The ambient infrasound noise environment is characterized for 21 globally distributed infrasound arrays in the frequency band of 0.03 to 7 Hz. Power Spectral Density (PSD) is measured for one site of each array for 21 intervals at each of four times of day from January 2003 through January 2004. The ambient noise at infrasound stations is highly variable by season, time of day and station. Noise spectra for an individual station may vary by four orders of magnitude at any given frequency. Preliminary infrasound noise models are defined, which can be used as baselines for evaluating ambient noise at current and new infrasound stations. Median noise levels in the microbarom band centered on 0.2 Hz vary smoothly in an annual pattern, with most stations observing maximum noise during local winter. Noise amplitudes do not have a normal or log-normal distribution, but rather are skewed to larger amplitudes.
 “Ambient noise” refers to all energy measured by an infrasound sensor that was not generated by sources of interest, which for our purposes are atmospheric explosions. Our objective is to characterize the ambient noise environment of all existing infrasound stations using a common methodology. An understanding of ambient noise is important because noise often limits the ability to detect and identify signals. Results of this study allow meaningful comparisons among stations and are an important input for simulating the capability of infrasound networks to detect atmospheric explosions.
 Ambient infrasound noise is dominated by long-range pressure fluctuations generated over the oceans as microbaroms and by short-range pressure fluctuations due to local eddies and winds. The ambient noise at infrasound stations is highly variable over time and among stations. Station-dependent factors that contribute to the noise include climate, station location relative to oceans, local topography, local noise sources, vegetation or snow cover at the sensor sites, and configurations of sensors and wind-noise reduction filters.
 Microbaroms are observed worldwide with frequencies of 0.1–0.4 Hz and amplitudes of a few tenths of a Pascal. Due to their pervasiveness, they routinely set the noise levels and thus determine the detection thresholds in that band. Their generation depends upon the amplitudes of ocean waves produced by storms [e.g., Arendt and Fritts, 2000], and their propagation depends strongly on stratospheric winds [Garcés et al., 2004; Le Pichon et al., 2004].
 Wind and eddies together are generally the dominant source of noise at frequencies higher than the microbarom band. Pressure variations across the ground result from turbulent eddies [e.g., Kaimal and Finnigan, 1994]. Large turbulent eddies are generated by buoyancy driven convection, typically due to heating of the surface layer over land. Smaller eddies are then generated as wind shear in the boundary layer breaks up the large eddies, which causes the power of wind-generated noise to decrease approximately inversely with frequency.
 Several studies have compared ambient infrasound noise levels of permanent stations. For example, Le Pichon  shows the broadband noise level as a function of surface wind velocity for five stations. In general, however, noise has been assessed only for single stations for limited observation periods. In contrast, this study applies a common methodology to data from all available infrasound stations for an extended period and presents the results uniformly, which allows direct comparison among stations, seasons and other factors. In addition, we define preliminary infrasound noise models for the infrasound network. Noise models for seismic networks [Peterson, 1993; McNamara and Buland, 2004; Berger et al., 2004] have been valuable for evaluating station performance and for helping to estimate network detection capability.
Figure 1 and Table 1 define locations of the 21 infrasound stations studied. We use data from all stations for which data were available in January 2004. All stations are arrays consisting of between 4 and 9 elements. Sixteen are part of the International Monitoring System (IMS) network, and five are experimental stations. All IMS arrays and three experimental arrays have apertures between 1 and 3 km, whereas two other arrays have apertures between 0.1 and 0.2 km.
Table 1. Infrasound Station Locations
New Mexico, USA
 Each site of an array consists of three main components: a wind-noise reduction filter, a microbarometer, and a digitizer. Wind noise is reduced through spatial filtering that suppresses uncorrelated, high-frequency noise. Wind filters are radial configurations of porous hoses or of non-porous pipes forming rosettes. Filters have spatial extents of 18 to 70 m. Some stations have high- and low-frequency sub-arrays with different noise suppression filters. This paper uses data only from the high-frequency sites.
