Diffusion-Based Super-Resolution of Sentinel-2 Imagery for Improved Forest Species Classification

The development and launch of satellites is a costly process that often imposes strict limitations on the size, weight, and capabilities of onboard sensors. For multispectral Earth observation, these constraints typically result in imagery with lower spatial resolution, reducing the accuracy of downstream analytical tasks such as environmental monitoring. Artificial intelligence (AI)-based super-resolution (SR) techniques offer a promising solution to mitigate these limitations by computationally enhancing image quality, thereby relaxing the hardware requirements for onboard sensors. In this study, we evaluate SR methods for multispectral satellite imagery, focusing on a diffusion-based approach (EDiffSR) adapted for multispectral data and comparing it to a conventional convolutional neural network (RCAN). Using paired datasets with 8 spectral bands at medium and high spatial resolutions, including Sentinel-2 and VENμS mission imagery, we demonstrate the effectiveness of these methods in restoring fine-grained spatial details. To validate the practical utility of SR-enhanced imagery, we apply it to a forest species classification task involving 16 tree types. We then combine original medium-resolution data and superresolution images to create a more informative feature space preserving spectral patterns and adjusting fine-grained details. Our experiments show that AI-driven SR, particularly diffusion models, can significantly improve input data quality for ecological analysis, offering a viable pathway to reduce sensor costs without compromising the accuracy of geospatial monitoring.