a. Wind-generated ambient sound

It has long been known that ambient sound levels in the ocean can be related to surface wind speeds (Knudsen et al. 1948). The ambient sound spectrum is dominated by low frequencies with sound levels decreasing at a rate of 20 dB decade⁻¹ with increasing frequency. Sound power levels at 8 kHz range from 35 to 60 dB (re 1 μPa at 1-m depth); Figure 1 displays ambient sound levels (measured with OASIS) at several different wind speeds. Ambient sound associated with wind is caused by oscillations of bubbles injected into the ocean surface as waves break (Medwin and Beaky 1989). In fact, Melville et al. (1988) demonstrated that the sound generated by a single breaking wave is proportional to the energy lost by that wave. This last point is noteworthy because it ties wind-generated ambient sound to energy dissipated by wind stress rather than to the wind itself.

b. Rain-generated ambient sound

Rain can represent a contaminant to the wind speed from ambient sound conversion, but it also represents a meaningful signal in itself. Medwin et al. (1992) give an overall description of the mechanisms that contribute to rain noise. Prior to Medwin et al.’s paper, these mechanisms were poorly understood. Figure 2 uses OASIS data to show how rain noise changes the wind-generated ambient sound spectra. This figure was generated by grouping sound spectra observations according to wind speed and weather conditions. Spectral differences were formed by subtracting averages from rain-free times from averages of data collected during periods of rain. We see that with 5–9 m s⁻¹ winds, rain increases sound levels at frequencies greater than around 10 kHz (this is the characteristic light rainfall signature), whereas with 15 m s⁻¹ winds, rain increases sound spectra at frequencies less than 10 kHz.

c. Wind direction measured by ADCPs

The concept of inferring wind direction from Doppler surface returns was first suggested by Schott (1989). Using data collected in the Gulf of Lions, Schott showed that ADCP-determined surface velocity was aligned with the wind direction. Brown et al. (1992) also reported on the use of this technique but found limited success.

Classical Ekman theory predicts surface drift currents to be directed at 45° to the wind velocity in contradiction to the results realized using (3). In classical Ekman theory, the eddy viscosity is assumed to be depth independent and this is not a realistic model near a boundary. If the eddy viscosity is considered to vary with depth and, in particular, if it should tend toward zero near the surface boundary (as expected in law-of-the-wall scaling), then surface velocities parallel to the wind velocity can be expected.

a. Accuracy considerations

The relationships between ambient sound and wind speed presented by Vagle (1990) and Evans et al. (1984) can be used as the basis for the design of a practical ambient sound recording system. Even if the calibrations provided by these publications suffer some inherent error, they can serve to establish bounds on the accuracy of ambient sound measurement that is needed to achieve a given accuracy in wind speed estimation. Figure 3 shows how uncertainty in wind speed estimate varies with actual wind speed given a 0.7-, 1-, and 1.5-dB error in ambient sound level [based on (2)]. For an accuracy of 10% in wind speeds up to a maximum of 20 m s⁻¹, sound level estimates must be made to within ±1 dB.

The figure of 1-dB uncertainty in ambient sound level provides a design point that must be met by all components of an ambient sound recording system. Aside from the obvious need for accuracy in electronic processing, any mechanism that can act to alter the recorded sound level by 1 dB must be accounted for in order to achieve a 10% system accuracy. Sound attenuation in water, bottom reflections, attenuation at the surface by bubble clouds, and interference by the measurement package can all contribute to degrade the quality of ambient sound measurements.

Bubble clouds created by breaking waves are likely to attenuate sound at the higher-frequency ranges of the OASIS system (Lemon et al. 1984). Compare the sound spectra at 7 and 15 m s⁻¹ wind speeds in Fig. 1. The apparent attenuation at 30–70-kHz frequency range for 15 m s⁻¹ winds is likely caused by bubble clouds occurring at higher sea states.

The effect of near-surface bubbles is at present not understood well enough to allow an adequate correction;the other sources of interference can (approximately) be accounted for. Figure 4 illustrates how acoustic absorption and bottom reflection affect the frequency response at 250-m depth in total ocean depths of 500, 2000, and 4000 m. The values presented in Fig. 4 were computed assuming bottom reflection losses of 5 dB, uniform in frequency and accounting for multiple bottom and surface reflections. At low frequencies, enhancements of about 2 dB are realized due to the multiple reflections from the bottom and the surface. At higher frequencies, the enhancement effect is overcome by acoustic absorption leading to a frequency-dependent variation in response.

b. Instrument design

The OASIS system is built around a conventional RD Instruments, 150-kHz BB-ADCP and the system uses the BB-ADCP computing and data storage capabilities. The only significant hardware change is the addition of an International Transducer Corporation (ITC) 6050C hydrophone used to measure sound levels and an interface capability to digitize the ambient sound signal. The hydrophone is placed on a 3-m-long cable so that it can be positioned away from any acoustic scattering that might occur due to the BB-ADCP instrument housing. Interference from the ADCP active sonar is avoided by appropriate timing of the data acquisition.

