Unsupervised Characterization of Rain-Induced Seismic Noise in Urban Fiber-Optic Networks Using Deep Embedded Clustering

Distributed acoustic sensing (DAS) with preexisting telecommunication optical fibers (dark fibers) has shown its ability to record rain-induced seismic noise with unprecedented high spatiotemporal resolution. This rain-induced noise exhibits strong correlations with rainfall intensity and rainwater discharge in pipeline sewers, highlighting its potential to infer rainwater flow characteristics. While raindrop impact models exist, a physical model linking stormwater discharge processes to DAS-recorded signals is still lacking. In this study, we introduce a data-driven method, deep embedded clustering (DEC), to automatically detect and classify rain-induced noise from massive DAS data, predicting the presence of moderate to heavy rain and the duration of stormwater discharge. We analyze continuous DAS recordings from 2019 to 2021 from a 4.2 km-long underground fiber-optic array in State College, PA. During training, the DEC model employs an autoencoder to learn the latent features from preprocessed spectrograms and then clusters these latent features into four clusters. Distinct features from spectrograms within each cluster reveal that four clusters correspond to background noise, rain-induced noise of varying rain intensities and stormwater discharge in sewers. Tests on unseen data sets in 2019 and 2021 demonstrate DEC's ability to not only predict rainfall rate levels but also indicate post-rain discharge durations. Furthermore, the model-derived post-rain discharge durations align with synthetic hydrograph estimates, yielding a drainage system time of concentration as 21 min in this region. Finally, we apply this workflow to two more locations to show the potential of spatial monitoring. Our results show that the combination of machine learning and fiber-optic sensing offers a scalable solution for improving stormwater management in urban environments.