Super-Resolution Of Geospatial Images Using Enhanced Gans

Abstract

High-resolution imagery is essential for critical applications such as environmental monitoring, urban planning, and disaster response, where accurate details support informed decision-making. However, limitations in available imaging systems for public use and resource constraints often result in low-resolution satellite images lacking the necessary detail. Super-resolution (SR) methods have emerged to address these limitations, with deep learning approaches like Generative Adversarial Networks (GANs) and Transformer-based models offering promising results. This study investigates a GAN-focused SR approach, linking Real-ESRGAN with Transformer-based methods such as SwinIR to obtain higher-resolution usable images. Real-ESRGAN’s multi-scale discriminators and Residual-in-Residual Dense Blocks (RRDB) effectively capture complex textures and mitigate noise, making it suitable for high-detail satellite imagery. Our results demonstrate significant improvements in image clarity and overall perceptual quality, supporting applications requiring precise, high-resolution images.