CicadaChorusDetection

This repository contains the source code for the paper "Simulation-based approach for enhancing automatic detection of cicada songs in challenging chorus conditions" (https://doi.org/10.1016/j.ecoinf.2025.103299).

Overview

Passive Acoustic Monitoring (PAM) has become an essential tool in ecological studies, but identifying individual species in complex choruses where multiple species vocalize simultaneously remains challenging. This project proposes a data augmentation method that uses acoustic simulation to train deep learning models capable of detecting cicada songs in chorus conditions with high accuracy.

Features

• Acoustic simulation to generate thousands of pseudo-chorus recordings
• Multi-label CNN-based classifier for detecting multiple cicada species
• Hyperparameter optimization for simulation parameters
• Application to phenology monitoring of five Japanese cicada species

Core scripts

scripts/
├── chorus_generator.py # Chorus simuration
├── cnn.py # CNN model definition
└── utils # Utilities
    ├── data.py # Data loaders
    ├── utils.py # Custom nn.Module
    └── whombat_utils.py # Utilities for treating whombat data

Project Structure

The project is organized into four main steps:

1. Real Data Annotation: Preparation and annotation of real-world field recordings
2. Training Data Preparation: Processing and segmentation of cicada songs and background sounds
3. Model Training: Simulation of chorus environments and CNN model training
4. Phenology Monitoring: Application of the trained model to analyze seasonal patterns

Technologies Used

• PyRoomAcoustics for acoustic simulation
• PyTorch for deep learning models
• EfficientNet architecture for classification
• MLflow for experiment tracking
• Hydra and Optuna for hyperparameter tuning

Requirements

Python and package versions are managed using uv

"hydra-core>=1.3.2",
"hydra-optuna-sweeper>=1.2.0",
"librosa>=0.11.0",
"mlflow>=2.21.3",
"omegaconf>=2.3.0",
"pandas>=2.2.3",
"pyroomacoustics>=0.8.3",
"scikit-learn>=1.6.1",
"tensorboardx>=2.6.2.2",
"torch>=2.6.0",
"torch-audiomentations>=0.12.0",
"torchaudio>=2.6.0",
"torchvision>=0.21.0",
"tqdm>=4.67.1",
"whombat>=0.8.3",

Citation

Ryotaro Okamoto, Hiroyuki Oguma,
A simulation based approach for enhancing automatic detection of cicada songs in challenging chorus conditions,
Ecological Informatics,
Volume 90,
2025,
103299,
ISSN 1574-9541,
https://doi.org/10.1016/j.ecoinf.2025.103299.