Unsupervised oil spill detection in SAR imagery through an estimator of local regularity
Marivi Tello(1), Carlos Lopez-Martinez(1), Jordi Mallorqui(1), Gerardo Di Martino(2), Antonio Iodice(2), Daniele Riccio(2) and Giuseppe Ruello(2)
(1) Universitat Politecnica de Catalunya, C/ Jordi Girona 1-3, 08034 Barcelona, Spain
(2) Universita di Napoli Federico II, Via Claudio 21, 80125 Napoli, Italy
Since surfactive substances damp the short-scale waves, the presence of oil in the sea surface is revealed as a dark patch in the SAR image. Hence, SAR systems are effectively suitable for oil spills identification on ocean surfaces. In fact, even if nowadays insufficient revisit time of satellites does not allow its use for permanent monitoring, SAR technology has been confirmed as one of the most appropriate for the detection of slicks.
Most often interpretation of SAR images is roughly performed manually. But this is an unacceptably slow, unpractical and hardly reproducible procedure; in order to assure its further usability, specific data mining methods are still to be developed to provide an efficient automatic interpretation of SAR data. Nevertheless, unsupervised interpretation is still troublesome, mainly because of speckle. In particular, in the context of pollution monitoring, automatic exploitation of SAR data is penalised by the incapacity of existing techniques to discriminate oil spills from look-alikes. More specifically, since SAR sensors are sensitive to changes of surface roughness, any other phenomenon producing a local damping of capillary waves, reduces the amount of energy backscattered to the radar and appears also as a dark area. Lack of wind, rain cells and banks of phytoplankton constitute the main sources of ambiguity. Therefore, the main drawback of computerised schemes is the discrimination of oil spills and look-alikes. This paper aims at shedding light on the possibility of solving the disambiguation issue.
From a computer vision point of view, algorithms designed to distinguish oil spills from false alarms in SAR images have to face two main difficulties. The first one is that the appearance of oil spills is subject to a great diversity. Therefore, the assumption of a priori models is not efficient, training of algorithms relying on neural networks is time consuming and techniques exclusively based on morphological features are not robust. The second one is that oil spills and look-alikes can present remarkable similarities at first sight.
According to these empirical observations and aiming at providing a deeper analysis of the statistical properties oil spill candidates, the objective of this paper is to provide a quantitative measure, as local as possible, of the regularity of the SAR signal. In order to do so, the use of a multiscale framework will be justified. Then the formalism of a new estimator of local texture will be introduced, based on the assumption that the decay of the wavelet transform amplitude across scales is related to the uniform and pointwise Lipschitz regularity of the signal. First, it will be confirmed through simulated images, that the parameter proposed is independent of the mean intensity value and that it is able to catch local differences of higher order moments. Then results will be evaluated of the application of this method on a set of simulated images produced by a SAR raw signal simulator which is able to generate SAR return relative to oil slicks on ocean surfaces. In particular, it will be shown that the processing technique proposed is able to discriminate dark patches simulated with the parameters corresponding to an oil covered surface from those corresponding to a lack of wind zone, even if they have the same damping. Even if more difficult to evaluate, results on a set of real images will also be analysed.
Keywords: ESA European
Space Agency - Agence spatiale europeenne,
observation de la terre, earth observation,
satellite remote sensing,
teledetection, geophysique, altimetrie, radar,
chimique atmospherique, geophysics, altimetry, radar,