On the Efficiency of Artificial Neural Networks for Detecting Natural Oil slicks on Copernicus Sentinel-1 Imagery

Abstract:

Synthetic Aperture Radar (SAR) imagery is the preferred data type for the detection and delineation of offshore oil slicks formed following the discharge of oil through human activities or natural occurrences. The preference has been fueled mainly by the inherent ability of the radar signal for cloud penetration, the large area coverage and the good contrast between dark surfactant slicks and the brighter surrounding sea surface on SAR backscatter images.

Discriminating between areas of high and low backscatter on a SAR image has proven to be a relatively straightforward problem for supervised classification algorithms, which are often used in operational scenarios in conjunction with ancillary datasets (e.g wind parameters). The results are often satisfactory with a good degree of accuracy in detecting dark slick formations and differentiating them from similar structures (also referred as “look-alikes”). However, the problem remains in that many false positives are detected, meaning that a final decision for classifying the slick’s composition and provenience is must be made by a trained human operator, adding to the time and resource consumption of the process.

With recent hardware advancements, attention has shifted towards adopting deep neural networks for analyzing satellite imagery. Convolutional Neural Networks (CNNs) and Generative Adversarial Neural Networks (GANs) have been successfully used for segmenting dark formations and extracting classified oil slicks from SAR imagery. Part of their success is due to the ability to use relatively limited amounts of data and perform tasks such as transfer learning, which is useful when considering the ability of the computer algorithm to distinguish between diverse meteo-ocean phenomena footprints. Nevertheless, deep neural networks require significant computational resources whilst providing similar classification accuracies.

In this study, we present the results on the development of an optimized automatic detection algorithm for natural oil slick detection using high-resolution, open-access Copernicus Sentinel-1 SAR imagery. This includes a comparison of different classification approaches, with the strengths and shortcomings of each type of technique analyzed in relation to the scenario in order to help optimize the final architecture of the algorithm. A case study over a set of known seepage sites and potential candidate sites in the Black Sea will be presented.