AquaScan-1K: Paving the Way to Robust Object Detection in Side-Scan Sonar

Abstract

Machine learning models for underwater perception face substantial challenges in turbid search-and-rescue scenarios, where optical sensing is ineffective and side-scan sonar (SSS) becomes the primary modality. Existing SSS datasets often lack variability in viewpoint and seabed structure, limiting robustness and generalization. We present AquaScan-1K, a curated dataset of approximately 1,000 SSS images acquired under multiple angles and diverse seabed conditions, with expert annotations provided by Red Cross personnel using ethical human surrogates. The dataset captures realistic operational variability relevant for small-object detection. Cross-dataset experiments show that models trained on AquaScan-1K generalize more reliably under domain shift than those trained on public baselines. Transformer architectures exhibit reduced performance degradation under domain shift, and saliency analyses confirm that they consistently attend to shadow–silhouette features that constitute the primary discriminative signal in side-scan sonar imagery.