Enhancing the Spatial Resolution of Sentinel-2 Images Through Super-Resolution Using Transformer-Based Deep-Learning Models

Abstract

Satellite imagery plays a pivotal role in environmental monitoring, urban planning, and national security. However, spatial resolution limitations of current satellite sensors restrict the clarity and usability of captured images. This study introduces a novel transformer-based deep-learning model to enhance the spatial resolution of Sentinel-2 images. The proposed architecture leverages multihead attention and integrated spatial and channel attention mechanisms to effectively extract and reconstruct fine details from low-resolution inputs. The model's performance was evaluated on the Sentinel-2 dataset, along with benchmark datasets (AID and UC-Merced), and compared against state-of-the-art methods, including ResNet, Swin Transformer, and ViT. Experimental results demonstrate superior performance, achieving a peak signal-to-noise ratio (PSNR) of 33.52 dB, structural similarity index (SSIM) of 0.862, and signal-to-reconstruction error ratio (SRE) of 36.7 dB on Sentinel-2 RGB bands. The proposed method outperforms state-of-the-art approaches, including ResNet, Swin Transformer, and ViT, on benchmark datasets (Sentinel-2, AID, and UC-Merced). The results demonstrate that the proposed method achieves superior performance in terms of PSNR, SSIM, and SRE metrics, highlighting its effectiveness in revealing finer spatial details and improving image quality for practical remote sensing applications.