Filtering Wind in Infrasound Data by Non-Negative Matrix Factorization

Abstract

Infrasonic data at volcanoes have been increasingly analyzed to get information about eruptive activity. Wind noise is an important problem, which cannot be solved using more classical seismological techniques such as deconvolution (Robinson, 1967), as the interaction between wind and original signal cannot be modeled simply as a convolution. The problem has therefore been tackled with a wide spectrum of original approaches, from the use of sensor arrays (Ripepe and Marchetti, 2002; Matoza et al. , 2011), to spatial filters consisting of a network of pipes (Hedlin et al. , 2003) or the location of the sensors in a densely forested area (Garces et al. , 2003). For arrays of infrasonic sensors, a pure state data‐adaptive polarization filter has been proposed by Olson (1982), dependent upon a measure of the multivariate coherence. This cannot, however, be applied to single sensor time series, whereas for monitoring purposes the availability of a single sensor very close to the crater is very important, as it was demonstrated, for example, in the 2011 Shinmoe‐dake eruption. To enhance the recognition of infrasound signals of small amplitude produced during that eruption, Ichihara et al. (2012) proposed a method for exploiting the use of a co‐located seismometer through the cross‐correlation function with seismic data. The appearance of characteristic patterns indicates an infrasound signal possibly originated at the volcanic vent. An alternative strategy was adopted in Cannata et al. (2013) to detect explosive activity of Mt. Etna; in that case the method is based on joint analysis of seismic and infrasonic data by wavelet transform coherence. Here, we apply the wind noise reduction procedure recently proposed by Cabras et al. (2014) for seismic data, based on non‐negative matrix factorization (NMF) with sparse coding (SC; Hoyer, …