Multi-sensor integration for an assessment of underwater radiated noise from common vessels in San Francisco Bay

Sound produced underwater from modern vessel activity is a primary component of anthropogenic noise in the marine environment (Wenz, 1962). Off the U.S. west coast, increased commercial shipping traffic since 1960 led to concurrent increases in underwater noise (McDonald et al., 2006). Growing evidence suggests that large commercial ships are not influencing underwater noise levels exclusively, especially in urbanized coastal areas (Pine et al., 2016; Hermannsen et al., 2019), where recreational and commuter vessels may also be common. Acoustic energy, a form of marine pollution (UNCLOS, 1982), can negatively impact marine mammals (Southall et al., 2019). Long-lived species, like cetaceans, evolved in the absence of this anthropogenic noise and are unlikely to adapt quickly as vessel noise is increasing in their environment (Tyack and Janik, 2013).

Anthropogenic noise can impact cetaceans through cumulative and synergistic pathways (Weilgart, 2018). Masking can disrupt how species sense their environment and communicate (Clark et al., 2009), and behavioral responses to noise can interrupt foraging (Blair et al., 2016), time resting, and increase respiration and swimming rates (Sprogis et al., 2020). These impacts potentially impose a fitness cost (Holt et al., 2015) or expose individuals to direct threats such as vessel collisions (Gannier and Marty, 2015). Case studies of subpopulations suggest that the sub-lethal effects of anthropogenic noise, such as disturbance and displacement from biologically important habitat, have a greater negative impact on population-level success over time than acute injuries (Forney et al., 2017).

Mitigating negative effects of vessel noise on cetaceans requires understanding the sources, levels, and distribution of that noise in their habitat. In this research, we assessed underwater radiated vessel noise in San Francisco Bay (SFB), a highly urbanized estuary where there is a wide range of vessel use, including commercial shipping, substantial ferry service, and recreational traffic (Cope et al., 2020). Cetacean species observed in SFB include seasonal gray (Eschrichtius robustus) and humpback (Megaptera novaeangliae) whales, year-round bottlenose dolphins (Tursiops truncatus), and harbor porpoises (Phocoena phocoena) (Stern et al., 2017; NOAA, 2019a). There has been an elevated rate of gray whale strandings off the U.S. west coast since 2019 (NOAA, 2019b), some of which were on SFB shores. Vessel collisions are a well-documented risk to cetaceans in the SFB area (Barcenas-De la Cruz et al., 2017) and along the California coast (Rockwood et al., 2017; Redfern et al., 2020).

The U.S. National Marine Fisheries Service (NMFS) classifies cetacean species into low- (7 Hz to 35 kHz), mid- (150 Hz to 160 kHz), and high-frequency (275 Hz to 160 kHz) groups based on documented hearing ranges (NMFS, 2018). Since low-frequency sound may be attenuated in shallow environments (Urick, 1983), mid- to high-frequency groups may be of specific concern in urbanized, enclosed, or partially enclosed coastal areas. Higher frequencies do not propagate as far as low-frequency sounds over large distances (Urick, 1983), but the concentration of vessels and cetaceans within short distances and frequent repeated exposures make these sounds significant in congested, urban areas (Williams et al., 2014a; Marley et al., 2017).

For continuous, non-impulsive sound, like that from a vessel passage, sound exposure level (SEL), defined as the time-integrated squared sound pressure level (ISO, 2017), is a metric that incorporates sound pressure level and duration (Southall et al., 2019) that potentially impact cetaceans. NMFS (2018) synthesized past research to develop cumulative SEL thresholds per frequency group for assessing potential injury to cetaceans, although threshold values are not recommended when accumulating SEL from multiple sources across space and time. A mobile source and receiver (i.e., vessel and individual animal) also compounds modeling complexity. The SEL metric is useful in modeling overall acoustic conditions over time to inform general spatial patterns of sound levels in cetacean habitat [e.g., Erbe et al. (2012), Ellison et al. (2016), and Joy et al. (2019)].

