Evaluating Lightweight Deep Learning Models for Satellite Image Super-Resolution: A Comparative Study with Traditional Interpolation Techniques

High-resolution satellite imagery is essential for a wide range of remote sensing applications, yet acquiring such data is often limited by hardware and cost constraints. Singleimage super-resolution (SISR) offers a promising solution by enhancing spatial detail as a post-processing step. This paper presents a comparative study evaluating the effectiveness of lightweight deep learning models-ESPCN and SRCNNagainst traditional interpolation techniques including Nearest Neighbor, Bilinear, and Bicubic methods. Using satellite imagery and scaling factors of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3 x, 4 x$</tex>, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$10 x$</tex>, we assess each method's performance through PSNR, SSIM, and inference time. ESPCN consistently achieved the highest PSNR and SSIM scores across all scales (36.21 PSNR and 0.9449 SSIM at 3x) while maintaining fast inference times, making it a strong candidate for real-time or resource-limited scenarios. Traditional methods offered faster processing but significantly lower image quality. Our results highlight the trade-offs between accuracy and computational efficiency and suggest ESPCN as the most balanced approach for practical satellite image enhancement tasks. Future work will explore perceptual optimization, advanced architectures, and domain adaptation across different satellite platforms.