Improving Road Semantic Segmentation Using Generative Adversarial Network

Road network extraction from remotely sensed imagery has become a powerful tool for updating geospatial databases, owing to the success of convolutional neural (CNN) based deep learning semantic segmentation techniques combined with high-resolution that modern remote sensing provides. However, most CNN approaches cannot obtain high precision maps rich details when processing imagery. In this study, we propose generative adversarial (GAN)-based approach road aerial part presented GAN approach, use modified UNet model (MUNet) map network. combination simple pre-processing comprising edge-preserving filtering, proposed offers significant improvement in compared prior approaches. experiments conducted on Massachusetts image dataset, achieves 91.54% and 92.92% recall, which correspond Mathews correlation coefficient (MCC) 91.13%, Mean intersection over union (MIOU) 87.43% F1-score 92.20%. Comparisons demonstrate framework outperforms CNN-based is particularly effective preserving edge information.