142
M. Vespe et al.
2007). The degree of confidence of such identification is proportional to the time lag
between the relevant information and the sensors intrinsic uncertainties. Ship detection is sometimes used (Solberg et al., 1999) as additional information to increase
the likelihood of the detection being an oil spill, although this does not increase
the probability of true positive. Conversely, connected ship and oil spill events shall
raise the alert level of the detection since it can be thought of as a relevant factor for
follow-up activities concerning the prosecution of potential polluters.
8.3.5 Data Fusion
The oil spill features extracted from SAR images can be combined with the ancillary
data described in the previous section using data fusion algorithms. This process
is expected to augment the “reliability” of the detection simply by adding into the
automatic application the information related to the likelihood of the specific context
to present oil spills and look-alikes. In this section, a possible way to approach data
fusion of SAR and ancillary data for oil spill detection is presented focussing on
the detection of operational oil spills, i.e. oily bilge water, deliberate tank washing
residues, and unsegregated ballast water discharge.
The ancillary dataset can be organised into homogeneous tiles presenting a certain degree of spatial uniformity in terms of information content. Artificial Neural
Networks can then be used to ultimately link the SAR data to the ancillary data, ultimately leading to the estimation of a reliability index of the oil spill classification.
This is based on the vector X whose elements X i are in the range [0,1] and represent
the different levels of information related to the area of interest. Specifically, X 1 is
the oil spill classification level derived from SAR image processing only. The particular classification algorithm implemented follows a fuzzy logic approach, where
the set of fuzzy rules correspond to each detected feature. The next element of the
information vector, X 2 , is based on the homogeneity of the backscattering surrounding the dark area, measured by the standard deviation of the NRCS data. Reduced
backscatter homogeneity would increase the likelihood of observing look-alikes.
However, high data homogeneity, in conjunction with low wind conditions, still
decreases the detection capabilities. This shows the non-linear relationship between
the elements of the vector X in the evaluation of the final output, motivating the
selection of associative mapping methodology for the data fusion algorithm.
The ancillary data information content (X 3 , X 4 and X 5 ) is then mapped using
empirical functions. For instance, the wind intensity information is retrieved from
QuikSCAT data and processed according to an empirical function taking into
account that:
• A threshold at 2 m/s is set since it is the minimum wind speed to generate gravitycapillary waves,
• 15 m/s can be thought of as the wind speed upper bound for oil spill detectability,
• At moderate to high wind speed conditions, the probability of observing lookalike decreases (Pavlakis et al., 2001).
M. Vespe et al.
2007). The degree of confidence of such identification is proportional to the time lag
between the relevant information and the sensors intrinsic uncertainties. Ship detection is sometimes used (Solberg et al., 1999) as additional information to increase
the likelihood of the detection being an oil spill, although this does not increase
the probability of true positive. Conversely, connected ship and oil spill events shall
raise the alert level of the detection since it can be thought of as a relevant factor for
follow-up activities concerning the prosecution of potential polluters.
8.3.5 Data Fusion
The oil spill features extracted from SAR images can be combined with the ancillary
data described in the previous section using data fusion algorithms. This process
is expected to augment the “reliability” of the detection simply by adding into the
automatic application the information related to the likelihood of the specific context
to present oil spills and look-alikes. In this section, a possible way to approach data
fusion of SAR and ancillary data for oil spill detection is presented focussing on
the detection of operational oil spills, i.e. oily bilge water, deliberate tank washing
residues, and unsegregated ballast water discharge.
The ancillary dataset can be organised into homogeneous tiles presenting a certain degree of spatial uniformity in terms of information content. Artificial Neural
Networks can then be used to ultimately link the SAR data to the ancillary data, ultimately leading to the estimation of a reliability index of the oil spill classification.
This is based on the vector X whose elements X i are in the range [0,1] and represent
the different levels of information related to the area of interest. Specifically, X 1 is
the oil spill classification level derived from SAR image processing only. The particular classification algorithm implemented follows a fuzzy logic approach, where
the set of fuzzy rules correspond to each detected feature. The next element of the
information vector, X 2 , is based on the homogeneity of the backscattering surrounding the dark area, measured by the standard deviation of the NRCS data. Reduced
backscatter homogeneity would increase the likelihood of observing look-alikes.
However, high data homogeneity, in conjunction with low wind conditions, still
decreases the detection capabilities. This shows the non-linear relationship between
the elements of the vector X in the evaluation of the final output, motivating the
selection of associative mapping methodology for the data fusion algorithm.
The ancillary data information content (X 3 , X 4 and X 5 ) is then mapped using
empirical functions. For instance, the wind intensity information is retrieved from
QuikSCAT data and processed according to an empirical function taking into
account that:
• A threshold at 2 m/s is set since it is the minimum wind speed to generate gravitycapillary waves,
• 15 m/s can be thought of as the wind speed upper bound for oil spill detectability,
• At moderate to high wind speed conditions, the probability of observing lookalike decreases (Pavlakis et al., 2001).