We present a new methodology for detection and space–time location of seismic sources based on multiscale, frequency-selective coherence of the wave field recorded by dense large-scale seismic networks and local antennas. The method is designed to enhance coherence of the signal statistical features across the array of sensors and consists of three steps: signal processing, space–time imaging, and detection and location. The first step provides, for each station, a simplified representation of seismic signal by extracting multiscale non-stationary statistical characteristics, through multiband higher-order statistics or envelopes. This signal processing scheme is designed to account for a priori unknown transients, potentially associated with a variety of sources (e.g. earthquakes, tremors), and to prepare data for a better performance in posterior steps. Following space–time imaging is carried through 3-D spatial mapping and summation of station-pair time-delay estimate functions. This step produces time-series of 3-D spatial images representing the likelihood that each pixel makes part of a source. Detection and location is performed in the final step by extracting the local maxima from the 3-D spatial images. We demonstrate the efficiency of the method in detecting and locating seismic sources associated with low signal-to-noise ratio on an example of the aftershock earthquake records from local stations of International Maule Aftershock Deployment in Central Chile. The performance and potential of the method to detect, locate and characterize the energy release associated with possibly mixed seismic radiation from earthquakes and low-frequency tectonic tremors is further tested on continuous data from southwestern Japan.