Description :
The overwhelming amount of multi-spectral EO images acquired consistently by Sentinel-2 has stimulated the development of new applications, some of them with viable long-term business interests. The interest in free and open data for routine monitoring applications is growing exponentially. However, the medium spatial resolution of most free and open EO data archives is a strong limitation for several applications that typically require Very High Resolution imagery (VHR). At the same time, the progress in Computer Vision, mainly driven by Deep Learning (DL) approaches, has greatly accelerated lately, and with it the ability Super Resolution (SR) algorithms.
There is however little consensus regarding the real value of SR algorithms in the Earth Observation domain. When ingesting scientifically calibrated Sentinel-2 measurements into a Deep Learning “black box”, can we expect trustworthy Super Resolution products that maintain the same level of data quality?
This agora aims to demystify the use and interest of Deep Learning techniques for Sentinel-2 Super Resolution: Is this all just another hype or is there really some hope to advance EO technologies and applications?
This forum will discuss the following questions:
• What are the limits of feasibility: Does it really make sense to resolve 10m Sentinel-2 data to 1-3m meter resolution?
• How to train SR algorithms integrating laws of remote sensing physics?
• Do we have enough dataset available to train SR algorithms? What is missing?
• What are the current trends in terms of DL architectures? Single Image versus Multi-Image SR?
• Are there already some success stories with end-user adoption?
• What is the robustness of the current SR techniques against hallucinations / artefacts / pixel fakeness?
• What are suitable quality metrics and quality assurance procedures for SR in EO?
• What should be the main research topics for Super Resolution Sentinel-2 in the years to come?
• Should ESA facilitate an inter-comparison exercise for super-resolution algorithms? What are the key points to consider?

Speakers:
-Freddie Kalaitzis, Univ. Oxford
-Julien Michel, CESBIO/CNES
-Jakub Nalepa, Kplabs
-Yosef Akhtman, Altyn