Speaker
Description
Effective anomaly detection in active sonar data is critical for the long-term monitoring of offshore Carbon Capture and Storage (CCS) sites, where early identification of leaks is essential for environmental safety and regulatory compliance. Multibeam sonar systems offer high-resolution imaging of the water column but produce large amounts of data, rendering conventional reconstruction-based anomaly detection methods computationally demanding and unsuitable for bandwidth-limited deployments.This study investigates the application of Siamese convolutional neural networks for anomaly detection in multibeam sonar imagery. Instead of reconstructing input images, the proposed method learns a similarity measure from feature embeddings derived from pairs of sonar observations, enabling anomaly detection based on distance comparisons. A dataset comprising labelled anomalous and non-anomalous multibeam sonar images was collected at an open-water test facility, with anomalies generated using controlled scattering and wake-producing events.This study explores how model configuration and training data composition influence Siamese-network-based anomaly detection in multibeam sonar imagery. Experiments were carried out using different network sizes and varying fractions of the available pairwise training combinations to examine their impact on model behaviour and training efficiency. By considering how changes in architectural complexity and combination count affect learning stability and generalisation, the work seeks to identify configurations that are well suited to resource-constrained, edge-computing environments typical of long-term CCS monitoring deployments