Volcanic Activity Detection and Forecasting Using a Diffusion Based Model

Recent machine learning methods automate volcanic event detection, yet many struggle to generalize across event types and often require extensive labelled, eruption-specific training data. We propose an unsupervised diffusion-based approach using a Denoising Diffusion Probabilistic Model (DDPM), formulating the detection as anomaly identification: volcanic events are deviations from the learned normal behaviour of continuous seismic data, represented by characteristic tremor features derived from the ambient seismic field. The method applies a forward diffusion that injects noise into observed seismic signals and a reverse denoising that reconstructs the clean counterpart. During this reconstruction, anomalous components are effectively replaced by the model's learned representation of normal seismic behaviour. We compute anomaly scores as residuals between reconstructed and original signals, highlighting abnormal/eruptive activity. Leveraging the generative capacity of DDPMs, the model captures complex variability in the ambient seismic field and its characteristic functions. Evaluations on real-world volcanic datasets show effective detection of anomalous events with state-of-the-art performance. Additional tests on benchmark multivariate anomaly datasets also outperform prior methods. Under a field-standard 48-h forecasting protocol, the label-free model attains AUC up to 0.87 on phreatic systems, matching a supervised state-of-the-art forecaster without using eruption labels. These findings show that diffusion-based models offer a robust framework for automated volcanic monitoring, providing anomaly evidence localised in time and across input channels and demonstrating short-term forecasting skill.