AO9783 - GENERATIVE ARTIFICIAL INTELLIGENCE FOR HIGH PERFORMING INVERSION MODELS - EXPRO +
Open Date: 15/03/2019
Closing Date: 27/05/2019 13:00:00
Reference Nr.: 19.1ET.06
Prog. Ref.: Technology Developme
Budget Ref.: E/0901-01 - Technology Developme
Special Prov.: BE+DK+FR+DE+IT+NL+ES+SE+CH+GB+IE+AT+NO+FI+PT+GR+LU+CZ+RO+PL+EE+HU
Tender Type: C
Price Range: 200-500 KEURO
Techology Domains: Electromagnetic Technologies and Techniques / Wave Interaction and Propagation / Wave Propagation Space System Software / Earth Observation Payload Data Exploitation / Data and Information Processing and Exploitation
Directorate: Directorate of Tech, Eng. & Quality
Department: Electrical Department
Division: RF Payloads & Technology Division
Contract Officer: Ferreol, Audrey
Industrial Policy Measure: C3 - Activities restricted to SMEs & R&D organisations, prefe...
The scarcity of adequate training data sets is a limiting factor for the complete exploitation of data acquired by Earth observation satellites. This data should be used to produce economically and socially valuable information when scientific inversion models are applied. However these models need to be trained with quality controlled and labelled datasets encountering the following issues: - Training data is scarce, costly and difficult to produce. - Labeling is difficult because instruments do not measure directly the parameters of interest (e.g. We get radiance but we are interested in vegetation parameters) - Most of AI experts do not have capabilities, knowledge and tools to execute EO simulations The "generative" Artificial Intelligence might provide a first solution. Not yet applied for space, these new kind of AI is gaining popularity and attracts massive investments in finance, drugs discovery and healthcare. In particular, the Generative Adversarial Neural Networks (GAN) look promising on generating drugs with desired medical and molecular structures. The novelty of this technique is in the ability of adjusting and learning from combinations of known features and to synthetically generate new ones according to some initial constraints. The activity will develop algorithms based on GAN able to enrich in quantity and in quality particular data sets from a limited number of initial EO data. Main tasks: - Review of the state-of-the art and selection of test cases - Select and collect data sets, from EO simulations/campaign and tentatively from ionosphere measurements - Derivation of EO focused GAN - Produce and evaluate the augmented data sets of signatures - Develop Deep Learning inversion models to be used as benchmark tools for the database - Apply DL inversion models to test the quality of the database enriched by GAN Procurement Policy: C(3) = Activity restricted to SMEs RD Entities. For additional information please go to EMITS news "Industrial Policy measures for non-primes, SMEs and RD entities in ESA programmes".