On the detection and segmentation of metocean features on SAR images using Deep Learning: perspectives for SWOT

• Main Authors: Longépé, Nicolas, Husson, Romain, Wang, Chen, Mouche, Alexis, Tandeo, Pierre
• Other Authors: Collecte Localisation Satellites (CLS), Institut Français de Recherche pour l'Exploitation de la Mer (IFREMER), Lab-STICC_IMTA_CID_TOMS, Laboratoire des sciences et techniques de l'information, de la communication et de la connaissance (Lab-STICC), École Nationale d'Ingénieurs de Brest (ENIB)-Université de Bretagne Sud (UBS)-Université de Brest (UBO)-École Nationale Supérieure de Techniques Avancées Bretagne (ENSTA Bretagne)-Institut Mines-Télécom Paris (IMT)-Centre National de la Recherche Scientifique (CNRS)-Université Bretagne Loire (UBL)-IMT Atlantique (IMT Atlantique), Institut Mines-Télécom Paris (IMT)-École Nationale d'Ingénieurs de Brest (ENIB)-Université de Bretagne Sud (UBS)-Université de Brest (UBO)-École Nationale Supérieure de Techniques Avancées Bretagne (ENSTA Bretagne)-Institut Mines-Télécom Paris (IMT)-Centre National de la Recherche Scientifique (CNRS)-Université Bretagne Loire (UBL)-IMT Atlantique (IMT Atlantique), Institut Mines-Télécom Paris (IMT), Département Signal et Communications (IMT Atlantique - SC), IMT Atlantique (IMT Atlantique), Institut Mines-Télécom Paris (IMT)-Institut Mines-Télécom Paris (IMT)
• Format: Conference Object
• Language: English
• Published: HAL CCSD 2019
• Subjects: [STAT.AP]Statistics [stat]/Applications [stat.AP]; [SDU.OCEAN]Sciences of the Universe [physics]/Ocean; Atmosphere; Sea ice
• Online Access: https://imt-atlantique.hal.science/hal-02156712; https://imt-atlantique.hal.science/hal-02156712/document; https://imt-atlantique.hal.science/hal-02156712/file/20190619_SWOTPoster_DLOcean.pdf

International audience The above classification model (one 20x20 km imagette = one label) is then used over a full Wide Swath S-1 image using local class estimates with a convolution approach to provide a semantic segmentation. Multi-class (and multi-label , i.e. one pixel several classes) semantic segmentation is now feasible with DL , pending starting with a good and sufficient training database, and active learning process to enrich it, To this end, annotation tools and adequate framework should be consolidated and tuned to our problematic: 1) massive processing needed to raise specific issues (for instance, "sea ice" North of Madagascar should be re-tagged as internal waves), 2) additional information from ancillary metocean data might be provided , … The entire S-1 Wave Mode archive from 2016 is being processed. Below the occurrence of each class as classified by the DL model is provided on a monthly basis. This classified database could be used as input for a systematic collocation process with SWOT data. That will 1) help to understand Ka-band near nadir imaging processes for a given phenomenon, 2) serve during the Cal/val campaign, and 3) be used to build a training database with tagged SWOT images serving also DL model. With the upcoming launch of SWOT, new Ka-band near-nadir SAR image will be produced. Whereas the legacy of ocean SAR imaging is huge for L-C-or X-band SAR sensors with intermediate incidence angle, the interaction of Ka-band near-nadir EM waves and its associated SAR image formation lead to some uncertainties on how metocean features will be imaged on SWOT image. To name a few, atmospheric fronts, ocean fronts, rain cells, convective microcells, internal waves, gravity waves, biological slicks, upwelling or wind trails are phenomena that will be imaged by SWOT. These phenomena could be a source of errors and bias for SSH products. Meanwhile, they are of potential interest for the scientific communities. In this study, we aim to propose a methodology to flag and detect these phenomena ...