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

1. OverviewThis repository contains both the <b>dataset and deep learning code</b> used in the study:<br><b>“</b><b>Listening to Whales with Island Seismometers: Year-Round Presence and Diel Rhythms of Bryde's Whales Unveiled by Deep Learning.</b><b>”</b>It provides a complete package for detecting <b>coastal Bryde’s whale vocalizations</b> using three-component seismic data collected at the <b>Xieyang Island (XYD)</b> station in the Beibu Gulf, northwestern South China Sea, during <b>January–December 2021</b>.<br>It includes:A preprocessed dataset of labeled spectrograms;CNN-ECA model source code and trained weights;Configuration and environment files for reproducible research.2. File StructurePathDescription<code>dataset/</code>Folder containing preprocessed spectrogram data and labels.├── <code>all_labels.xls</code>Metadata for all samples (timestamps, labels, data source).├── <code>split_info.xls</code>Summary of data split ratios.├── <code>*_indices.npy</code>Index files for train/val/test subsets.├── <code>*_spectrograms.npy</code>3-channel normalized spectrogram arrays for each subset.├── <code>y_*.xls</code>Label files for train/val/test sets.<code>config.json</code>Configuration file for data paths and hyperparameters.<code>train.py</code>CNN-ECA model training script.<code>test.py</code>Model evaluation script.<code>optimized_whale_detector_best.pth</code>Trained model weights (best validation F1).<code>requirements.txt</code>Python dependencies for environment setup.3. Dataset Description<b>Sampling rate:</b> 100 Hz<b>Frequency band:</b> 3–20 Hz (Butterworth band-pass filtered)<b>Channels:</b> North, East, Vertical (three-component seismic data)<b>Sample length:</b> 10 seconds<b>Format:</b> Log-scaled, z-score normalized spectrograms (<code>[N, 3, F, T]</code>)<b>Labels:</b><code>1</code> = Bryde’s whale vocalization<code>0</code> = Background / non-vocalization<b>Split ratio:</b> 70% train / 15% validation / 15% test<br>