Description
This work explores the application of Denoising Diffusion Probabilistic Models (DDPMs) to super-resolution tasks in remote sensing imagery. Many research domains within Earth observation face a scarcity of high-quality data, largely due to the complexity and cost of acquiring satellite imagery. To address this challenge, we use Sentinel-2 data from the Copernicus Programme of the European Space Agency (ESA) as a testbed. Diffusion models are particularly attractive in this context because of their versatility: once trained for a specific task—such as image super-resolution—they can be adapted to related tasks with minimal architectural or code modifications. These models have already demonstrated strong potential in applications such as crop generation, cross-domain image translation (e.g., radar-to-optical), and a wide range of generative modeling scenarios.
In this study, we adopt the Latent Diffusion Model (LDM) framework, in which the diffusion process operates within the latent space of a Variational Autoencoder (VAE). Our results show that diffusion-based approaches offer a promising path for enhancing remote sensing data quality. Specifically, the proposed method successfully reconstructs satellite images at a spatial resolution of 10 m × 10 m (100 m² per pixel) from inputs at 20 m × 20 m resolution (400 m² per pixel), highlighting the potential of diffusion models for practical super-resolution applications in Earth observation.
