Multi Image Super Resolution using Deep Learning Techniques for Satellite Images
K. Deepthi, B. L. Nethish, R. Shravan Kumar, Akshay Palya, Surag S. Kshari, K. Aditya Shastry · 2025
Remote sensing image analysis presents significant challenges due to resolution limitations. This paper explores Deep Learning approaches for Multi-Image Super-Resolution (MISR) of satellite imagery. Current Convolutional Neural Network (CNN) methods demonstrate strong performance but primarily focus on Single-Image Super-Resolution (SISR), which limits the potential benefits of utilizing multi-temporal satellite data. The proposed Residual Attention Multi-Image Super-Resolution (RAMS) framework effectively performs MISR by integrating spatial-temporal patterns across multiple low-resolution images. The model enhances data fusion and recovers essential details through 3D convolutional operations with attention mechanisms, while eliminating redundant low-frequency signals via layered residual connections. Experimental results on the Proba-V dataset demonstrate that this approach outperforms existing SISR and MISR techniques, achieving superior PSNR and SSIM metrics for remote sensing applications, particularly in environmental monitoring, satellite image acquisition, and urban development solutions.