DGNet: Distribution Guided Efficient Learning for Oil Spill Image Segmentation

Successful  implementation  of  oil  spill  segmentationin  Synthetic  Aperture  Radar  (SAR)  images  is  vital  for  marineenvironmental  protection.  In  this  paper,  we  develop  an  effectivesegmentation   framework   named   DGNet,   which   performs   oilspill  segmentation  by  incorporating  the  intrinsic  distribution  ofbackscatter  values  in  SAR  images.  Specifically,  our  proposedsegmentation   network   is   constructed   with   two   deep   neuralmodules  running  in  an  interactive  manner,  where  one  is  theinference module to achieve latent feature variable inference fromSAR images, and the other is the generative module to produce oilspill segmentation maps by drawing the latent feature variablesas  inputs.  Thus,  to  yield  accurate  segmentation,  we  take  intoaccount  the  intrinsic  distribution  of  backscatter  values  in  SARimages  and  embed  it  in  our  segmentation  model.  The  intrinsicdistribution originates from SAR imagery, describing the physicalcharacteristics of oil spills. In the training process, the formulatedintrinsic  distribution  guides  efficient  learning  of  optimal  latentfeature variable inference for oil spill segmentation. The efficientlearning  enables  the  training  of  our  proposed  DGNet  with  asmall  amount  of  image  data.  This  is  economically  beneficial  tooil spill segmentation where the availability of oil spill SAR imagedata is limited in practice. Additionally, benefiting from optimallatent feature variable inference, our proposed DGNet performsaccurate  oil  spill  segmentation.  We  evaluate  the  segmentationperformance of our proposed DGNet with different metrics, andexperimental evaluations demonstrate its effective segmentations