 Thirteen stations use Tekelec MB2000 microbarometers, and eight use Chaparral Physics Model 2 or Model 5 microbarometers. The MB2000 and Chaparral 5 sensors are flat to pressure over the frequency band considered in this study (0.03 Hz to 7 Hz), and the Chaparral 2 is flat to pressure above 0.1 Hz.
 Digitizers at all stations sample data at 20 samples/s, except for DLIAR (10 samples/s), and NVIAR and TXIAR (40 samples/s). Most stations apply anti-aliasing filters with corners between 7 and 10 Hz.
 Power Spectral Density (PSD) is estimated at 21 stations (Figure 1) for data from January 20, 2003 through January 31, 2004, from 21 consecutive 3-minute data segments taken four times daily, beginning at 00:00, 06:00, 12:00, and 18:00 local time. Three-minute windows are used to minimize smoothing of the amplitude distributions, while permitting estimation of the longest periods of interest. A Hanning taper is applied to the outer 10% of each data window. Spectral amplitudes are estimated using the method of overlapping fast Fourier transforms. Spectra are then corrected for the instrument responses provided by station operators.
 Plots of noise PSD for the four seasons and at four times of day were generated for all stations and used for data quality control and interpretation. Seasons are delineated astronomically and are reversed between the southern and northern hemispheres.
 We convert the PSDs into empirical noise-amplitude Probability Density Functions (PDFs) for each station. The noise PDFs illustrate the distribution of noise and are valuable for predicting signal-to-noise ratio for signals of specified amplitude for estimating station and network detection capability.
Figure 2 shows a sample PSD plot for station I08BO in Bolivia. All spectra calculated for each time and season interval are plotted as yellow lines, the median for each interval as a black line, and the 5th and 95th percentiles of the distribution as red lines. Green lines show the median of all spectra for all times and seasons for 10 stations having a full year's data, and serve as references for comparison among intervals and stations. At all frequencies the power varies by 4 orders of magnitude. The median noise level is similar to the network median for some times, such as 6 PM in winter, but is 2 orders of magnitude higher than the network median at others, such as noon in spring.
 The distribution of individual PSD curves is skewed toward higher power in Figure 2. Excursions of individual curves are greater above the 95th percentile than below the 5th percentile, as seen most dramatically for 6 AM in the spring. The noise amplitude PDFs are also strongly skewed to the right and have heavy tails (Figure 3). In the past, noise amplitude has been approximated as log-normal in many estimations of station detection capability, but such extreme features may invalidate such approximations or at least complicate their application.
 The variability in median noise levels among 18 stations is illustrated in Figure 4 for the winter months from 6 to 7 AM. Insufficient winter data are available for three other stations. The median power for a given time of day varies among stations by 1 to 2 orders of magnitude at frequencies of 0.2 to 1 Hz. At frequencies lower than 0.2 Hz, the power varies by 3 to 4 orders of magnitude, and above 1 Hz, by 4 to 5 orders of magnitude. Microbaroms, with a peak at 0.2 Hz, define a lower bound on the noise level at most stations.
 The median PSDs for several stations are distinct from the general population in Figure 4. I07AU (Australia) and I33MG (Madagascar) exhibit the highest noise levels in the 0.2 to 1 Hz band. Microbaroms are not visible for I07AU, I31KZ (Kazakhstan) or I33MG for this season and time of day, suggesting that, while microbaroms may be present, other noise sources are dominant. Noise levels at TXIAR (Texas) appear to be the lowest at all frequencies.
 The noise level at I55US (Antarctica) falls off steeply above 0.4 Hz, possibly owing to the generally low winds or the spatial filters at the site. Station operators dismiss snow cover as a likely cause, because noise levels do not change when the filters are unburied and replaced on the snow surface each year (D. Osborne and J. Olson, personal communication, 2004).
 Power spectral density begins to flatten out for frequencies above 2 to 3 Hz and then increases above 5 Hz for I10CA (Canada), I26DE, I34MN (Mongolia), and I56US (Washington). Flattening could be a consequence of the noise floor of the MB2000 sensors used at these stations or a resonance from the pipe rosette filters.
