Impacts of Anthropogenic Noise on the House Wren’s Song: An XAI Approach to Bioacoustic Insights

This paper presents an automated explainable AI (XAI) method for characterizing variations in the calls of the House Wren (Troglodytes aedon musculus) under the influence of anthropogenic noise. We propose an End-to-End (E2E) Deep Learning model utilizing the pre-trained Biophony CNN architecture alongside Kapre, a Keras Audio Preprocessor. The outputs are interpreted through SHAP (SHapley Additive exPlanations) values, enabling the identification of fine-grained bioacoustic changes in vocalizations that differ between urban and rural environments. Our contributions include (i) the application of SHAP to bioacoustics data to interpret the impact of environmental noise on Wren’s vocalizations and (ii) novel visualizations using SHAP values, including the SHAP Spectrogram, SHAP Spectrum, and SHAP Time. These methods provide a transparent and detailed understanding of how urban noise affects bird vocalizations, bridging gaps in traditional statistical and signal processing approaches and offering a foundation for future studies in computational bioacoustics.