Abstract
Sea-surface petroleum pollution is observed as "oil slicks" (i.e., "oil spills" or "oil seeps") and can be confused with "look-alike slicks" (i.e., environmental phenomena, such as low-wind speed, upwelling conditions, chlorophyll, etc.) in synthetic aperture radar (SAR) measurements, the most proficient satellite sensor to detect mineral oil on the sea surface. Even though machine learning (ML) has become widely used to classify remotely-sensed petroleum signatures, few papers have been published comparing various ML methods to distinguish spills from look-alikes. Our research fills this gap by comparing and evaluating six traditional techniques: simple (naive Bayes (NB), K-nearest neighbor (KNN), decision trees (DT)) and advanced (random forest (RF), support vector machine (SVM), artificial neural network (ANN)) applied to different combinations of satellite-retrieved attributes. 36 ML algorithms were used to discriminate "ocean-slick signatures" (spills versus look-alikes) with ten-times repeated random subsampling cross validation (70-30 train-test partition). Our results found that the best algorithm (ANN: 90%) was >20% more effective than the least accurate one (DT: similar to 68%). Our empirical ML observations contribute to both scientific ocean remote-sensing research and to oil and gas industry activities, in that: (i) most techniques were superior when morphological information and Meteorological and Oceanographic (MetOc) parameters were included together, and less accurate when these variables were used separately; (ii) the algorithms with the better performance used more variables (without feature selection), while lower accuracy algorithms were those that used fewer variables (with feature selection); (iii) we created algorithms more effective than those of benchmark-past studies that used linear discriminant analysis (LDA: similar to 85%) on the same dataset; and (iv) accurate algorithms can assist in finding new offshore fossil fuel discoveries (i.e., misclassification reduction).
