IRT Saint Exupéry boosted your data experience at Big Data from Space 2017

IRT Saint Exupéry was partner of the 3rd edition of the 2017 Conference on Big Data from Space (BiDS’17) held from 28 to 30 Nov. at Toulouse. It was co-organised by ESA, the Joint Research Centre of the European Commission, and the European Union Satellite Centre (SatCen) and hosted by the CNES. This conference has brought together more than 300 researchers, engineers, users, infrastructure and service providers from 40 countries, interested in exploiting Big Data from Space and who presented 90 papers.

On its booth, the Intelligent Systems & Data Competence Centre team of IRT Saint Exupéry has successfully presented 5 demonstrations and launched innovative services to boost data experience through its Citadel Platform. IRT Saint Exupéry gathers all the process pipelines, from data collection to benchmarking, aiming to boost application development, research in artificial intelligence, and data processing for all-sizes actors.


Today’s Earth observation systems are facing major evolutions: more and more satellites, customers, requests. Thus, there is a need to develop new planning systems which can provide mission plans for satellite constellations in a bounded time (<5 minutes) and providing more reactivity to the whole system. To answer those requirements, we rely on adaptive multi-agent’s systems introducing dynamic planning and develop the ATLAS (adaptive satellite planning for dynamic earth observation) planning system. ATLAS has been tested on real spot 6/7 & Pleiades scenarios. Moreover, we have built scenarios with varying number of satellites per constellation to demonstrate the ATLAS scalability.


There is a growing operational demand for the acquisition of large area zones (up to several millions of km²) through high resolution observation satellites. Such large coverage projects require a huge number of elementary acquisitions from satellites whenever they fly over the zone, and take up to several weeks or months to reach completion. The objective for large coverage management is to speed up the completion of large coverage projects through the long-term optimization of the acquisition strategy and the combination of multiple Earth Observation (EO) systems. Large coverage management is relevant for system exploitation like Astoterra and Pléiades. Control policies are integrated into existing systems while future EO systems benefit from decisional autonomy through reinforcement learning.


We have investigated solutions to the imagery classification problem in a massive data context. In this case massive means more than 200.000 images. Our work heavily relies on the most recent machine learning technologies giving computers the ability to learn from series of examples. The quantity of data implies the use of scalable and specific technologies such as cloud computing during the design of such framework. Learning as well as testing are based on real spot 6/7 image data sets. The processing chains have been evaluated on tera-scaled Spot 6 database. Two disruptive applications are developed: an automatic cloud detector and an automatic land cover classifier.


Within the class of object detection problems, we have built a specific use case whose objective is to enhance existing oil slick detection using a massive radar imagery dataset. We evaluate a machine learning approach to answer this problem. It is implemented in a cloud distributed framework to manage the quantity of data. Learning as well as testing are based on real ENVISAT ASAR and Sentinel 1 image data sets. In order to increase the learning data set with an affordable contribution of human operators, we make use of simulations and characterize the achievable performances. Uses cases include the automatically pointing geographical areas where there are cumulative oil slick detections (natural seep sources).


Sequential observation data are now made available thanks to existing systems like Spot 6/7, Pleïades, Cosmo-skymed, Terrasar and Sentinel. This tendency will be reinforced in a near future with the development of constellations or geostationary observation systems. In this dynamic context, the change detection class of algorithms appears to be well suited to ease the end user task: focus on significant changes. We developed new concepts of change detection chains based on dictionary learning and deep learning state of the art methods, with leading edge robustness characteristics. It has been tested on real Spot 6/7 & Pleiades temporal series. It can be applied to big cities expansion, defense, oil and gas or harbor surveillance.

IRT Saint Exupéry boosted your data experience at Big Data from Space 2017
Scroll to top