A radiometrically and spatially consistent super-resolution framework for Sentinel-2

Deep learning-based super-resolution (SR) models offer a promising approach to enhancing the effective spatial resolution of optical satellite images. However, existing SR implementations have shown that, while these models can reconstruct fine-scale details, they often introduce undesirable artifacts, such as nonexistent local structures, reflectance distortions, and geometric misalignment. To mitigate these issues, fully synthetic data approaches have been explored for training, as they provide complete control over the degradation process and allow precise supervision and ground-truth availability. However, challenges in domain transfer have limited their effectiveness when applied to real satellite images. In this work, we propose SEN2SR , a new deep learning framework trained to super-resolve Sentinel-2 images while preserving spectral and spatial alignment consistency. Our approach harmonizes synthetic training data to match the spectral and spatial characteristics of Sentinel-2, ensuring realistic and artifact-free enhancements. SEN2SR generates 2.5-meter resolution images for Sentinel-2, upsampling the 10-meter RGB and NIR bands and the 20-meter Red Edge and SWIR bands. To ensure that SR models focus exclusively on enhancing spatial resolution, we introduce a low-frequency hard constraint layer at the final stage of SR networks that always enforces spectral consistency by preserving the original low-frequency content. We evaluate a range of deep learning architectures, including Convolutional Neural Networks, Mamba, and Swin Transformers, within a comprehensive assessment framework that integrates Explainable AI (xAI) techniques. Quantitatively, our framework achieves superior PSNR while maintaining near-zero reflectance deviation and spatial misalignment, outperforming state-of-the-art SR frameworks. Moreover, we demonstrate maintained radiometric fidelity in downstream tasks that demand high-fidelity spectral information and reveal a significant correlation between model performance and pixel-level model activation. Qualitative results show that SR networks effectively handle diverse land cover scenarios without introducing spurious high-frequency details in out-of-distribution cases. Overall, this research underscores the potential of SR techniques in Earth observation, paving the way for more precise monitoring of the Earth’s surface. Models, code, and examples are publicly available at https://github.com/ESAOpenSR/SEN2SR .