8 Perspectives on Oil Spill Detection Using Synthetic Aperture Radar
133
of the sea surface. Several research investigations on techniques and algorithms to
detect oil spills from SAR data have been conducted in the last years (e.g. Solberg
et al., 1999; Del Frate et al., 2000; Fiscella et al., 2000; Pavlakis et al., 2001;
Karathanassi et al., 2006), as also recently reviewed by Topouzelis (2008). The typical functions implemented by SAR based oil spill detection are based on magnitude
detected products and can be summarised as follows.
1. Image pre-processing: the image is pre-processed through well known algorithms especially tailored on the particular application. For oil spill detection,
the radiometric resolution of the image is a key parameter and often is advantaged against spatial resolution using 2-D edge preserving filters (e.g. Lee, Local
Region, Frost etc). Such filters increase the Equivalent Number of Looks (ENL)
with the effect of reducing the speckle noise and therefore enhancing the contrast
of dark patches over the image. The image is commonly land-masked in order to
reduce the number of processed pixels, thus decreasing the computational burden
of the algorithm.
2. Image segmentation: the pre-processed image is then segmented in order to
extract the dark patches of interest. The segmentation process can be thought of
as a Constant False Alarm Rate (CFAR) algorithm that, on the basis of the local
sea backscattering statistics, applies a NRCS (Normalised Radar Cross Section)
threshold below which the pixels are indicated. The pixels are then organised in
adjacent sets commonly referred to as image objects.
3. Feature extraction: the image objects are analysed in detail, leading to the estimation of a set of features previously selected. Such features are related to shape,
internal texture and neighbourhood contrast. The feature selection process is
a key process for any classification algorithm. As regards oil spill classification, the problem is often reduced to the discrimination between “oil spill” and
“look-alike”. Separating these two classes in feature space requires an accurate
optimisation.
4. Classification: for each object, the values related to the selected features are
compared with a database of templates in order to evaluate the class to which
the object belongs. A preliminary screen on size of the objects is commonly
performed in order to exclude residual speckle objects that are of no interest
to the analysis. The classification function has been approached using a variety
of algorithms spanning from statistical, to associative mapping techniques. An
overview, as well as a review of the oil spill classification literature, can be found
in Brekke and Solberg (2008), and Topouzelis (2008).
The use of satellite based SAR oil spill detection has been considered taking into
account operational sensors, e.g. Radarsat-1 and 2, ERS-2, Envisat, Alos, TerrasarX and CosmoSkyMed. The primary products used are ScanSAR magnitude detected
images. However, the advances made by polarimetric oil spill detection foresee the
developing of relevant algorithms for operational use. Among them, fully polarimetric data have been demonstrated to be useful in discriminating look-alikes from
oil spills using circular polarization coherence (CPC) and polarimetric anisotropy
133
of the sea surface. Several research investigations on techniques and algorithms to
detect oil spills from SAR data have been conducted in the last years (e.g. Solberg
et al., 1999; Del Frate et al., 2000; Fiscella et al., 2000; Pavlakis et al., 2001;
Karathanassi et al., 2006), as also recently reviewed by Topouzelis (2008). The typical functions implemented by SAR based oil spill detection are based on magnitude
detected products and can be summarised as follows.
1. Image pre-processing: the image is pre-processed through well known algorithms especially tailored on the particular application. For oil spill detection,
the radiometric resolution of the image is a key parameter and often is advantaged against spatial resolution using 2-D edge preserving filters (e.g. Lee, Local
Region, Frost etc). Such filters increase the Equivalent Number of Looks (ENL)
with the effect of reducing the speckle noise and therefore enhancing the contrast
of dark patches over the image. The image is commonly land-masked in order to
reduce the number of processed pixels, thus decreasing the computational burden
of the algorithm.
2. Image segmentation: the pre-processed image is then segmented in order to
extract the dark patches of interest. The segmentation process can be thought of
as a Constant False Alarm Rate (CFAR) algorithm that, on the basis of the local
sea backscattering statistics, applies a NRCS (Normalised Radar Cross Section)
threshold below which the pixels are indicated. The pixels are then organised in
adjacent sets commonly referred to as image objects.
3. Feature extraction: the image objects are analysed in detail, leading to the estimation of a set of features previously selected. Such features are related to shape,
internal texture and neighbourhood contrast. The feature selection process is
a key process for any classification algorithm. As regards oil spill classification, the problem is often reduced to the discrimination between “oil spill” and
“look-alike”. Separating these two classes in feature space requires an accurate
optimisation.
4. Classification: for each object, the values related to the selected features are
compared with a database of templates in order to evaluate the class to which
the object belongs. A preliminary screen on size of the objects is commonly
performed in order to exclude residual speckle objects that are of no interest
to the analysis. The classification function has been approached using a variety
of algorithms spanning from statistical, to associative mapping techniques. An
overview, as well as a review of the oil spill classification literature, can be found
in Brekke and Solberg (2008), and Topouzelis (2008).
The use of satellite based SAR oil spill detection has been considered taking into
account operational sensors, e.g. Radarsat-1 and 2, ERS-2, Envisat, Alos, TerrasarX and CosmoSkyMed. The primary products used are ScanSAR magnitude detected
images. However, the advances made by polarimetric oil spill detection foresee the
developing of relevant algorithms for operational use. Among them, fully polarimetric data have been demonstrated to be useful in discriminating look-alikes from
oil spills using circular polarization coherence (CPC) and polarimetric anisotropy