RoadNeXt-CSS: ConvNeXt-Adapted Encoder-Decoder Road Extraction for Robust Spatial Information Systems

Background: Road extraction from remote sensing imagery is essential for geospatial information systems but faces challenges of thin structures, severe class imbalance and similarity between roads and non-roads.Objective: We propose RoadNeXt-CSS, integrating a ConvNeXt encoder, inverse frequency weighting (IFW) loss and a scale-sensitive decoder to address these challenges.Methods: RoadNeXt-CSS was trained and evaluated on the Massachusetts Roads and DeepGlobe benchmarks using intersection over union (IoU) as the primary metric.ConvNeXt extracts rich spatial-semantic features, IFW mitigates foreground-background imbalance and the scale-sensitive decoder adaptively weights multi-level features.Component-wise ablation experiments were conducted to isolate the contribution of each proposed component and paired t-tests were applied to assess statistical significance against baseline models.Results: RoadNeXt-CSS achieves IoU of 0.6779 (Massachusetts Roads) and 0.6984 (DeepGlobe), outperforming RCFSNet and other baselines (p < 0.001).Ablation analysis confirms that the ConvNeXt encoder, IFW loss and scale-sensitive decoder contribute independently to performance gains.Conclusion: The three components together provide robust and statistically significant road extraction improvements, supporting transportation planning, disaster response and navigation systems.