The reconstruction of seismic Green’s functions from correlations of ambient seismic noise has recently developed as a promising approach for exploring the earth’s interiors without the requirement of costly active seismic sources. This approach is widely used for imaging the crust at a kilometer scale. However, few studies report noise-based Green’s function reconstruction at smaller scales in industrial environments. We have investigated the possibility of constructing seismic Green’s functions between sensors in an active underground mine (Garpenberg, Sweden) by crosscorrelating seismic noise generated by mining activities. We have determined with realistic numerical simulations that the mining excavations in an underground mine lead to a regime of strong scattering, which is favorable for constructing seismic Green’s functions by crosscorrelating seismic noise. One month of continuous data was recorded by 18 seismic sensors located more than 1 km below surface. A variety of broadband (10–3000 Hz) seismic sources were present, but the seismic wavefields are directional and often monochromatic, so that the conditions for constructing Green’s functions by crosscorrelating ambient seismic noise were not ideal (isotropic illumination and spectrally white). We developed a stacking scheme that dismissed data during periods when the seismic noise was dominated by monochromatic signals or when noise sources were not in stationary phase locations. Estimates of the seismic Green’s functions were retrieved for a broad frequency range (20–400 Hz) for almost all of the correlation pairs when we used the selective stacking scheme. We used the direct body waves present at low frequencies (less than 100 Hz) in the reconstructed seismic Green’s functions to invert for the 3D S-wave velocity structure of the mine. Our results revealed the existence of a high-velocity zone and a low-velocity zone that corresponded with known ore bodies.