How Do Ground Conditions Shape DAS Signals? An Experimental Study with Telecom Fibers and Machine Learning.

Deployed telecommunication fiber optic cables and distributed acoustic sensing (DAS) are increasingly emerging as a versatile sensing platform in geosciences, offering high-resolution dynamic environmental monitoring across large areas. However, in practical deployments, the measured DAS signals are not entirely determined by the event source. Variations in geological conditions, fiber optic cable installation, and fiber-medium coupling can significantly alter the recorded waveform, complicating the interpretation of vibration signals.

This study aims to investigate the differences in DAS signals under different ground conditions during the same controlled impact event. We analyzed repeated-impact experiments using ∆ϕ-OTDR (Differential Phase Optical Time Domain Reflectometry) measurements acquired from deployed telecommunication fiber optic cables under three different ground conditions. By integrating time-domain and frequency-domain features, we extract discriminative patterns from raw signal fluctuations. We employed metrics such as Euclidean distance, dynamic time warping, and Area Under the Receiver Operating Characteristic Curve (AUC-ROC) to compare these differences. Based on these distance metrics, we present preliminary classification attempts for ground type configuration by applying a K-Nearest Neighbors (KNN) algorithm.

While these initial classification attempts show promise, extracting physically meaningful features from DAS data remains a challenge. We propose a novel feature extraction method that leverages the spatio-temporal coherence of DAS signals through correlation matrices used in a previous work to identify fiber layout. Here, we use it for a new target: identification of the ground configuration. This proposed approach effectively demonstrates the ability to locate known impact events and reveals unique wave propagation patterns across different ground configurations.

This preliminary study highlights the potential of coherence-based feature extraction for DAS signal analysis under different ground configurations. Future work will focus on expanding datasets, optimizing feature extraction pipelines, and integrating further machine-learning techniques for classification.