East Antarctica constitutes two-thirds of the Antarctic continent, where glacier systems have been thought to be more stable than those in West Antarctica. However, the stability could be increasingly undermined by global warming, intensifying local and regional glacial activities. Here, using deep unsupervised learning, we analyze seismic signals recorded by a dense nodal array near Dalk glacier in the Larsemann Hills, East Antarctica, in austral summer (6 Dec 2019 – 2 Jan 2020). We apply an autoencoder to automatically extract event features and input them to a Gaussian mixture model for clustering. During the operation period, three main types of seismic signals are identified: high-frequency monochromatic events, broadband short-duration icequakes, and low-frequency long-duration events. By comparing these events to environmental observations (local wind speed, temperature, tide level, and satellite imagery), we infer that the first type was wind-induced vibration, the second type thermal contraction/basal slip, and the third type water-filled crevassing/iceberg calving. The latter two glacial activities appear to be modulated by temperature and tide, respectively, implying the susceptibility of Dalk glacier to environment conditions in East Antarctica. Our results demonstrate that deep clustering is an effective means to identify diverse glacial seismicity and contributes to the rapid growth of passive glacier seismic monitoring.