dc.rights.license | CC-BY-NC-ND | |
dc.contributor.advisor | Krempl, G.M. | |
dc.contributor.author | Meyer, Samuel | |
dc.date.accessioned | 2021-10-29T11:01:08Z | |
dc.date.available | 2021-10-29T11:01:08Z | |
dc.date.issued | 2021 | |
dc.identifier.uri | https://studenttheses.uu.nl/handle/20.500.12932/162 | |
dc.description.abstract | Detecting occurrences of ships discharging waste into the sea is important to reduce sea pollution,
but difficult due to data and resource limitations. The act of inspecting whether a ship has discharged
waste is expensive and true occurrences are expected to be rare. This makes it difficult to collect
enough labels to use for classification by supervised machine learning. This thesis investigated the
use of several active learning approaches (uncertainty sampling, density-weighted sampling, QBC
sampling and xPAL sampling) to help increase the rate of learning using fewer training instances to
classify looping behavior (a proxy variable for waste discharging). Trajectories were summarized to
single instances to allow established active learning methods to select them to be queried. Experi-
ments were performed for different selection/learning pipelines to classify both complete trajectories
(post hoc) and partial (initial steps to real time) trajectories. Almost all active learning methods
significantly improved learning for complete trajectory classification, on average reaching a macro F1
performance plateau of at least ∼90% within 50 queried instances, compared to ∼78% for random
sampling after 100 instances. Different models were trained for different points in elapsed time in
the trajectories for partial trajectory classification. Most active learning approaches either matched
or outperformed random sampling for partial trajectory classification, depending on the evaluated
time point. At the best time point the well performing methods, on average, reached a macro F1
performance plateau of at least ∼60% within 100 queried instances, compared to ∼50% for random
sampling after the same amount of instances. These results suggest that active learning methods
are a suitable approach to decreasing labeling efforts for the problem of looping detection for both
complete and partial trajectories, and possibly for similar problems involving trajectories and/or
high class imbalance. | |
dc.description.sponsorship | Utrecht University | |
dc.language.iso | EN | |
dc.subject | This thesis investigated the use of active learning for the classification of behavior related to waste discharging of ships in the North Sea. Different active learning approaches were tested, as well as different methods of applying active learning approaches to trajectory data. The approach taken was chosen to be general to allow inference of the more general use of active learning for similar problems. Classifications was done for full and partial trajectories. | |
dc.title | Investigating the Use of Active Learning for Classification of Ship Waste Dumping in the North Sea | |
dc.type.content | Master Thesis | |
dc.rights.accessrights | Open Access | |
dc.subject.keywords | Supervised learning, active learning, trajectories, time series, anomaly detection, high class imbalance, waste discharge detection | |
dc.subject.courseuu | Artificial Intelligence | |
dc.thesis.id | 623 | |