STCN-SR: A Spatial Transformer-CNN Framework for Super-Resolution of Sentinel-2 satellite imagery

The management and mitigation of disasters increasingly rely on remote sensing data. However, the spatial resolution of publicly available satellite imagery often limits the detection of fine-scale features, reducing the accuracy of small-scale environmental analysis. This paper proposes a deep learning-based super-resolution (SR) framework to enhance the spatial resolution of Sentinel-2 imagery. The proposed method, termed STCN-SR (Spatial Transformer-CNN Network for Super-Resolution), integrates convolutional neural networks and transformer-based attention mechanisms to jointly capture local textures and global spatial relationships. The model is trained on low and high-resolution image pairs and reconstructs super-resolved images from 10 m to 5 m resolution. The generated outputs are georeferenced using metadata from the original tiles to maintain alignment within GIS workflows. Quantitative evaluation using peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) demonstrates improved reconstruction quality, achieving a maximum PSNR of 43.91 dB on the test dataset. Qualitative comparison with UAV reference imagery indicates improved boundary sharpness and clearer small-scale structures, suggesting potential benefits for environmental monitoring tasks, particularly in flood-affected landscapes. Overall, the results demonstrate the potential of deep learning-based super-resolution to improve the utility of Sentinel-2 imagery for geospatial applications.