Ambient sound over the frequency range recorded by OASIS can span 70 dB in sound level. This dynamic range requirement presents a data processing and recording challenge if the required accuracy levels (±1 dB) are to be preserved. Our approach has been to prewhiten the spectrum to remove the 20 dB decade⁻¹ spectral rolloff; firmware is then used to adjust gain levels allowing data to be digitized using a 12-bit analog-to-digital converter. The choice of a 12-bit digitizer in this case was largely motivated by the favorable power requirements of this unit.

Firmware controls the OASIS gain level by monitoring the root-mean-square (rms) value of the digitized signals. When the rms value exceeds 2¹⁰, the gain is reduced, and if the rms value is below 2⁸, the gain is increased. Ideally, the gain control should maintain the digitized data with an rms value of 2⁹ = 512. The digitized time domain signal is transformed to a frequency representation using a 16-bit integer fast Fourier transform (FFT) algorithm to allow rapid computation. In this case, computational speed is not required for data throughput but rather to reduce power consumption. Overflow and underflow conditions are regulated through the FFT by scaling spectra at each pass through the FFT butterfly. The product is a 16-bit spectrum with a scaling value that also includes the hardware gain setting. For the Ocean Weather Station (OWS) Mike observations, we used the full bandwidth of the hydrophone (75 kHz) combined with the maximum frequency resolution available to the OASIS system (i.e., a full 8192-point ambient sound time series sampled at 300 kHz). Ambient sound was sampled every minute and averaged into 3-min ensemble intervals. Data storage requirements were limited by averaging the 4096 bands of power spectra into 129 logarithmic frequency bins so that the relative bandwidth (ΔF/F) of each bin is held constant.

c. Calibration

The 6050C hydrophone comes from ITC with a factory calibration but we repeated the calibration using OASIS hardware. A series of calibration tests using ITC’s facilities was undertaken by transmitting pure tone pulses toward the hydrophone in a calibration tank. Plots of observed signal level versus the source level at each calibration frequency were produced. Figure 5a provides an example taken from the 8-kHz calibration:to highlight the accuracy of the results, the difference between the measured and known source level is indicated on an expanded scale in Fig. 5b. In all cases the calibrations came out linear with some offset from the (mean) hydrophone sensitivity of −162 dB (generally less than 2 dB). The tests confirmed the ITC calibration and verified that the OASIS calibration tracked properly through changes in the gain settings. Results for both the factory and OASIS calibrations are shown in Fig. 6.

An additional calibration can be achieved by assuming that the natural ambient sound spectrum averaged over an entire deployment has little frequency structure. In that case, by applying an offset to the prewhitened spectrum, the deployment average spectrum can be treated as a calibration relative to a flat frequency response. The observed mean spectrum (also shown in Fig. 6) shows a structure similar to the laboratory calibrations but with many more data points made possible by the OASIS sampling scheme. Notable differences include a well-defined minimum at about 20 kHz in the field data that is not seen in either tank calibration. In addition, the resonance peak seen in all data at about 50 kHz is much higher in the field data. These differences can be accounted for by the narrow spectral resolution of the OASIS system. In the calibration facility, frequency resolution is limited by the need to use short tone pulses with a finite bandwidth. Both the ITC and subsequent OASIS calibrations are restricted to a bandwidth of 1000 Hz. The OASIS system, however, resolves spectral structure with a resolution of 55 Hz.

a. Deployment

The OASIS system was deployed at OWS Mike, located 210 km west of Bergen at 66°N, 2°E (Gammelsroed et al. 1992). The water depth at the site is about 2000 m and the ADCP was deployed at a nominal depth of 250 m in 1996 and 100 m in 1997. Figure 7 indicates the deployment location relative to the Norwegian coastline. A weather ship is maintained at this location within 3 km of the OASIS mooring. As part of the services of the weather ship, routine observations of atmospheric conditions are made using World Meteorological Organization standards and are recorded hourly. Wind observations are made continuously and hourly averages corrected to an observation height of 10 m are recorded. Comparisons in this study are based on 12-h average wind conditions so that any lag between OASIS observations and weather ship observations resulting from the 3-km separation will be unresolvable.

