Intelligent earthquake prediction using animal vocal behavior analysis based on machine learning and deep learning approaches.

Earthquakes are natural calamities that are destructive and result in the loss of lives and economic losses globally. It is vital to identify earthquakes at an early stage to reduce the damage caused to the environment and infrastructure. But, traditional monitoring much lacks timely warnings. A study has suggested an innovative framework that utilizes Machine Learning (ML) and Deep Learning (DL) to identify animal behaviour to predict earthquakes. This research proposes a novel framework to identify earthquake precursors using the field of bioacoustics, and ML and DL models. Animal's audio recordings of historical animal vocalizations were enhanced using data augmentation techniques to increase the audio dataset. The library utilized Librosa for extracting temporal and spectral features from the audio data. The ML algorithms used in the model were XGBoost, Random Forest, and Multi-Layer Perceptron. The DL algorithms used were Recurrent Neural Networks, Long Short-Term Memory, Bidirectional LSTM, and Gated Recurrent Units. The model was trained using a ratio of 80:20. Hyperparameter tuning and early stopping were used to improve the model's performance. DL methods have shown their potential to go beyond the conventional ML methods in handling temporal dependencies of the sequential audio data. The Bi-LSTM model obtained a test accuracy of 98.87%, and the AUC was found to be close to 1.00, which shows the effectiveness of the model in terms of generalization and robustness to environmental noises. The scalability of the model was tested using unseen datasets. The proposed approach permits a cost-effective early warning system. The approach is more beneficial in areas where there is no advance infrastructure available for seismic activity detection. Future work will be focused on the integration of IoT technology with edge computing technology to improve the scalability of the system.