Speckle reduction, High and Very High resolution SAR imagery
A new generation of SAR sensors is born in 2008 and 2009 with the launch of TerraSAR-X, CosmoSkyMed, ALOS, RadarSat-2. Their resolution has considerably improved passing from 10m to 1m. Today, commercial sensors like those from Iceye or Capella also acquire very high resolution SAR images. In parallel, aerial sensors like SETHI or RAMSES-NG or UAV (ONERA), allow the acquisition of very high resolution data. To deal with these new images, different works are developped:
Speckle reduction for multi-channel-SAR data: many methods are investigated : deep learning methods (SAR2SAR, MERLIN), patch-based approaches and previously markovian modeling with different priors and optimization tools (graph-cut, ... ). Speckle correlation and side-lobe reduction are also studied through spectrum exploitation. Extension to multi-channel SAR data like polarimetric or interferometric acquisitions are also developed (PolSAR2PolSAR, InSAR2InSAR, MuChaPro).
Statistical modeling of SAR signals: new distributions can be derived using Mellin transforms ; besides log-cumulant approaches give better estimates of the parameters ; many applications based on Fisher and Meijer distributions have been developped (filtering, classification, etc.).
Local descriptors for SAR data: a descriptor adapted to SAR statistics and multiplicative noise has been derived. It is called SAR-SIFT and has well grounded key-point and descriptor definition based on log ratio edge detector.
Dictionaries: dictionaries of SAR amplitude signals are studied with the introduction of some geometric or radiometric invariances.
Multi-temporal series processing and change detection
This axis has gained an increased importance in the last few years with a huge amount of data available (increase of the number of sensors, constellations of sensors, sensor agility). The launch of Sentinel-1 and 2 and ESA data policy have provided long and stable temporal series with middle resolution. New issues on data combination, fusion, change detection, mining on these series have emerged. The SWOT mission launched in 2023 also provides multi-temporal single-pass interferometrci acquisitions.
Multi-temporal estimation: different approaches have been proposed to improve SAR signal estimation by exploitation the multi-temporal information. Multi-temporal MERLIN (based on deep learning) and RABASAR method (based on a "super-image") provide state of the art multi-temporal denoising results.
Multi-temporal change detection Different approaches based on statistical tests using Meijer distribution (MIMOSA) or generalized likelihood ratio tests (NORCAMA) have been proposed. Another method relying on local descriptor and a contrario theory is also developed. The exploitation of RABASAR results allow the use of classical change detection approaches. New approaches based on speckle reduction and uncertainty estimation are also investigated.
3D reconstruction: multi-view, interferometry and tomography
Although interferometric and radargrammetric data give an information about ground elevation or movement, there are many difficulties to exploit them (speckle, phase wrapping, shadows, ...). Tomographic approaches exploiting a set of images with different baselines allows 3D reconstruction.
3D reconstruction of buildings : developpment of a tomographic processing chain exploiting regularization prior on the urban surfaces. A graph-cut baesd framework is used to introduce iteratively surface information.
Multi-channel phase unwrapping : development of approaches for the joint unwrapping and denoising phase when multi-channel images are available (multi-baseline or multi-frequency) ; development of dedicated graph-cut based methods improving the memory requirement.
Forest reconstruction : development of deep learning approaches for the tomographic inversion of forested areas.
Fusion of optic and SAR data
The difficulty of SAR image interpretation has led to the developpment of fusion approaches using both optic and SAR data.
SAR / optic registration : This is a preliminary step, necessary to be able to merge the data. It can be done using sensor knowledge and fine feature matching or using manual ground control point. It can also be done in an automatic way by feature matching.
Joint filtering or interpretation of SAR / optic features
3D reconstruction of buildings: After the detection of building outlines in the optical image, a step to recover the elevation of buildings by likelihood maximization in the SAR data has been proposed.
Multi and hyperspectral imaging
Optical sensors acquire multiple spectral bands and hyperspectral imaging reaches more than a hundred of bands providing rich spectral signatures of the objects on the earth surface. These data raise many challenges.
Hyperspectral unmixing : Because the spatial resolution of hyperspectral images is rather low, it usually mixes some materials inside the pixels. Therefore a preliminary step is the unmixing of the material endmembers contributing to a pixel signature. Unrolling methods like LPALM and methods based on variational endoders are currently developed.
Super-resolution of hyperspectral images
Deep learning methods allowing to improve the spatial resolution of the hyperspectral images are investigated leveraging the "dead-leaves model" of image synthesis to define a supervised framework.