Resumo:
The appearance of oil sheens in the ocean is a challenge for companies that perform
primary oil processing on offshore platforms. After the separation of the gas, oil and water
that are present in crude oil, part of the water is returned to the oceans with a certain
content of oils and greases. The value of the total oil and greases (TOG) associated with
the values of metoceanographic variables such as: wind direction (WD), wind speed
(WS), current direction (CD), current speed (CS), wind wave direction (WWD) and peak
period (PP) create scenarios that favor or hinder the appearance of oil sheens. In Brazil,
these oil sheens can lead to sanctions for companies if they exceed 500 meters in length.
In view of this, the present work conducts a study about how such variables influence the
probability of occurrence and detection of oil sheens via satellite, as well as their extent,
applying machine learning techniques (random forest, k-nearest neighbors, artificial
neural networks, logistic regression, and support vector machines), factor analysis, design
of experiments (DOE) and the optimization algorithm desirability. The main conclusions
of the study were: (i) random forest outperformed the other analyzed classifiers and a
model whose area under the Receiver Operating Characteristic Curve (ROC curve) was
0.93 was achieved; (ii) the methodology used, combining the classifiers with the
aforementioned techniques proved to be satisfactory; (iii) the higher the values of WS,
WD and CS, the lower the probability of occurrence and detection of oil sheens, whereas
the higher the values of TOG, PP, WWD and CD the higher the value of this probability;
(iv) variables such as CS and TOG contribute positively to increasing the extension of the
oil sheens, while high values of WD, WS and PP reduce the extension of the features.
