International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
images to demonstrate the generalization efficiency of oil 
slick classification algorithm. 
STUDY AREA and DATA USED 
The Deepwater Horizon is a 9-year-old semi-submersible 
mobile offshore drilling unit that could operate in waters up 
to 2,400 m deep and drill down to 9,100 m. At the time of the 
explosion, it was drilling an exploratory well at a water depth 
of approximately 1,500 m in the Gulf of Mexico about 66 km 
off the Louisiana coast of the United States (Figure 1). 
  
ALABAMA 
UNITED STATES LOUISIANE MISSISSIPPI 
   
  
Suton Rouge, 
Dries 
   
    
Daepwater Horizon Tf. 
oil rig 
   
  
Gulf of Mexico 
   
      
   
Berids Yucatan 
zt Peninsula 
  
© 2010ER fa | 
  
  
  
Figure 1. Map of the study area (Encyclopaedia Britannica). 
In this study, as an active sensing system, 2 SAR (Radarsat -2 
and ALOS PALSAR) images acquired on the same date were 
used in the analyses (Table 1). 
Table 1. The characteristics of the radar dataset used. 
  
  
  
  
  
  
  
  
  
  
  
  
  
Spatial 
Satellite Ace. Beam Res. Band | Pol. Cov. 
date mode (sq km) 
(m) 
1 
ScanSAR HH 300 x 
Radarsat-2 | May Narrow 50 C HV 300 
2010 
Alos 
Pals May FBS 9 L HH 70 x 70 
sat 2010 
METHODOLOGY 
Radar images have an advantage for oil spill detection due to 
the dampening effect of oil on capillary waves causing them 
to be detectable as black patches on images. However, SAR 
images must be processed carefully since the dark areas 
might occur because of some natural phenomena without oil 
like smooth water (low wind areas), organic films, wind front 
areas, areas sheltered by land, rain cells, grease ice, internal 
waves and shallow bathymetric features (Sabins, 1997; 
Alpers et al, 1991; Hovland et al., 1994). The procedure 
steps of oil spill detection in SAR data can be generalized as 
segmentation (dark object extraction), feature extraction and 
classification (determination oil) stages (Pavlakis et al., 2001; 
Brekke and Solberg, 2005; Solberg et al., 2007; Shi et al., 
2008; Topouzelis et al., 2009). 
Before the procedure steps, all SAR imagery used in this 
study was calibrated to obtain Normalized Radar Cross 
Section (NRCS, 60) values. The NRCS calibrations correct 
68 
the radar imagery for the effects of antenna pattern and local 
incidence angle. Calibration of ALOS SAR images was done 
using Gamma Interferometric SAR Processor, while the 
RADARSAT-2 images were calibrated using the look-up 
table provided in the products meta-data. 
— Segmentation 
Segmentation can be defined as the subdivision of original 
image into small and homogeneous regions that correspond 
to individual surfaces, objects, or natural parts of objects 
(Figure 2). Since the segments, composed by pixels, provide 
spectral and geometric information, the parameters extracted 
from shape (area, length, etc.) and neighbor's relationship can 
be included on classification strategies to promote better 
discrimination of objects with similar spectral responses. In 
this study, segmentation of dark objects from the others was 
done by object-based classification using eCognition 
software. 
      
A 
) and 
Figure 2. Objects from ALOS PALSAR (left 
RADARSAT-2 (right). 
In the segmentation process, “multi-resolution segmentation” 
procedure of object based image analysis concept (OBIA) 
was applied. For the SAR images used, scale factor was 
chosen very small to obtain correct oil spill features in the 
images; (25 for ALOS and 15 for RADARSAT-2). After 
segmentation process, classification process was done by 
using class features defined in class hierarchy. For 
classifications of the oil objects, mean layer value of each 
object was used with a defined threshold. By using process 
based classification tree, all processes were programmed in a 
suitable order and then the mask image of oil objects was 
exported as a final step. 
— Feature Extraction 
In general, the features of oil spill are its geometric 
characteristics (such as area, perimeter, complexity), physical 
behaviours (such as mean, standard deviation or max 
backscatter value) and context (such as number of other dark 
formations, presence of ships and proximity to and route of 
ships) (Del Frate et al., 2000, Karathanassi et al., 2006, 
Brekke and Solberg, 2005). Moreover, Haralick textural 
features computed based on Haralick's cooccurence matrix 
are widely used (Haralick, 1979), such as angular second 
moment, contrast, correlation, dissimilarity, entropy, mean 
and standard deviation (Assilzadeh and Mansor, 2001). 
The widely used textural features are; 
- SP2 (Shape factor 2) describes the general shape of the 
object which is also called as ‘first invariant planar 
moment’, ‘form factor’, and ‘asymmetry’. 
- BSd (Background Standard deviation) is the standard 
deviation of the intensity values of the pixels belonging