PORTIA Dataset

PORTIA: A Large-Scale Port Acoustic Dataset with AIS-Grounded Annotations for AI-Based Vessel Monitoring PORTIA is a passive acoustic monitoring dataset collected in the Port of Valencia, Spain, using a fixed hydrophone array and synchronized Automatic Identification System (AIS) metadata. The dataset is designed to support the development and benchmarking of AI methods for vessel detection, vessel-type classification, and multi-label acoustic monitoring in dense, realistic port environments. The dataset consists of 3-second multi-channel WAV audio windows sampled at 32 kHz and stored as 32-bit PCM, together with CSV metadata files linking each acoustic window to AIS-derived vessel information. Each valid 3-second audio file contains 96,000 samples per channel. Recordings were acquired using three hydrophones connected to a four-input audio interface. Consequently, the released WAV files contain four channels. Using zero-based indexing, channels 0, 2, and 3 contain hydrophone signals, while channel 1 corresponds to an unused input and is expected to be silent or near-silent. The released public version contains 95,022 unique audio windows in a shared audio/ directory. These windows are referenced by three task-specific benchmark subsets: Benchmark Samples Approx. duration Purpose Binary Detection 36,148 ~30.12 h Vessel presence versus background Single-Vessel Classification 39,125 ~32.60 h Classification of isolated vessel windows Multi-Label Monitoring 34,996 ~29.16 h Concurrent vessel monitoring in multi-vessel scenes The three benchmarks are not mutually exclusive: some audio windows are shared across more than one task-specific subset. The total number of benchmark memberships is 110,269, while the union of all referenced audio files contains 95,022 unique windows. Files included in the release The dataset is distributed through separate archives for the audio data, benchmark definitions, and general metadata: File Contents PORTIA_audio.z01 to PORTIA_audio.z06 First six parts of the split audio archive. PORTIA_audio.zip Final part and main entry point of the split audio archive. PORTIA_Binary_Detection.zip Training, validation, and test manifests for the Binary Detection benchmark. PORTIA_Single_Vessel_Classification.zip Training, validation, and test manifests for the Single-Vessel Classification benchmark. PORTIA_Multi_Label_Monitoring.zip Training, validation, and test manifests for the Multi-Label Monitoring benchmark. PORTIA_Metadata.zip General dataset metadata. Dataset structure After all archives have been extracted into the same destination directory, the recommended structure is: PORTIA/ ├── audio/ │ ├── 20250516_165455.610896209.wav │ ├── 20250516_191000.729903060.wav │ └── ... ├── Binary_Detection/ │ ├── train.csv │ ├── val.csv │ └── test.csv ├── Single_Vessel_Classification/ │ ├── train.csv │ ├── val.csv │ └── test.csv ├── Multi_Label_Monitoring/ │ ├── train.csv │ ├── val.csv │ └── test.csv └── metadata.json The audio/ directory contains the shared pool of unique WAV files. The benchmark directories contain CSV manifest files defining the training, validation, and test partitions and the corresponding labels for each task. CSV metadata fields The released CSV files use a harmonized schema. Depending on the benchmark, some fields may be task-specific. Field Description window_id Name of the corresponding WAV file in the shared audio/ directory. start_utc Start timestamp of the 3-second acoustic window in UTC. label_vessel Binary vessel-presence label. 1 indicates at least one AIS-linked vessel and 0 indicates background. This is the ground-truth target for Binary Detection. n_vessels Number of AIS-linked vessels associated with the acoustic window. primary_class Class of the primary or nearest AIS-linked vessel. This is the target field for Single-Vessel Classification. primary_dist Distance, in kilometres, from the hydrophone array to the primary or nearest AIS-linked vessel. primary_mmsi MMSI identifier of the primary or nearest AIS-linked vessel. tug, pilot, cargo, passenger, support, tanker, other Per-class binary indicators used in the Multi-Label Monitoring benchmark. dist_tug_km, dist_pilot_km, dist_cargo_km, dist_passenger_km, dist_support_km, dist_tanker_km, dist_other_km