ASTRID - Accompagnement spécifique des travaux de recherches et d’innovation défense

Statistical learning for multi-dimensional SAR imagery – ASTRAL

Submission summary

SAR (Synthetic Aperture Radar) imagery is based on the emission of electromagnetic microwaves that are then backscattered by the elements on the ground surface, thus allowing the measurement of their physical properties. The relative phase shift of the waves received by the radar, measured by combining pairs of images (interferometry) or a stack of images (tomography) acquired under close incidence angles, gives access to the height of the elements of the scene or their displacement. Polarimetry, which emits waves with different polarizations, gives access to richer information on the electro-magnetic properties of the backscatterers. These imaging techniques are now well established, but the full automatic exploitation of the acquired data remains difficult due to the very high speckle noise that degrades the data and the geometric distortions (related to lateral viewing: shadowing and overlay). At the same time, artificial intelligence techniques and in particular deep learning have revolutionized the field of image analysis and interpretation in the last 5 years. Some problems have made such striking advances that they can now be considered as almost solved under certain conditions (denoising, super-resolution, classification, ...). Yet, the exploitation of these approaches still requires significant research efforts in different situations: scarcity of labeled data, specific nature of the data (such as vectors of complex numbers in SAR imagery), need to characterize the reliability of the result ... A particularly active area of research at present is the articulation between physical knowledge, often available in the form of models, and data-driven learning approaches.

The general objective of this project is to develop approaches for analyzing and interpreting scenes from SAR data that differ from existing ones by including the physics of SAR within the learning technique.
It is built according to the 3 following axes. A first axis is dedicated to the representation of knowledge in deep networks, in particular by integrating the physics of SAR acquisition (complex vector data, parameters of interest in Hermitian positive definite covariance matrices), and acquisition geometry (taking into account the influence of geometry, in particular the relative positions of sensors in interferometry or tomography and the spatial relationships between objects in the scene). A second axis is devoted to learning strategies in the case of a small number of labeled data (simulation of training bases, self-supervision techniques) and to the confidence that can be associated to the results. The last axis is devoted to the development of applications for the characterization and monitoring of the urban environment in SAR imagery: the classification of urban scenes, their three-dimensional reconstruction and the detection of changes.

One of the originalities of the project lies in the choice not to be restricted to a particular application or modality of radar imagery, but rather to study generic frameworks for the representation of SAR data by exploiting the knowledge and physical constraints specific to this field. The objectives are thus twofold: on the one hand, to provide methodological contributions on the integration of physical knowledge specific to SAR imagery within the architecture of networks and their training (axes 1 and 2); on the other hand, to provide application contributions for the interpretation of urban scenes using SAR imagery (characterization, 3D reconstruction, change detection) (axis 3).

Project coordinator

Madame Florence TUPIN (Laboratoire Traitement et Communication de l'Information)

The author of this summary is the project coordinator, who is responsible for the content of this summary. The ANR declines any responsibility as for its contents.

Partner

LTCI Laboratoire Traitement et Communication de l'Information
CEDRIC CENTRE D'ETUDES ET DE RECHERCHE EN INFORMATIQUE ET COMMUNICATIONS
LabHC Laboratoire Hubert Curien
ONERA - DEMR ONERA

Help of the ANR 297,000 euros
Beginning and duration of the scientific project: - 36 Months

Useful links

Explorez notre base de projets financés

 

 

ANR makes available its datasets on funded projects, click here to find more.

Sign up for the latest news:
Subscribe to our newsletter