oil spill; look-alike; support vector machine; convolutional neural networking; multilabel classification, variational autoencoder.
The search for increasingly intelligent and automated solutions has generated a growing demand for digitalization in various sectors. It is also noticeable that society has become more aware of the need to deal with climate change and environmental risks in recent decades. The oceans, one of the planet's main ecosystems, face several threats to their balance, with oil spills being one of the most serious. This study proposes to use machine learning as a tool for quickly identifying these leaks with the aim of minimizing the damage and associated costs. Three models are explored for this purpose: Support Vector Machines (SVM), a very effective algorithm in classification tasks, Convolutional Neural Networks (CNN), which are highly specialized in computer vision tasks, and Variational Autoencoder (VAE), method recognized for its ability to detect anomalies. The first two models were built to perform a multi-label classification of images in a contextualized approach, taking into account elements such as coastal areas, ships and look-alikes. Look-alikes represent a challenge to the ultimate objective of this research, since they appear as dark spots in images captured by satellites, similar to oil spill spots, however, their origin is not related to human activities. The perspective of multi-label classification is crucial in the quest to make targeted decisions and avoid false positives. CNN stood out for its ability to identify multiple multi-label tags. SVM, on the other hand, tends to generate higher hit proportions for each specific tag. And the VAE results demonstrate a highly accurate reconstruction. In the end, it is expected that these models can act in an integrated way in a leak identification system.