Distance, in kilometres, from the hydrophone array to the nearest vessel of each class. Empty values indicate that the corresponding class is absent from the window. Benchmark definitions Binary Detection The Binary Detection benchmark is intended for acoustic vessel-presence detection. The target field is: label_vessel A value of 1 indicates vessel presence, while 0 indicates background. Single-Vessel Classification The Single-Vessel Classification benchmark contains only windows associated with exactly one AIS-linked vessel. The target field is: primary_class This subset is intended for vessel-type classification under isolated-vessel conditions. Multi-Label Monitoring The Multi-Label Monitoring benchmark is intended for dense port scenes where several vessels may be present in the same acoustic window. The targets are the per-class binary indicators: tug, pilot, cargo, passenger, support, tanker, other Per-class proximity information is provided through the corresponding distance fields: dist_tug_km, dist_pilot_km, dist_cargo_km, dist_passenger_km, dist_support_km, dist_tanker_km, dist_other_km Download and extraction instructions 1. Download all required files Download all parts of the audio archive and keep them together in the same directory: PORTIA_audio.z01 PORTIA_audio.z02 PORTIA_audio.z03 PORTIA_audio.z04 PORTIA_audio.z05 PORTIA_audio.z06 PORTIA_audio.zip The .z01 to .z06 files and the final .zip file form a single split ZIP archive. All seven files are required to extract the complete audio collection. Do not rename the archive parts and do not attempt to extract the .z01 files separately. Extraction must be started from PORTIA_audio.zip. 2. Extract the audio archive using 7-Zip The recommended extraction method is 7-Zip. From a terminal, run: 7z x PORTIA_audio.zip 7-Zip will automatically locate and read PORTIA_audio.z01 through PORTIA_audio.z06, provided that all parts are stored in the same directory. On Windows, the same operation can be performed by right-clicking PORTIA_audio.zip, selecting 7-Zip, and then selecting Extract here or Extract to PORTIA_audio\. 3. Alternative extraction using Info-ZIP Some versions of the standard unzip command do not directly support split ZIP archives. In that case, first convert the archive into a single ZIP file: zip -s 0 PORTIA_audio.zip --out PORTIA_audio_full.zip unzip PORTIA_audio_full.zip The temporary file PORTIA_audio_full.zip may be deleted after successful extraction. 4. Extract the benchmark and metadata archives The smaller benchmark and metadata archives can be extracted independently: unzip PORTIA_Binary_Detection.zip unzip PORTIA_Single_Vessel_Classification.zip unzip PORTIA_Multi_Label_Monitoring.zip unzip PORTIA_Metadata.zip Extract all archives into the same parent directory to obtain the recommended PORTIA/ dataset structure. 5. Storage requirements The complete split audio archive is approximately 75 GB. Users should ensure that sufficient disk space is available for both the downloaded archive parts and the extracted WAV files. Approximately 150 GB or more of free space is recommended when retaining both the compressed and extracted copies. 6. Verification after extraction After extraction, users are encouraged to verify that: The audio/ directory contains 95,022 unique WAV files. Each valid file represents a 3-second audio window. The sampling rate is 32 kHz. The sample format is 16-bit PCM. Each WAV file contains four channels. The CSV manifests reference files available in the shared audio/ directory. Notes on annotation PORTIA labels are derived from synchronized AIS records rather than exhaustive manual acoustic annotation. Therefore, the labels represent vessel metadata grounded in spatiotemporal co-occurrence with the acoustic recordings. This design enables large-scale benchmarking under realistic port conditions. However, users should consider the AIS-derived nature of the labels when interpreting experimental results, particularly in relation to vessels without valid AIS transmissions, positioning uncertainty, and the potential acoustic contribution of vessels located outside the annotation area. Recommended citation Please cite the associated manuscript when using the dataset: [Add final citation here once available]