This repository contains the code implementation of the methodology proposed in the paper "Automated Delphinid Click Identification in Seismic Acoustic Surveys for Environmental Impact Assessment". The project aims to support environmental monitoring by automatically identifying delphinid clicks in acoustic data collected during seismic surveys using passive acoustic monitoring (PAM).

Passive acoustic monitoring is an important technique for studying marine mammal behavior; however, it presents challenges due to the need for trained specialists to manually classify audio recordings—a time-consuming process. In this study, we propose a machine learning-based approach tailored to analyze audio data collected during seismic surveys, where environmental conditions result in a low signal-to-noise ratio when considering cetacean signals. Our goal is to develop an algorithm capable of detecting delphinid echolocation clicks within acoustic pre-watches—a mandatory procedure in seismic surveys for environmental protection. The data were obtained from seismic research campaigns conducted in Brazil, which follow strict environmental regulations. To address these challenges, we constructed a dataset for binary classification, distinguishing dolphin echolocation clicks from background noise. Spectrograms were generated from the audio recordings and subsequently split into training, validation, and test sets where undersampling was applied. We evaluated the performance of eight convolutional neural network architectures, with MobileNetV2 achieving the best performance (accuracy: 0.9159, recall: 0.9230, AUC: 0.9188, F1-score: 0.7878). Our analysis also highlights how seismic equipment can affect acoustic monitoring operations carried out onboard by environmental monitoring teams, due to the proximity of the hydrophones to the vessel and the noise caused by cavitation. This study contributes to the development of a robust dataset and demonstrates the potential of deep learning models to improve environmental auditing processes. By automating the analysis of passive acoustic data, our approach enhances oversight of cetacean acoustic monitoring during seismic research activities and supports broader conservation efforts.

This research was funded by public resources allocated to the project supporting the Ocean Dynamics Laboratory at the Federal University of Santa Catarina (UFSC).
It was developed as part of the master's research of Flora Medeiros Sauerbronn, under the supervision of Professor Antonio Härter Fetter (UFSC) and co-supervision of Professor Andrea Dalbel Soares from BIOSONAR.

We would also like to acknowledge the essential contributions of Pedro Igor de Araújo Oliveira and Ingridy Moara Severino, who were actively involved in the development and execution of this project.

Audio recordings and manual annotations (labels) used in this research can be accessed at:
** Mendley Data**

• Sauerbronn, Flora (2025), “PAM Recordings of Seismic Surveys - IBAMA Regulations in Brazil - part1”, Mendeley Data, V1, doi: 10.17632/w9b5z7znst.1
• Sauerbronn, Flora (2025), “PAM Recordings of Seismic Surveys - IBAMA Regulations in Brazil - part2”, Mendeley Data, V1, doi: 10.17632/8srrb2rcw6.1

• Install uv

• Create virtual environment:

  uv venv  # Create a virtual environment at .venv.

• Activate virtual environment:

  source .venv/bin/activate

• Install the project itself:

  For minimal dependencies:

  uv pip install -e .

  Or with development dependencies:

  uv pip install -e .[dev]

• The config.py sets all the configurations for the project.
• Audio files should be placed in the audios folder as specified in the config.
• Annotation CSVs should be placed in the labels directory as specified in the config. Obs.: each CSV should have a prefix equal to the name of the corresponding audio file plus a suffix separated by a dot (.).

python -m dolphins

uv pip compile pyproject.toml -o requirements.txt

uv pip compile pyproject.toml -o requirements_dev.txt --extra dev

• Make sure to have the development dependencies installed.
• Run tests with:

pytest

• Make sure to have the development dependencies installed.
• Run linter with:

ruff check src/ tests/

• Run formatter with:

ruff format src/ tests/

Below is an overview of the main folders included in this repository:

Main directory for input and output data used throughout the project.

• audios/: Contains raw audio files, as collected during the pre-watch phase of seismic surveys by the onboard environmental monitoring team. No preprocessing is applied at this stage.
• chunks/: Output directory containing audio segments ("chunks") extracted from the raw recordings. These are the samples used to train the machine learning model.
• labels/: Contains .txt files with manual annotations and cataloging of the raw audio data.
• npys/: Stores .npy files that compile training, validation, and testing datasets. These are created from the chunked samples and used directly by the model.
• tables/: Tabular metadata files that describe the content of each chunk, indicating whether it contains clicks, whistles, or background noise.

This folder contains all source code necessary to reproduce the pipeline proposed in the paper, from audio preprocessing to dataset preparation and model training.

• sql/
  Contains a DuckDB query that applies the thresholding rule described in the paper. It defines how many delphinid calls (clicks) a chunk must contain to be labeled as a positive sample.

• __init__.py and __main__.py
  Initialization and entry point for the package.

• audio_processing.py
  Includes all functions related to raw audio preprocessing, such as filtering, chunking, and signal transformation.

• config.py
  The only file intended to be modified by the user. It stores the configuration parameters such as chunk size, spectrogram characteristics, and file paths.

• controllers.py
  Manages and orchestrates the execution of different modules in the pipeline.

• data_Split.py
  Responsible for generating the training, validation, and test datasets from the preprocessed data.

• image_generation.py
  Handles the creation and formatting of spectrogram images used as model input.

• model_data_loader.py
  Implements the DataLoader used during model training and evaluation.

• targets.py
  Parses and processes label files, mapping delphinid events to corresponding audio chunks.

• undersampling.py
  Applies undersampling rules to balance the dataset by reducing overrepresented classes.

• utils.py
  Contains auxiliary functions used across different modules.

• Flora Medeiros Sauerbronn: [flora.ufsc24@gmail.com]
• Prof. Antonio Fetter: [antoniofetter@gmail.com]
• Andrea Dalben: [biosonardalben@gmail.com]

Federal University of Santa Catarina      BIOSONAR