From satellite to shapefiles. ZymbaNet for memory‐efficient automated road mapping

Road network extraction from very-high-resolution (VHR) satellite imagery is a fundamental task in geospatial applications such as urban planning, navigation, and geographic information systems (GIS). Despite significant progress achieved by convolutional and attention-based architectures, existing methods still face two critical challenges: high computational complexity and the absence of complete automated pipelines that generate GIS-ready vector products. In this paper, we introduce ZymbaNet, a memory-efficient deep neural architecture that integrates an enhanced State Space Model (E-SSM) with Res2NeXt multiscale feature modules and PseudoMamba blocks for efficient long-range dependency modeling. ZymbaNet reduces memory requirements by more than 50% compared to existing attention-based networks while maintaining competitive segmentation accuracy. Beyond segmentation, we propose a fully automated road-to-GIS pipeline, which first georeferences the predicted road masks and then converts them into topologically valid shapefiles through skeletonization and Douglas-Peucker geometric simplification. Extensive experiments on the Massachusetts Roads and DeepGlobe datasets demonstrate the superiority of our approach, achieving an F1-score of 76.23% and 78.80% respectively, while significantly reducing the number of parameters and floating-point operations (FLOPs). This framework bridges the gap between pixel-level predictions and operational GIS-ready cartographic products, providing an efficient and scalable solution for automated road mapping at large scale.