 The noise at I59US flattens out between 2 and 5 Hz and then falls off rapidly. The spectral flattening is the result of surf-generated infrasound [Garcés et al., 2003].
 Instrument responses and electronic filters are not consistent among stations. The noise PSD for DLIAR (New Mexico) drops abruptly near 3 Hz owing to an anti-aliasing filter not described in the response file. The noise PSD for PDIAR flattens out below 0.1 Hz, in contrast to other stations. This may be the result of its Chaparral 2 sensor, which has a low-frequency corner at 0.1 Hz. Many stations appear to use low-pass, presumably anti-aliasing, filters with different high-frequency corners that are not properly documented. For this reason, PSD is plotted only to 7 Hz. No correction has been made for the spatial wind-noise reduction filters, which may affect details of the frequency response as well as the calibration value at the calibration period.
 We examined variations in amplitudes among seasons and stations at 0.2 Hz, the peak frequency for microbaroms, and at 1 Hz, a frequency likely to be valuable for detection of atmospheric explosions. Median noise varies at a given station by 2 orders of magnitude depending on time of day or season. Seasonal variations are generally stronger at 0.2 Hz than at 1 Hz.
 The ratios of maximum to minimum noise levels among stations range from 11 to 14 at 0.2 Hz and range from 30 to 45 at 1 Hz, depending on the season. The narrower range of noise amplitudes for 0.2 Hz may be a consequence of the broad regional effect of microbaroms, compared to the more localized effect of wind at 1 Hz. The noise is generally higher in winter than summer by an average factor of 3 at 0.2 Hz and 2 at 1 Hz.
Figure 5 shows the median noise amplitude at 0.2 Hz for each month for all 21 stations. Noise at most stations varies smoothly in an annual cycle. The smooth variation validates the partitioning of noise estimates by season as in Figure 2.
 Seven of the northern hemisphere stations (DLIAR, I10CA, I17CI, I26DE, I53US, I56US and I59US) exhibit clear U-shaped patterns in Figure 5, with the lowest noise in the northern summer. Three of the southern hemisphere stations (I07AU, I08BO, and I55US) show humps in the middle of the year. This is mid-winter in the southern hemisphere, so this corresponds with the prevailing U-pattern in the north. These observations are consistent with our expectations that microbaroms are the dominant source of noise at 0.2 Hz and are larger during the winter in each hemisphere, because of the presence of large storms over the oceans.
Figure 6 presents preliminary models for low-, high-, and median-noise conditions for the infrasound network. The models were defined using noise observations for all times and all seasons from the ten stations for which data were available for a full year. The low-noise model (heavy red line) is defined as the minimum at each frequency of the 5th percentiles of all ten PSD distributions from individual stations. The high-noise model (heavy blue line) is defined similarly. The all-time network median (green) is the median of all PSDs at each frequency. Although subject to further refinement as additional data become available, the infrasound noise models obtained can be used as a baseline against which relative changes in noise levels are measured. The microbarom peak is seen clearly in the low- and median-noise models but is washed out by wind in the high-noise model. The 95th percentile noise for some stations is lower than the 5th percentile noise for other stations above 1 Hz.
 Ambient infrasound noise is highly variable by station, season and time of day. Median noise amplitude varies among stations by a factor of about 10 at 0.2 Hz and about 35 at 1 Hz. Median noise varies at a given station by 2 orders of magnitude depending on time of day or season. Noise at 0.2 Hz varies smoothly annually and is higher during the local winter at most stations as a result of microbaroms generated by winter storms over the oceans. Noise amplitudes at each station are heavier tailed than normal or log-normal distributions. Instrument response anomalies were identified for nine stations. Our preliminary infrasound noise models can be used as baselines for measuring relative noise levels at different times, seasons and stations, and for evaluating stations.
 Further investigation of ambient infrasound noise is warranted. Noise estimates should be extended to include newer stations of the planned 60-station IMS network and to cover two years or more to better quantify variations and to better define the noise models. The physical basis for ambient infrasound noise should be explored, including the relationship of ambient noise to surface meteorological conditions, local topography and vegetation.
 We thank Paul Piraino for calculating noise spectra.