National Technical University of Athens (NTUA)
Marine pollution detection from multispectral remote sensing data
Abstract
dc:descriptionMonitoring marine pollutants, such as Marine Debris and Oil Spills is of fundamental importance for protecting the marine environment, public health and maritime activities. The deployment of the latest-generation Earth Observation satellite platforms has led to an increasing volume and variety of geospatial data, facilitating continuous monitoring of the marine ecosystem and operational systems development. However, the accurate detection and tracking of marine pollutants and other features on the sea surface from the currently broadly available satellite data is still a major challenge due to the properties of current sensors, the complex optical properties of specific floating objects, the variability of the sea conditions and the monolithic approaches employed so far (e.g., binary classification tasks). To this end, this dissertation is concerned with the mapping of marine pollution, i.e., Marine Debris and Oil Spill, and other sea surface features using multitemporal high-resolution multispectral data at a global scale. The first topic of this dissertation is the investigation of the capabilities of the current Sentinel-2, Landsat-8 and Planet satellite sensors in detecting plastic debris. Since at the beginning of this thesis, the research on marine litter detection from satellite imagery was in its very infancy, the first step was to collect reliable in situ data of Marine Debris to verify multispectral satellite observations. Thus, focusing on Honduras Gulf (Caribbean Sea), a highly affected by plastic pollution region, significant insights about plastics' spectral behavior, sources, travelling trajectories and tracking were extracted, contributing to fundamental open issues regarding Marine Debris monitoring. Secondly, this dissertation contributes to benchmarking Marine Debris detection using high-resolution multispectral Sentinel-2 satellite data. Well-established Random Forest classifier and Deep Learning U-Net model were developed and evaluated, and the labelled data were openly provided, triggering the research community to create and validate current and future Marine Debris detection solutions. The considered created database, named MARIDA (Marine Debris Archive), was the first open-access benchmark dataset that discriminates Marine Debris from other features that co-exist, such as floating macroalgae, ships, water-related classes, etc. A detailed spectral and statistical analysis of MARIDA was also offered through interactive online material, enabling the users to explore the spectral behaviour of Marine Debris and other recorded classes. MARIDA FAIR data and software are open and re-usable and can be integrated to several marine debris detection solutions based on artificial intelligence, or satellite pre-processing pipelines. The third contribution of this dissertation is the development of a mapping framework for marine pollutants and sea surface features detection, including globally distributed Sentinel-2 satellite observations, aiming to support future operational solutions development. Towards this direction, a new state-of-the-art convolutional model was proposed, named MariNeXt, which led to overall F1, mIoU and Overall Accuracy scores of 76%, 64% and 89%, outperforming all baselines by at least 12% in F1 and mIoU metrics. A Marine Debris and Oil Spill (MADOS) dataset was also offered, composed of 174 Sentinel-2 scenes captured between 2015-2022, with approximately 1.5M annotated pixels globally distributed and collected under various weather conditions, forming 15 different thematic classes. Finally, important findings regarding the discrimination potentials of the considered thematic classes are discussed in detail based on an extensive quantitative and qualitative evaluation. To the best of our knowledge, there is no other available mapping framework and open-access dataset for marine pollutants and other sea surface features detection at 10m spatial resolution with such thematic analysis and quite promising mapping outcomes. By considering, without exceptions, the competing marine pollutants, other sea surface features and water-related classes, reliable operational monitoring solutions can be developed using the proposed framework and dataset.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kikaki, Aikaterini
- Κικάκη, Αικατερίνη
Subjects
dc:subject × 18- Θαλάσσια απορρίμματα
- Πλαστική ρύπανση
- Πετρελαιοκηλίδες
- Μηχανική μάθηση
- Δορυφορικά δεδομένα
- Υψηλή ανάλυση
- Marine debris
- Plastic pollution
- Oil spills
- Machine learning
- Satellite data
- High resolution
- Γεωεπιστήμες και Επιστήμες Περιβάλλοντος
- Φυσικές Επιστήμες
- Ωκεανογραφία
- Earth and Related Environmental Sciences
- Natural Sciences
- Oceanography
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/54940