Abstract
In habitats that are encroaching on humans, human-wildlife conflict is an increasing global challenge. There is a significant risk of human injury and retaliatory action being taken if humans encounter dangerous animals. This work presents a novel approach to automated detection and classification of dangerous animals using audio signals, with a focus on model interpretability. This work introduces the Convolutional Interconnected Layer Neural Network (CILNN), a deep learning architecture designed to effectively process and classify animal vocalizations. Our method leverages a comprehensive set of audio features, including Mel-frequency cepstral coefficients (MFCCs) and spectral characteristics, optimized through SHAP-based feature selection. The CILNN incorporates interconnected layers and attention mechanisms to enhance feature extraction and model performance. It evaluates proposed approach on a diverse dataset of vocalizations from five dangerous animal species: bears, bison, cheetahs, elephants, and wild boars. Experimental results demonstrate that the CILNN outperforms traditional machine learning models such as Random Forests and Decision Trees in classification accuracy and robustness. Crucially, it employs Explainable AI (XAI) techniques, including SHAP values and decision tree visualizations, to interpret the decision-making processes of both our CILNN (90.6% accuracy) and other models. This interpretability analysis provides insights into feature importance and model behavior, enhancing trust and understanding in the classification process. Our work contributes to wildlife monitoring and human-wildlife conflict mitigation by offering an efficient, accurate, and interpretable method for acoustic-based animal detection.
Keywords: Acoustic of Animals, CILNN, Dangerous Wild Animals, Explainable AI (XAI), MFCC, SHAP