Seismic-Acoustic Dataset of Coastal Bryde’s Whales in the Beibu Gulf

1. OverviewThis repository contains the dataset and deep learning framework used in the study:“Listening to Whales with an Island Seismometer: Year-Round Presence and Diel Rhythms of Bryde’s Whales Unveiled by Deep Learning.”It provides a complete and reproducible pipeline for detecting coastal Bryde’s whale vocalizations using three-component seismic data recorded at the XYD (Xieyang Island) station in the Beibu Gulf, northwestern South China Sea, during January–December 2021.The repository includes:A preprocessed dataset of labeled spectrogramsCNN-ECA model source code and trained weightsConfiguration and environment files for reproducible researchA manually annotated catalog of whale vocalizationsA full-year model inference catalog of predicted whale vocalizations2. File StructurePathDescriptiondataset/Folder containing preprocessed spectrogram data and labels.├── all_labels.xlsMetadata for all samples (timestamps, labels, data source).├── split_info.xlsSummary of dataset split ratios.├── *_indices.npyIndex files for train/validation/test subsets.├── *_spectrograms.npy3-channel normalized spectrogram arrays for each subset.├── y_*.xlsLabel files for train/validation/test sets.Dataset_S1_annotated_catalog.csvManually annotated catalog of Bryde’s whale vocalizations (see Section 4).Dataset_S2_dl_catalog.csvFull-year inference catalog of predicted whale vocalizations (see Section 5).config.jsonConfiguration file for data paths and hyperparameters.train.pyCNN-ECA model training script.test.pyModel evaluation script.optimized_whale_detector_best.pthTrained model weights (best validation F1-score).requirements.txtPython dependencies for environment setup.3. Dataset DescriptionSampling rate: 100 HzFrequency band: 3–20 Hz (Butterworth band-pass filtered)Channels: North, East, Vertical (three-component seismic data)Sample length: 10 secondsData format: Log-scaled, z-score normalized spectrogramsData shape: [N, 3, F, T]Labels1 = Bryde’s whale vocalization0 = Background / non-vocalizationData Split70% training15% validation15% test4. Manually Annotated CatalogDataset_S1_annotated_catalog.csvThis file contains the manually annotated Bryde’s whale vocalizations used to construct the training dataset.Annotations were performed by visually inspecting spectrograms and identifying characteristic low-frequency pulse sequences associated with Bryde’s whale calls.All timestamps are reported in Beijing Time (UTC+8).Columnsstart_time_bjt – Start time of the annotated vocalization segment (Beijing Time, UTC+8)end_time_bjt – End time of the annotated vocalization segment (Beijing Time, UTC+8)duration_s – Duration of the vocalization segment in secondsfreqmin_hz – Minimum frequency of the vocalization (Hz)freqmax_hz – Maximum frequency of the vocalization (Hz)Each row corresponds to one annotated Bryde’s whale vocalization event.This catalog forms the ground-truth reference used for dataset construction and model training.5. Full-Year Inference CatalogDataset_S2_dl_catalog.csvThis file contains the deep learning model inference results for the entire year of 2021.Using the trained CNN-ECA model (optimized_whale_detector_best.pth), continuous seismic data from January–December 2021 were segmented into 10-second windows and processed sequentially. All segments predicted as class 1 (Bryde’s whale vocalization) were extracted and compiled into this catalog.Columnsstart_time_bjt – Start time of the 10-second segment (Beijing Time, UTC+8)end_time_bjt – End time of the 10-second segment (Beijing Time, UTC+8)Each row corresponds to one 10-second segment classified as containing a Bryde’s whale vocalization.This catalog represents the basis for analyzing:Year-round presenceSeasonal patternsDiel (daily) vocal activity rhythms