Forest Sound Classification Dataset: FSC22

Area of Science:

• Environmental Science
• Computer Science
• Machine Learning

Background:

• Environmental Sound Classification (ESC) is increasingly important, with forest ESC applications in monitoring illegal activities.
• Existing generic datasets limit the accuracy of deep learning models for specific forest sound identification.
• A specialized benchmark dataset is needed to improve the reliability of forest sound classification.

Purpose of the Study:

• To introduce FSC22, a novel benchmark dataset for forest environmental sound classification.
• To provide a comprehensive resource for training and validating deep learning models in forest sound analysis.
• To facilitate research in forest observatory tasks and the identification of specific forest acoustic events.

Main Methods:

• Compilation of 2025 sound clips categorized into 27 distinct acoustic classes representative of forest environments.
• Detailed documentation of the dataset preparation procedure.
• Validation of the FSC22 dataset using various baseline deep learning sound classification models.

Main Results:

• The FSC22 dataset effectively addresses the gap in specialized forest sound data.
• Baseline models demonstrate the utility of FSC22 for sound classification tasks.
• Comparative analysis highlights FSC22's advantages over existing generic environmental sound datasets.

Conclusions:

• FSC22 serves as a crucial benchmark for advancing forest environmental sound classification research.
• The dataset enables more accurate and reliable deep learning predictions for forest sound events.
• FSC22 is a valuable resource for researchers and developers in forest monitoring and related applications.