Code for "Zero-shot Detection of Baleen Whale Calls with Vision Foundation Models and Prompt Engineering"

This repository contains the implementation code for zero-shot detection of baleen whale calls using the Segment Anything Model (SAM) and prompt engineering techniques. OVERVIEW The detection framework consists of two main components:1. DenoiseSpectro.py - Preprocessing and spectrogram generation from seismic data2. WhaleDetect.py - SAM-based whale call detection with intelligent gap-filling REQUIREMENTS - Python >= 3.8- PyTorch >= 1.7- TorchVision >= 0.8- CUDA support (strongly recommended) INSTALLATION 1. Download and extract this code package from Zenodo2. Install dependencies: pip install -r requirements.txt3. Install Segment Anything: pip install git+https://github.com/facebookresearch/segment-anything.git4. Download SAM model checkpoint (see below) SAM MODEL CHECKPOINT Download the ViT-H model and place it in the project root directory:wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth USAGE Step 1 - Generate spectrograms:python DenoiseSpectro.py <input_mseed_file> -o <output_directory> Step 2 - Detect whale calls:python WhaleDetect.py DATASET This work uses the seismic-acoustic dataset of coastal Bryde's whales published by Wang and Xiao (2025) on Figshare: https://doi.org/10.6084/m9.figshare.30363799 REFERENCES Dataset: Wang, Y., & Xiao, Z. (2025). Seismic-acoustic dataset of coastal Bryde's whales. Figshare. https://doi.org/10.6084/m9.figshare.30363799 Segment Anything Model: Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A. C., Lo, W. Y., Dollár, P., & Girshick, R. (2023). Segment Anything. arXiv preprint arXiv:2304.02643.