Deep Learning-Based Road Traffic Noise Annoyance Assessment

Area of Science:

• Environmental Science
• Acoustics
• Artificial Intelligence

Background:

• Urban road traffic noise pollution is a growing public health concern.
• Assessing subjective annoyance levels is crucial for traffic noise management.
• Existing methods include time-consuming subjective experiments and less accurate objective predictions.

Purpose of the Study:

• To develop a rapid and accurate objective method for evaluating traffic noise annoyance.
• To leverage deep learning for direct mapping between noise characteristics and subjective annoyance.
• To enhance the model's robustness using transfer learning.

Main Methods:

• A deep learning model was developed to map acoustic features to annoyance levels, trained on listening experiment data.
• The model's performance was compared against regression algorithms and standard neural networks.
• Transfer learning was applied to address performance gaps in data-sparse annoyance intervals.

Main Results:

• The deep learning model achieved a 30% reduction in mean absolute error compared to existing algorithms.
• Transfer learning further reduced mean absolute error by 30% and improved correlation by 5%.
• The model showed limitations in annoyance intervals with insufficient training samples.

Conclusions:

• Deep learning offers a promising approach for rapid and objective noise annoyance assessment.
• Transfer learning significantly enhances the robustness and accuracy of noise annoyance prediction models.
• Further research is needed to address limitations and generalize the model across diverse populations.