Distributed acoustic systems (DAS) offer a unique opportunity to sense seismic and acoustic waves at high sampling rates and spatial resolutions using fiber optic cables. The prevalence of fiber optic cables crossing seabeds around the world make them particularly attractive sensors of opportunity for acoustical and geophysical applications. However, the high spatial and temporal resolutions of these systems pose significant challenges to efficient data processing and analysis. To overcome this, we propose a two-step exploratory analysis procedure in which the dimensionality of the data is first reduced into a latent space containing its salient information, and then perform clustering analysis of the data in the latent space. Spectrograms of detected events are encoded into the latent space using a convolutional autoencoder, and clustering is performed using Gaussian mixture model clustering. We demonstrate our technique on experimental data collected from a DAS array deployed at a geothermal energy research site. With this technique, we show a data-driven approach to exploratory data analysis, enabling identification of dominant types of detections in the data and identification and removal of anomalous detections due to instrument noise.