Past research on vessel noise has been largely focused on the low-frequency (<1 kHz) contributions from vessels in the open ocean and their potential to impact low-frequency cetaceans [e.g., Hatch et al. (2008) and Redfern et al. (2017)], who typically vocalize in this range (Clark et al., 2009). While much of the research has focused on commercial shipping, fishing and whale-watching vessels that are smaller in size also have the potential to mask communication for these species (Cholewiak et al., 2018). The primary source of broadband underwater noise from vessels is cavitation at propeller blade tips which can reach frequencies higher than 1 kHz, up to 100 kHz (Ross, 1976; Hildebrand, 2009). There is documentation of this higher frequency sound (>1 kHz) from large ships [e.g., Arveson and Vendittis (2000), Hermannsen et al. (2014), and Veirs et al. (2016)] and smaller vessels with outboard or inboard engines (Erbe et al., 2016; Hermannsen et al., 2019; Wladichuk et al., 2019).

Recreational craft have received little attention in underwater noise studies despite a growing number of registered vessels (Erbe et al., 2019). One complication when incorporating recreational craft into a systematic analysis of vessel noise in cetacean habitat is data availability. Commercial vessels are required to broadcast vessel and voyage information via the Automatic Identification System (AIS) (Tetreault, 2005). AIS data consisting of known vessel locations are used to construct vessel trajectories through space and time and have facilitated vessel noise modeling in cetacean habitat [e.g., Hatch et al. (2008), Redfern et al. (2017), and Cholewiak et al. (2018)]. Recreational vessels and those used by small-scale fisheries may not meet the minimum size at which AIS data transmission is required by law. Ignoring the noise contribution from these smaller vessels could lead to significant underestimation of anthropogenic noise in cetacean habitat in urbanized coastal areas (Hermannsen et al., 2019).

Previous studies on noise emitted from non-AIS vessels have employed manual methods for determining vessel location, such as theodolite tracking (Williams et al., 2014b; Hermannsen et al., 2019) and known waypoints (Pine et al., 2016), both requiring human effort that can limit data collection opportunities. The Marine Monitor (M2), a proprietary coastal monitoring system developed by ProtectedSeas (2019), provides a tool for tracking non-AIS vessels as it utilizes a shore-based commercial marine radar system to autonomously track all vessel activity (Cope et al., 2020). Equipped with an Automatic Radar Plotting Aid, the M2 tracks vessels as unique targets whose attributes, such as location, speed, course, and heading, are stored for post hoc analysis. A high-definition pan-tilt-zoom camera captures photos of targets tracked by radar so that these vessels can be classified. Finally, the M2 system integrates data collected via AIS receiver to provide additional information from vessels broadcasting AIS data, such as vessel identity and size.

The objective of this research was to utilize this vessel tracking tool in an assessment of radiated vessel noise from all relevant vessel types in a specific potential cetacean habitat area in central SFB. This opportunistic research estimated source levels (SL) of unique vessels passing through the study area (Scrimger and Heitmeyer, 1991; Wales and Heitmeyer, 2002) and contributes to the growing number of unique vessel SL estimates on the U.S. west coast [e.g., Bassett et al. (2012), McKenna et al. (2012), Veirs et al. (2016), and MacGillivray (2019)]. Employing radar allowed for the inclusion of non-AIS vessels, which makeup a substantial portion of vessel traffic in the study area (Cope et al., 2020). We first employed traditional methods of calculating SLs based on acoustic recordings and vessel passages. Next, we modeled radiated noise propagation from vessels for comparison with ambient noise levels. Finally, we modeled cumulative SEL across the study area. The M2 system employed multi-sensor integration to provide the vessel tracking data necessary for acoustic analysis of the most common vessel types in SFB, regardless of size and consequent AIS data availability, although non-AIS vessels could be confirmed only during daylight hours. With both vessels and cetaceans present in SFB, it is vital to understand the underwater soundscape of this habitat. This research provides a first assessment of underwater noise from vessels in SFB, for use in evaluation of habitat quality and management and as a baseline for further study.