Our first dataset was collected over a 40-day period starting on 17 May 1996. For this first deployment, we were concerned about the ability of the gain control algorithm to keep track of sound level changes from sample to sample. The gain level for a given sample is initialized to the gain value used for the last ensemble of the previous sample. If significant changes in sound levels occur between sample times, inappropriate gain values would be used and data faults would occur. Our approach to avoid this difficulty was to select a sampling rate that was rapid compared to short-term fluctuations in wind speed. The sampling rate we arrived at was one sample every 3.5 min. As it turned out, the gain control system worked well with 90% of the computed FFTs having no saturation events. Our experience with the algorithm indicates that this level of data rejection results in no detectable change in resolved spectral levels.

An unexpected problem did complicate the data analysis from the first deployment. The instrument did not measure the surface drift velocity because the surface was located beyond the current profiling range. (The ADCP depth was 250 m and the profiling range was typically less than 200 m). We had selected BB-ADCP mode 4, which requires a minimum contiguous range of good data or it rejects the resulting velocities. The relatively small range of good data received from the surface echo was insufficient to satisfy the data quality check. We avoided the problem in our second deployment by positioning the instrument closer to the ocean surface, thereby ensuring that the surface echo was within the current profiling range. We would likely have obtained good surface data from the first deployment had we used BB-ADCP mode 1. Future firmware modifications in these instruments should enable reliable measurement of the surface drift velocity, even if the surface is beyond the profiling range.

Based on the favorable sound-level statistics from the first deployment, we relaxed the sampling rate to once every 72 min. Also, the instrument was deployed at a depth of 100 m to ensure that velocity measurements of the surface were realized. The second deployment commenced on 16 April 1997 and collected data for 80 days.

The critical difference between the two deployments is that for the first deployment, no wind direction estimates can be made from the OASIS data. The range of wind speeds that were observed during the first deployment is a little greater than those observed during the second deployment. For this reason, where comparisons of wind speed measurement performance are of interest, we use data from the first deployment.

The change in deployment depth does not significantly affect the ambient sound observations. The ocean surface acts as an infinite plane acoustic source and there is no change in intensity with distance from such a source (see Evans et al. 1984). Slight changes in intensity do occur at higher frequencies due to acoustic absorption but corrections are applied for these effects.

b. Computed wind speed

We computed wind speed using both the Vagle et al. (1990) algorithm [(2)] and the model given by (1). Constants for (1) were determined by correcting Evans et al. (1984) data for depth effects and using a 19 dB decade⁻¹ spectral slope to determine coefficients at various frequencies. At 8 kHz, the values used were

View Expanded

Rain noise (see Fig. 2) and shipping noise (Vagle 1990) have the potential to degrade wind speed estimates; we, however, have made no attempt to identify periods of such data contamination and all data are included in the analysis. Wind speed estimates are made by averaging over 1 h of OASIS ambient sound data.

The overall performance of the algorithms is determined by considering errors in OASIS velocity estimates as determined relative to the direct OWS Mike observations. Two aspects are important: the mean errors and the variance or standard deviation of the errors. Mean errors as a function of acoustic frequency are shown in Fig. 8a for the 1996 deployment. Results for wind speed estimates made using (1) are indicated as a dashed line, and those made using (2) are indicated using a solid line. For the OWS Mike data, mean errors are about the same for both methods and are generally less than 1 m s⁻¹, increasing somewhat at higher frequencies. Figure 8b shows how the estimated standard deviations vary with frequency. The algorithm based on (1) provides wind speed estimates with lower variance than those of (2). Considering Figs. 8a and 8b, the optimal frequency range for estimating wind speeds is from 2 to 6 kHz.

An alternative view of the data can be presented through scatterplots between OASIS estimates and direct wind speed observations. Figure 9 presents scatterplots for hourly OASIS wind speed estimates [as based on (1)] at 2, 4, 8, and 16 kHz for the 1996 deployment. The 16-kHz data stand out in that they fail to accurately record wind speeds above 15 m s⁻¹. This limit to the range of the high-frequency response is most likely due to high concentrations of subsurface bubbles that would occur at higher wind speeds. It is also noteworthy that there is no indication of a speed threshold as was seen by Wille and Geyer (1984).

c. Wind direction

Wind directions for the 1997 OASIS deployment were computed using (3) with velocities for V_m(z_m) taken from a depth of 64 m. Each wind speed estimate is based on an average over 12 h of data. Figure 10 compares OASIS-estimated wind directions with weather ship observations as a scatterplot. Notice that while some points might be considered as outliers, a few of these are simply values near 0° or 360° that have been wrapped around the axes.

The mean difference between OASIS and ship-determined wind directions is 0° ± 25°. The uncertainty here is the standard deviation of the difference when wrapped data points are not included. If surface drift direction alone is used as a measure of wind direction, the mean difference between OASIS and direct observations is −3° ± 30°. At this location, only a small benefit is realized by using the correction given by (3). It is also worth noting that for this deployment we found that the choice of reference depth [V_m(z_m) in (3)] did not greatly affect the results.