DFGNet: A Hybrid Approach for Enhanced Road Extraction with Satellite Imagery and Global Fusion

Road network extraction from satellite imagery plays a vital role in applications like autonomous navigation and urban development. This paper presents DFGNet, an advanced hybrid architecture designed to enhance the precision and scalability of road segmentation tasks. DFGNet combines the strengths of D-LinkNet and FuseNet while introducing a novel Global Fusion Module (GFM) for effective feature integration. The encoder utilizes hybrid blocks with dilated convolutions to capture features across multiple scales, while residual connections ensure efficient feature propagation. In the decoder, the GFM optimally aligns spatial and feature information from the encoder, leading to superior segmentation accuracy. Additional refinement processes, including noise suppression and vectorization, further enhance the quality of extracted road networks. Evaluations on real-world remote sensing datasets highlight that DFGNet achieves competitive performance with an F1-score of 0.96 and an IoU of 0.79, surpassing several state-of-the-art approaches. This research offers a practical and robust solution for high-resolution road extraction, advancing the capabilities of remote sensing and geospatial analysis.