Source Variability as a Limiting Factor in Seismic Velocity Monitoring

Industrial facilities can act as persistent, high-energy seismic noise sources and are increasingly exploited for passive time-lapse monitoring. In geothermal settings, vibrations generated by power plants provide spatially stable sources that are well suited for long-term interferometric analyses. However, temporal variations in operational conditions may lead to changes in source spectral content, potentially biasing time-lapse measurements derived from ambient noise cross-correlation. Assessing and mitigating the effects of such source nonstationarity is therefore essential for reliable monitoring.

We investigate continuous seismic noise recorded in 2008 at the Salton Sea geothermal field (southern California) and apply a Variational Autoencoder (VAE) to characterize temporal variability in source spectra. The VAE is trained on frequency spectra from relatively stable periods and subsequently used to identify time intervals exhibiting anomalous time–frequency behavior, interpreted as changes in the industrial noise source. We then compute noise cross-correlation functions and corresponding travel-time variations (dt) for both normal and anomalous periods.

Our analysis reveals systematic differences in dt behavior associated with source spectral changes, including abrupt offsets at transitions between normal and anomalous intervals and increased high-frequency fluctuations during anomalous periods. These effects occur despite stable source locations, demonstrating that spectral variability alone can significantly contaminate time-lapse measurements.

To reduce these biases, we construct separate correlation reference functions for distinct source regimes. This adaptive strategy suppresses spurious dt fluctuations during anomalous intervals and yields more physically interpretable travel-time variations. The results highlight the importance of explicit source characterization in passive seismic monitoring and demonstrate how machine learning–based approaches can enhance the robustness of time-lapse interferometry in geothermal fields and other environments dominated by industrial noise sources.