Seismic arrays for detection of small earthquakes benefit from array processing aimed at reducing noise levels. We present a frequency-dependent multichannel Wiener filtering (MCWF) technique, which employs an adaptive least-squares method to remove coherent noise in seismic array data. The noise records on a number of reference channels are used to predict the noise on a primary channel, which can then be subtracted from the observed data. A sequence of aftershocks caused by the Mw 6.1 21 February 2008 mainshock in Spitsbergen was recorded by the ARCES array in northern Norway. This aftershock sequence was filtered using the multichannel Wiener filters in both triggered and continuous modes. The Spitsbergen (SPITS) array, at a much closer distance to the source region, provides reliable reference information on the true number of detectable aftershocks. The conventional delay-and-sum beamforming combined with a band-pass filter could detect only 513 aftershocks with 181 false alarms, using a series of constraints comprised of signal-to-noise ratio, back azimuth, and slowness; the multichannel Wiener filtered results found 577 aftershocks with 165 false alarms using the same constraints. A complete automatic multichannel Wiener procedure is developed for event detection on continuous data. An appropriate signal-to-noise ratio threshold for aftershock detection of 2.7 is suggested. Compared to the beamforming method, the MCWF also reduces false alarms when detecting the same number of aftershocks.