ai-pam-pipeline — Reference Manual

This document is the reference manual for ai-pam-pipeline (v1.1.0), an open-source Python framework for CNN-based detection and classification of cetacean vocalizations in passive acoustic monitoring (PAM) recordings. The manual covers all six pipeline stages in technical depth: S1 — Software installation and environment: system requirements, virtual environment setup, dependency groups, GPU configuration S2 — Configuration system: two-level YAML architecture, experiment reproducibility via run_config.yaml, global parameter blocks S3 — Stage 1 (Audio I/O): seek-based memory-efficient reading, channel selection, resampling, amplitude normalization S4 — Stage 2 (Segmentation and windowing): timeline discretization, class definitions, exclusion zone logic, binary erosion, scroll augmentation S5 — Stage 3 (Spectrogram generation): STFT parameters, frequency cropping, image output parameters, colormap and channel mode S6 — Stage 4 (Image preprocessing): per-image normalization, OpenCV filter pipeline (CLAHE, Sobel, and others) S7 — Stage 5 (CNN training): architecture plug-in convention, stratified k-fold cross-validation, training hyperparameters, output artefacts S8 — Stage 6 (Cross-domain evaluation): annotation protocol, spectrogram generation for external datasets, evaluation modes, classification threshold S-UTIL.1 — Spectrogram PDF catalog: headless quality-control utility for HPC environments Related resources Source code (Apache 2.0): https://github.com/roccodm/ai-pam-pipeline Software archive: https://doi.org/10.5281/zenodo.19860736 How to cite If you use this manual or the ai-pam-pipeline software, please cite the companion software archive: De Marco, R. (2026). ai-pam-pipeline framework, v1.0.0. Zenodo. https://doi.org/10.5281/zenodo.19860736 License This documentation is released under Creative Commons Attribution 4.0 International (CC BY 4.0). The software itself is released under the Apache License 2.0.