2:45 PM - 3:00 PM
[ATT35-05] Artificial Intelligence for Anomaly Detection in Sea Surface Optical Imagery
Keywords:floating marine litter, environment monitoring, artificial intelligence, object detection, machine learning, statistical data exploration
Marine litter represents a global environmental challenge affecting nearly all of the World Ocean. Particles of anthropogenic waste, varying in size, origin, and composition, are now widely distributed across marine environments, occurring at all depths and latitudes. To effectively combat oceanic pollution, establishing standardized monitoring protocols capable of reliably assessing marine litter density and origins remains a critical first step. Among practical observational approaches, visual imaging of ocean surface combined with subsequent image analysis offers an accessible yet objective methodology.
In our study, we aim to develop a framework for detecting floating marine litter through automated analysis of sea surface imagery. We propose an algorithmic pipeline for processing visual data that identifies surface anomalies, including potential marine litter, bird presence, atypical glare patterns, and other visual irregularities distinct from typical sea surface characteristics.
Our methodological foundation lies in training artificial neural networks via self-supervised contrastive learning, alleviating the issue of the necessity of large labeled datasets. A novel aspect of our approach involves optimized sampling of sea surface image patches for contrastive learning. This optimization exploits the ergodic properties of ocean wave fields—specifically, the high spatial autocorrelation of surface features with large spatial correlation radii — to enhance training efficiency.
The developed anomaly detection framework shows potential for adaptation to diverse object recognition tasks in spatially distributed data, ranging from satellite imagery analysis to microscopic examination of field-collected samples.
In our study, we aim to develop a framework for detecting floating marine litter through automated analysis of sea surface imagery. We propose an algorithmic pipeline for processing visual data that identifies surface anomalies, including potential marine litter, bird presence, atypical glare patterns, and other visual irregularities distinct from typical sea surface characteristics.
Our methodological foundation lies in training artificial neural networks via self-supervised contrastive learning, alleviating the issue of the necessity of large labeled datasets. A novel aspect of our approach involves optimized sampling of sea surface image patches for contrastive learning. This optimization exploits the ergodic properties of ocean wave fields—specifically, the high spatial autocorrelation of surface features with large spatial correlation radii — to enhance training efficiency.
The developed anomaly detection framework shows potential for adaptation to diverse object recognition tasks in spatially distributed data, ranging from satellite imagery analysis to microscopic examination of field-collected samples.