Satellite images: when vectorial rhymes with large-scale
Yuliya Tarabalka - Inria - Gilles Scagnelli
Yuliya Tarabalka, a researcher with the Titane team at the Inria Sophia Antipolis-Méditerranée centre, is the winner of the ANR JCJC (French National Research Agency's Young Researchers) grant. Her project for a compact and generic representation of satellite images opens the door to numerous applications.
Every year, over a hundred observation satellites are put into orbit to produce new images of the Earth, which are increasingly rich in content thanks to sensors that are more and more sophisticated. As a result, between them the French system Pléiades twin satellites launched in 2011 produce thousands of images daily, with a resolution of 70cm/pixel, covering the entire Earth's surface. DigitalGlobe's American satellites go even further, as they are capable of providing images with details in the order of 30cm in a vast range of spectral bands. But what happens with these petabytes of data produced each day above our heads? It should be noted that the majority of them are never used, due to a lack of available resources in order to extract knowledge from unstructured raw files.
Numerous researchers are currently endeavouring to find a solution to this state of affairs, including Yuliya Tarabalka, a specialist in the automatic analysis of remote sensing images. "Following a post-doc at NASA and the French government space agency CNES, I first of all joined the Ayin team at Inria - which no longer exists - before joining Titane in 2015, she explains. Over the years, I felt a growing need to develop my own research project. As a result, in 2016, after the initial industry transfer experience with the CNES, I have initiated my search for funding." This was a success as, in the space of six months in 2017, the researcher was awarded an ANR JCJC grant for her EPITOME project, as well as funding from other partners (the CNES and Acri-ST, on the one hand and, on the other hand, Thales and the Géoazur laboratory) who enabled the launch of three PhD theses.
A bridge between deep learning and geometric modelling
And so, in total, there is a team of eight people looking to take up the challenges of information extraction from large-scale satellite data. "The name of the ANR project is no coincidence, Yuliya Tarabalka adds. If, officially, EPITOME is the acronym of 'Efficient rePresentation TO structure large scale satellite iMagEs', it is also an ancient Greek word meaning 'abridgement' and describes a summary of a work. This sums up perfectly what we want to do: design a reliable and compact generic representation for satellite images." In concrete terms, the team is moving towards a multiresolution vector-based representation, liable to provide different significant levels of detail, depending on the resolution. In order to do this, it will bridge the gap between advanced machine learning and geometric modeling tools to devise algorithms that would provide representations containing the essential information from the raw files provided by the satellite sensors.
Monitoring...and why not predicting?
The potential prospects for this work are numerous, and in varied fields such as urban planning, precision agriculture, the monitoring of systems and terrestrial resources or the monitoring of natural or anthropogenic disasters. "In the long term, our research could also contribute to the construction of better prediction models for natural disasters. Moreover, that is what led me to collaborate with the Géoazur laboratory within the context of another ANR prizewinning project dedicated to the dynamics of seismic faults and earthquakes, évoque Yuliya Tarabalka en conclusion.
Deep learning at the service of image quality
In parallel with its research on effective representations, the team led by Yuliya Tarabalka is also working on the development of a new optimisation method for the quality of satellite images, also based on deep learning. In practical terms, this means optimising the pansharpening techniques that aim to produce very high-resolution colour images by merging two types of data produced by the sensors: very high-definition panchromatic (grayscale) images on the one hand and, on the other hand, lower resolution colour images. A major challenge, since the 'pansharpened' images currently used do not lend themselves to a reliable classification of their content, thereby naturally limiting their use.