to accurately extract different types of roads from high-resolution satellite imagery. The model follows classic encoder-decoder architecture with skip connections, trained and tested on the DeepGlobe Road Extraction dataset. The encoder utilizes convolutional and max-pooling layers to capture context, while the decoder employs transposed convolutions for precise localization, leveraging skip connections to recover spatial detail. Quantitative evaluations on this benchmark establish that our model achieves a higher IoU (66.62%) and Precision (87.01%) than existing state-of-the-art methods, while maintaining a comparable F1-Score (71.11%), indicating superior detection accuracy with fewer false positives. The results confirm the effectiveness of the proposed approach for robust road extraction.
Jabbar et al. (Mon,) studied this question.