Optimized U-Net Model for Advanced Road Extraction Using Satellite Images

Extracting accurate road networks from high-resolution satellite images is the key feature for updating geospatial databases and supporting autonomous navigation. However automatic road extraction is a difficult task due to various aspects like shadows, occlusion of buildings, trees, and varying road conditions that hinder precise extraction. This paper focuses on using the hybrid loss function that integrates dice loss, binary crossentropy, and focal loss that enable the optimized U-Net architecture to counter the class imbalance and mitigate the impact of small road segments. The experiment is conducted on a DeepGlobe dataset that demonstrates improved performance metrics and road connectivity compared to state-of-the-art methods. Overall, our model enhances the ability to extract the road network accurately.