JOINT SUPER-RESOLUTION AND IMAGE RESTORATION FOR PLÉIADES NEO IMAGERY
Keywords: Single-Image Super-Resolution, Pléiades Neo, Remote Sensing, Deep Learning, Convolutional Neural Networks
Abstract. Modern Earth Observation optical satellite systems, such as Airbus’s Pleiades Neo (PNeo) push the boundaries of high spatial resolution by providing commercial imagery products with up to 30cm ground sampling distance (GSD). To further enhance the quality of the images, the in-space imaging system is usually complemented by on-ground image restoration processing, such as deconvolution and denoising (Latry et al., 2012). Recent advances leverage Convolutional Neural Networks (CNNs) to improve the image restoration quality (K. Zhang et al., 2021a).
Single Image Super-Resolution (SISR), or Zoom, the process of obtaining a higher resolution (HR) image from a lower resolved (LR) source, has recently gained traction for both medium resolution sensors such as Sentinel 2 (Lanaras et al., 2018) and high resolution such as Pléiades and GeoEye-1 (Zhu et al., 2020). This process further enhances the resolution of the image to improve downstream applications such as mapping (L. Zhang et al., 2021) and small objects recognition (Shermeyer and Van Etten, 2019). While SISR for remote sensing has been successfully tackled using CNNs (Rohith and Kumar, 2021) the main challenge for reaching acceptable image quality performance lies in the generation of realistic LR/HR training pairs (K. Zhang et al., 2021b). In this paper, we propose:
- a dedicated simulation chain leveraging extremely-high-resolution (EHR) aerial imagery to generate realistic 30cm Pléiades Neo images and their corresponding fully restored HR equivalent at 15cm GSD
- A residual-based CNN architecture which we train to jointly restore and zoom the images All contributions are assessed on real PNEO images.
We deployed the trained models in a production context, to enhance the full Pléiades Neo products – with a swath of 47k pixels – in an efficient and scalable manner.