We present a fully automated, zero-cost pipeline for fine-grained road segmentation from medium-resolution Sentinel-2 imagery, using all 12 spectral bands and binary supervision from OpenStreetMap (OSM). Unlike prior methods relying on high-resolution imagery or manual labels, our approach operates entirely on public data and requires no human annotation. A key contribution is the use of all Sentinel-2 spectral bands (10–60m), enabling multi-spectral road segmentation from medium-resolution imagery using only open data. We compare a baseline U-Net trained using full-image segmentation with a ResNet34-based U-Net that employs a tile-based training and inference strategy, decomposing each scene into fixed-size tiles to improve optimization under extreme class imbalance. Training is supervised via automatically rasterized OSM road vectors. Evaluation across two diverse datasets shows that regional diversity and data augmentation improve generalization, with the tile-based ResNet34 U-Net achieving higher IoU and F1-score alongside substantially improved recall, reflecting more complete and continuous road network extraction under severe class imbalance. The only manual input is the latitude/longitude of training regions; all other steps—from GeoJSON tiling to image download, mask generation, and alignment—are fully automated. This makes our method highly scalable and well-suited to mapping underrepresented or resource-constrained regions. We further demonstrate that applying the model across multiple time points enables identification of newly constructed or evolving road infrastructure absent from both OSM and Google Maps, including unmapped access routes in conflict-adjacent regions.