SRCNet: Seminal Image Representation Collaborative Network for Oil Spill Segmentation in SAR Imagery

Effective oil spill segmentation in Synthetic ApertureRadar (SAR) images is critical for marine oil pollution cleanup,and proper image representation contributes to effective learningfor accurate oil spill segmentation. In this paper, we propose aneffective oil spill segmentation network named SRCNet, which isconstructed  by  leveraging  seminal  SAR  image  representation  toempower  the  learning  capability  of  the  proposed  segmentationnetwork for accurate oil spill segmentation. Specifically, the imagerepresentation  utilised  in  our proposed  SRCNet originates  fromSAR  imagery,  modelling  with  the  internal  characteristics  of  oilspill SAR image data, which therefore promotes effective learningfor accurate oil spill segmentation in the training process. Besides,to  conduct  enhanced  oil  spill  segmentation,  we  construct  theproposed  SRCNet  with  a  pair  of  deep  neural  nets  that  work  ina  competition  manner,  where  one  neural  net  strives  to  produceaccurate  oil  spill  segmentation  maps  by  drawing  samples  fromthe  collaborated  seminal  image  representation,  while  the  othertries  its  best  to  distinguish  between  the  produced  and  the  truesegmentations. It is the competition and the image representationcollaborated that drives the proposed SRCNet to operate accurateoil  spill  segmentation  efficiently  with  small  amount  of  trainingdata. This establishes an economical and efficient way for oil spillsegmentation. Additionally, to further improve the segmentationperformance of the proposed SRCNet, a regularisation term thatpenalises the segmentation loss is devised, which encourages theproduced  segmentation  to  approach  the  ground-truth  segmen-tation,  promoting  the  segmentation  capability  of  the  proposedSRCNet  for  accurate  oil  spill  segmentation.  Experimental  eval-uations  from  different  metrics  validate  the  effectiveness  of  theproposed SRCNet for oil spill segmentation.