| dc.description.abstract | While there is a considerable number of machine learning and anomaly detection studies as far as the maritime domain is concerned, the particular topic of ship waste and residue discharge remains an interesting but understudied problem. In this study, we explore the potential of machine learning, more specifically anomaly detection, in order to detect ship waste discharge at the North Sea, where several techniques ranging from supervised to unsupervised are investigated. One of the main problems we face in this study is the small amount of labeled data of the anomalous class in our disposition, leading to an extreme class imbalance in favor of the normal class. In order to tackle this problem, different class imbalance handling techniques are experimented with, one of them being the utilization of ensemble learning techniques. Our main contribution involves the development of a time series supervised learning pipeline that is used to discriminate between normal and anomalous behavior effectively (with performance significantly higher than chance level) and efficiently (with minimal cost and preferably real-time). The final performance results show that our model achieves a Macro Recall of 80\%, and does so at $t = 18$ hours. Given that in perfect circumstances it would only be possible to detect our targeted behavior around or after 12 hours, $t=18$ hours is considered relatively quick. Another, albeit smaller, aspect of our research is the development of an unsupervised learning model that aims to find previously undetected anomalous instances residing in our AIS dataset. The results of this approach give positive early indicators that the model is able to detect our targeted behavior, however, further refinements and a higher amount of anomalous data are necessary to make a case for the deployment of this model. | |
