Work

GAIA: A Physics-Informed Spatio-Temporal Neural Operator with Attention Gating for Geohazard Early Warning and Landslide Risk Prediction

Year: 2026
Type: preprint
Abstract: GAIA: A Physics-Informed Spatio-Temporal Neural Operator with Attention Gating for Geohazard Early Warning and Landslide Risk Prediction Abstract: Landslides represent one of the most catastrophic geohazards globally, causing severe loss of life, infrastructure degradation, and economic disruption, particularly across ecologically sensitive tropical highlands such as the Western Ghats of Kerala, India. Conventional landslide susceptibility modeling predominantly relies on either empirical limit equilibrium equations or purely data-driven statistical machine learning models, such as Random Forest and XGBoost. While data-driven models achieve high predictive correlation, they frequently lack physical interpretability and fail to generalize under extreme unseen meteorological events. In this work, we propose GAIA (Geophysical Artificial Intelligence Assistant), a novel Spatio-Temporal Physics-Informed Neural Operator with Attention Gating (ST-PINO with AG) tailored for operational geohazard early warning. GAIA employs a modular Mixture of Experts (MoE) architecture comprising: Hydrological Expert: A Bidirectional Long Short-Term Memory (Bi-LSTM) network for modeling sequential precipitation accumulation and soil moisture temporal dynamics. Geotechnical Expert: A Graph Convolutional Network (GCN) for modeling topographical terrain topology and slope proximity. Land-Use Expert: A Multi-Head Self-Attention Transformer for analyzing land cover and vegetation dynamics. Differentiable Physics Layer: An analytical Mohr-Coulomb Factor of Safety (FOS) equation embedded directly into neural attention gating to enforce mechanical slope stability laws. Key Findings and Performance: Extensive empirical benchmarks on a multi-source geohazard inventory from the Western Ghats demonstrate that GAIA achieves state-of-the-art predictive performance, registering an Accuracy of 99.30%, an F1-Score of 0.9915, a Recall of 99.40%, and an AUC-ROC of 0.998. These results significantly outperform traditional XGBoost (98.00%) and Random Forest (97.00%) baselines while maintaining strict geotechnical physical consistency. Keywords: Landslide Prediction, Physics-Informed Neural Operators (PINO), Factor of Safety, Mohr-Coulomb Failure Criterion, Attention Gating, Mixture of Experts, Geohazard Early Warning Systems. Companion Code and Repository: The end-to-end operational code, pretrained weights, and interactive Streamlit early-warning dashboard are publicly available on GitHub: https://github.com/benAJoseph/GAIA_Landslide_Predictor
Language: English
Open Access status: green
OpenAlex ID: W7234170104
Cites:
Cited by:
Related to: 10