Extended and sufficiently dense seismic networks capture spatiotemporal properties of the continuously recorded wavefields and can be used to compute the level of their coherence at different frequencies via the analysis of the network covariance matrix, which has been successfully used to study volcanic seismicity. Here, we present an application of the covariance matrix method in a subduction zone environment. We show that most coherent signals identified through the covariance matrix analysis are related to regional earthquakes with the wavefield properties affected by the scattering, which depends on the source location. Tectonic tremors, on the other hand, are not characterized by a high level of coherence. We compare real data results with a set of synthetic tests aimed at mimicking the properties of seismic sources and the main features of wave propagation. We conclude that highly coherent volcanic tremor wavefields could be produced in two ways: by a spatially localized group of monochromatic seismic sources or by a single source located in a highly heterogeneous medium. In both cases, the stability of the source position is a necessary condition to reproduce the observations in volcanoes. On the other hand, the low coherence of tectonic tremor wavefields can be explained by a spatially extended distribution of sources, in agreement with large portions of the subduction interface being nearly simultaneously involved in the episodes of slow deformation.