Fast and Automatic Oil Spill Segmentation in SAR images

1

• Fast and Automatic Oil Spill Segmentation in SAR
  images
• F. Galland1, O. Germain2, N. Bertaux1 and Ph.
  Refregier1
• 1 Fresnel Institute, UMR CNRS 6133, Marseille,
  France, http//www.fresnel.fr/phyti
• 2 Starlab, Barcelona, Spain, http//www.starlab.es

2
Introduction

• Issue
• Any oil spill detection scheme would need a step
  of candidate identification, i.e. a
  pre-selection of dark spots in the image. This is
  is the purpose of image segmentation
• Difficult task, often carried out in a supervised
  manner the user selects the region of interest,
  set thresholds, refine spill boundaries, ...
• Efficient, fast and automatic segmentation
  algorithms are therefore needed to improve oil
  spill detection
• Proposed method
• Method based on Maximum Likelihood (ML), Minimum
  Description Length (MDL) and Active Contours (AC)
• Fully automatic. The only input parameter is the
  required number of segmentation labels
• Fast less than 10 s on a 2.8 GHz PC for a
  600x900 image

3
Overview of the method
Optimal segmentation in the MDL sense corresponds
to the best compression with a Shannon code. This
principle requires

• a probabilistic image model
• the definition of the code length associated to
  a given segmentation
• a code length minimization scheme

4
Probabilistic image model
noisy image
Probabilistic model tessellation of statistical
homogeneous regions where the intensity follows a
parametric PDF (for SAR image, Gamma PDF is a
good speckle model)
5
Polygonal grid model
Parametric estimator of the partition
grid with arbitrary topology and number of
nodes defining homogenous regions
6
MDL criterion
Grid encoding

• With
• N Nb of pixels in the image
• R Nb of regions in the grid
• n, p, mx and my descriptors of the polygonal
  grid
• Nr Nb of pixels in the region r
• alpha Nb of parameters in the image intensity
  PDF (1 for a Gamma distribution)
• L likelihood of region r intensity sample
  (Assumed PDF is a Gamma distribution)

No free parameter (fully automatic)
7
Criterion minimization
Start with a regular grid and
1) Merge regions to estimate the number of
regions, 2) Move the nodes to estimate their
location, 3) Suppress nodes to estimate their
number, 4) Go to (1) until convergence.
Scheme driven by the minimization of the MDL
criterion.
8
Fast implementation
Need to compute statistics of the regions (number
of pixels, the mean and variance) after each
modification of the grid.
Time consuming
To solve this problem 2D summations are replaced
by 1D contour summation
9
Some examples
Example on ENVISAT-ASAR images of the Prestige
disaster, (Galicia, Spain, november 2002)
10
ASAR-Galicia1 (581x801)
Three class segmentation(Low-Medium-High
backscatter)
11
(No Transcript)
12
(No Transcript)
13
(No Transcript)
14
ASAR-Galicia2 (598x875)
Three class segmentation(Low-Medium-High
backscatter)
15
(No Transcript)
16
(No Transcript)
17
ASAR-Galicia3 (601x878)
Two class segmentation(Low-High backscatter)
18
(No Transcript)
19
(No Transcript)
20
(No Transcript)
21
Conclusion and perspective

• This segmentation algorithm is highly relevant
  for the SAR oil spill detection problem. Could be
  integrated into an oil spill classifier to
  improve the step of candidate identification.
• Perspective 1 test the method on a large
  database (on the way)
• Perspective 2 identify other SAR applications
  (agricultural classification, cartography, )
• Perspective 3 adapt the method to other EO
  sensors (ex MERIS with HAB application)

22
Bibliography

• F. Galland, N. Bertaux and Ph. Refregier. Minimum
  Description Length SAR Image Segmentation, IEEE
  IP 12-9, 2003
• F. Galland, N. Bertaux and Ph. Refregier. Merge,
  Move and Remove MDL Segmentation for SAR Images,
  Proc. of ACIVS, Ghent, Belgium, Sept. 2002.