This paper proposes a novel framework, AL-PV-SHA, that integrates Bayesian neural networks with spatial active learning to enable adaptive, uncertainty-aware hazard mapping. It employs Monte Carlo dropout to decompose predictive uncertainty into aleatoric and epistemic components, guiding optimal sensor placement and data collection. Experiments show superior hazard prediction accuracy with fewer labeled samples compared to passive learning baselines.
Key findings
AL-PV-SHA achieves 40% fewer labeled samples for superior hazard prediction accuracy.
The framework provides calibrated uncertainty estimates for risk-informed decision-making.
Establishes a new paradigm for adaptive seismic monitoring through principled uncertainty-guided data acquisition.
Limitations & open questions
The framework's scalability to larger datasets and more complex geological conditions needs further investigation.
The integration with existing seismic monitoring infrastructures presents practical challenges.