ForNet: classification of critical forest acoustic events using discriminative CNN representations and ensemble...

Area of Science:

• Acoustics
• Machine Learning
• Environmental Monitoring

Background:

• Sound event classification research has primarily focused on urban environments, with limited work on forest ecosystems.
• Accurate classification of forest sounds is crucial for biodiversity monitoring and ecological research.

Purpose of the Study:

• To develop a robust framework for forest sound event classification.
• To introduce FSM5, a new dataset for forest acoustic research.
• To evaluate the effectiveness of different feature extraction methods and classification strategies.

Main Methods:

• A two-stage framework, ForNet, was proposed, utilizing Convolutional Neural Networks (CNNs) for audio embedding extraction followed by ensemble classifiers (XGBoost, Random Forest).
• The performance of Mel-Frequency Cepstral Coefficients (MFCC), Log-Mel, and Mel spectrogram features was systematically evaluated.
• The framework was tested on the newly curated FSM5 forest dataset and the benchmark UrbanSound8K dataset.

Main Results:

• MFCC and Log-Mel features significantly improved classification performance.
• Combining handcrafted acoustic features with CNN-derived embeddings outperformed end-to-end CNN classification.
• ForNet achieved 91.4% accuracy on the FSM5 dataset and 94% accuracy on the UrbanSound8K dataset via 10-fold cross-validation.

Conclusions:

• The proposed ForNet framework demonstrates high efficacy in classifying critical forest sound events.
• The study highlights the benefit of integrating traditional acoustic features with deep learning embeddings for improved sound classification.
• ForNet shows strong generalization capabilities, performing well on both specialized forest and general urban sound datasets.