Towards Explainable AI - Trustworthy Super-Resolution (OpenSR)

Satellite remote sensing is important for the monitoring of human impact on the planet (e.g. deforestation, biomass estimation), the mapping and monitoring of natural resources (e.g. inland water streams, crops), and urban development (e.g. roads network, land use/land cover). These all depend on reliable satellite imagery to bridge the observation gap endemic in in-situ data.  

This project aims to use OpenSR to bring robust, accountable, and scalable multi-spectral super-resolution techniques to the Earth Observation (EO) community for the ubiquitous L2 and L3 pre-processing of the Sentinel-2 (S2) revisits archive. OpenSR will generate SR and xAI tools, along a WebGIS platform for illustration in challenging EO applications. 

Super-resolution (SR) is a nascent technology and the roadmap to maturity will require insights from many disciplines. Super-resolution is not just about image generation, but also degradation: how much is lost in pixelation.  

To shift the public perception on the safety of SR-S2 products, this project will provide uncertainty and quality metrics along with the SR products, and establish and disseminate best practices through new methods and tools that will be open to everyone. 

This OpenSR proposal introduces many innovative elements: 

• State of the art SR solutions especially adapted to Sentinel data, reproducible and computationally efficient code. 
• A full framework of perceptual metrics and information criteria for model evaluation. 
• A toolbox for xAI methods that serve attribution of features, model predictions and saliency maps, which collectively allow for scrutinising the SR products objectively. 
• An operational WebGIS platform to showcase challenging applications of SR and XAI in case studies involving coastal biodiversity, fire risk assessment, solar energy and urbanisation. 
• Harmonisation of SR products to Sentinel-2. 
• Documentation and synthesis of breakthroughs in techniques for SOTA SR and XAI, as well as recommendations, lessons learned and good practices.