Super-Resolution Enhancement of Landsat Satellite Data Using Generative Adversarial Networks (GANs) and UNET Deep Learning Algorithms
Sai Balakavi · 2025
Super-resolution methods enhance the spatial resolution of remote sensing imagery, extracting finer details from lower-resolution data. While traditional techniques relied on signal processing and frequency-domain methods, the potential of deep learning for super-resolution remains underexplored for medium-resolution datasets like Landsat. This study evaluates two deep learning-based approaches, the Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) and the Enhanced Super-Resolution U-Net (ESRUNet), to super-resolve Landsat data. Using Landsat and PlanetScope imagery from February 25, 2019, in Bandipur, India, the study highlights each model’s strengths and limitations. ESRGAN excelled in generating visually realistic outputs with improved smoothness and perceptual quality similar to very high resolution Planetscope images, while ESRUNet prioritized spatial integrity and structural fidelity. Patch-level statistics from the 3×3 moving window analysis showed ESRGAN’s superior performance in reconstruction accuracy and perceptual fidelity, with significantly lower mean squared error (MSE) and mean absolute error (MAE) compared to ESRUNET. Additionally, ESRGAN achieved a markedly higher peak signal-to-noise ratio (PSNR), reflecting less noise and greater visual fidelity. These findings, consistent in whole image analysis, demonstrate ESRGAN’s reliability in producing high-quality super-resolved images, making it valuable for land cover/land use change (LCLUC) studies.