Super-Resolution of Sentinel-2 Images Using a Spectral Attention Mechanism

Abstract

Many visual applications require high-resolution images for an adequate interpretation of the data stored within them. In remote sensing, the appearance of satellites such as Sentinel or Landsat has facilitated the access to data thanks to their free offer of multispectral images. However, the spatial resolution of these satellites is insufficient for many tasks. Therefore, the objective of this work is to apply deep learning techniques to increase the resolution of the Sentinel-2 Read-Green-Blue-NIR (RGBN) bands from the original 10 m to 2.5 m. This means multiplying the number of pixels in the resulting image by 4, improving the perception and visual quality. In this work, we implement a state-of-the-art residual learning-based model called Super-Resolution Residual Network (SRResNet), which we train using PlanetScope-Sentinel pairs of images. Our model, named SARNet (Spectral Attention Residual Network), incorporates Residual Channel Attention Blocks (RCAB) to improve the performance of the network and the visual quality of the results. The experiments we have carried out show that SARNet offers better results than other state-of-the-art methods.

Keywords

• remote sensing
• image super-resolution
• deep learning
• channel attention
• Sentinel-2
• PlanetScope