PORTIA Dataset

This dataset contains synchronized underwater acoustic recordings and AIS-derived vessel metadata collected in a busy port environment, designed to support research in passive acoustic monitoring (PAM), vessel detection, and multimodal maritime analysis. The dataset comprises 98,297 audio windows, including curated subsets for benchmarking: 36,778 binary-labeled samples, 39,996 single-vessel samples, and 34,996 multilabel samples. Vessel classes include tug, pilot, cargo, passenger, support, tanker, and other, with distributions preserved to reflect real-world traffic. For example, in the single-vessel subset, pilot (11,661) and tug (7,784) vessels are most represented, while multilabel samples include strong co-occurrence patterns (e.g., tug: 22,022; pilot: 16,910). Labels are derived from AIS-based spatiotemporal proximity, linking vessel positions to the recording location. The vessel–sensor distance ranges from 0.02 km to 1.98 km (mean: 0.6 km, median: 0.48 km). The dataset captures realistic port conditions, including overlapping sources and class imbalance, making it suitable for robust evaluation of machine learning models and acoustic scene analysis methods.