Large-Scale High-Resolution Essential Urban Land Cover Category Mapping Using a Semantic-Augmented and Noise-Tolerant Approach

High-resolution urban land cover information is essential for understanding the complex urban environment. Although publicly available land cover products offer valuable opportunities to minimize labor-intensive annotation efforts in deep learning-based mapping, effective strategies to address their inherent uncertainties remain underdeveloped. To mitigate the negative impact of label noise and enhance segmentation accuracy, we propose a Semantic-augmented and noise-tolerant Network (SantNet) for high-resolution urban land cover mapping using public products as imperfect labels. SantNet incorporates three key components: 1) a high-resolution-preserving backbone (HPB) to maintain detailed image resolution; 2) a noise-handling module (NHM) to enhance robustness against erroneous labels; and 3) a semantic augmentation module (SAM) to improve feature representation through noise-adapted contrastive learning. Benchmark comparisons with six general semantic segmentation networks and three noise-robust models demonstrate the superiority of our proposed model. Extensive ablation studies further confirm the effectiveness of each module within SantNet. Additionally, we validated SantNet’s transferability across different cities, where it consistently outperformed other models, demonstrating its suitability for large-scale mapping applications. We employed this model to generate 1-m high-resolution Essential Urban Land Cover Category (EULCC) mapping for all urban areas across the Conterminous United States (CONUS). This dataset, with a validated overall accuracy (OA) of 88.77%, surpasses existing large-scale urban land cover products in spatial coverage, resolution, detail richness, and category granularity, positioning EULCC as a foundational resource for precise urban studies and applications. The data are publicly available at https://doi.org/10.5281/zenodo.13987465