| Conventional geological monitoring systems may fail to provide reliable short-term early warnings for natural disasters, leaving populations vulnerable during critical pre-event windows. Prior observational studies documented correlations between animal behavior and seismic events, but, they lacked quantitative predictive frameworks. This research presents an integrated cross species machine learning framework that uses wildlife as biological sensors for disaster prediction. This framework integrates LSTM autoencoders for temporal pattern recognition, Isolation Forest algorithms for anomaly detection, and Hidden Markov Models for behavioral state modeling. Multi-species validation across farm animals (cattle, sheep), Arctic foxes, kinkajous, and white-tailed deer demonstrates robust generalization with F1-scores of 94.9%, 91.0%, and 87.0%, respectively. The framework achieves advance warning lead times of 6–48 hours before disaster onset, extending the 20-hour pre-seismic warning reported by Wikelski et al. by up to 2.4×. This represents one of the first scalable, data-driven system providing reproducible metrics for animal-behavior-based disaster early warning, offering evacuation windows that exceed existing sensor-based technologies. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.