Vessel data were collected continuously for 11 days (1 February 2018 through 11 February 2018) by the M2 system, sited at the Estuary & Ocean Science Center in Tiburon, California along the SFB shoreline in the northern Central Bay (Fig. 1). See Cope et al. (2020) for a full description of detecting and processing vessel positions and trajectories from both AIS and radar data via M2. For non-AIS vessels detected by radar, we manually classified vessel types using photos collected by the M2 system and made available in a web-based data portal. Since photo classification was restricted by daylight, we were not able to incorporate non-AIS vessels transiting at night, which likely influenced SELs as a result of those vessels detected by radar relative to those providing AIS data (discussed further in Sec. V A).

We limited analyses to those vessels most common and largest in size: ferries, motorized recreational craft, crude oil tankers, bulk carriers, oil/chemical tankers, and vehicle carriers (Fig. 2). Vessel type classification for the large commercial ships was consistent with ship-type categories in Lloyd's Registry of Ships, and consistent acoustic qualities within each respective ship-type category make it appropriate to group vessels at this scale (McKenna et al., 2012). Motorized recreational craft were those vessels transiting using outboard or inboard engines with no commercial markings or obvious industrial utility. Inconsistency of photo quality prevented more narrow classification of these vessels. Vessel characteristics related to identity and design were primarily provided in AIS data but unavailable for motorized recreational craft since these vessels were predominantly (86%) detected by radar.

Some common vessel types were excluded from our analysis. We did not include tugboats as we did not observe isolated passages near the M2 during the study period. Since these vessels typically escorted large ships while traveling to and from port, their sound contribution is incorporated into the overall sound observed from the ship with which they were traveling. Sailing vessels were also excluded from our analysis due to uncertainty of operating conditions. While sailing vessels transiting under power can have SLs similar to those of small motorized recreational craft (Wladichuk et al., 2019), we were unable to consistently determine if vessels were operating under power or sail. Recreational fishing vessels were excluded due to inconsistency of operating conditions. These vessels were classified as recreational fishing vessels (instead of motorized recreational craft) if photographs showed clear fishing activity. This hook and line activity primarily occurred in the northwest corner of the study area, with no photographic evidence of these vessels passing by the hydrophone location while actively fishing. Despite the large size of container ships, we only observed a single passage of this ship-type by the hydrophone and overall presence in the study area was low, so they were not included in analysis.

For use in determining vessel SLs, we used the known geolocation of the hydrophone used to collect acoustic recordings (Fig. 1) and vessel positions provided by the M2 system to calculate distance from source to receiver. AIS and radar detection points were received and processed on average every 7 and 24 s, respectively, along a vessel's trajectory. For each vessel passage, we identified the point with the minimum distance to the hydrophone as the closest point of approach (CPA) and used the associated distance in SL calculations. Commercial traffic in this region is spatially consistent (Cope et al., 2020), so records with an extended pause or strong deviation from expected paths of travel discovered in manual review were not considered. We prioritized vessels closest to the hydrophone ignoring motorized recreational craft beyond 2 km due to the potential for their sound to be masked by other noise in the greater area (Hermannsen et al., 2019). We only considered trajectories running parallel to the shoreline and passing by the hydrophone. There were no violations of wind speed conditions required by ANSI/ASA (2009) and ISO (2016) (<10.28 m s⁻¹), and wave heights are likely insignificant in SFB compared to open ocean conditions. Each passage was considered an independent event for each unique vessel.

While speed over ground (reported in AIS and radar data) was used to select the time window defining the bounds of acoustic data collection, speed-through-water was reported in summary tables, removing the effect of current direction and speed [e.g., McKenna et al. (2012) and Joy et al. (2019)]. Current and tidal data were collected at National Oceanic and Atmospheric Administration (NOAA) stations s08010 and 9414863 located at 37° 54.975′ N, 122° 25.340′ W and 37° 55.4′ N, 122° 24.6′ W, respectively (Fig. 1) (NOAA/Center for Operational Oceanographic Products and Services, 2021).

When modeling radiated noise from vessels in SEL calculations, we considered the complete trajectories across the study area of those vessel types for which we calculated SL. We defined study area boundaries by the maximum spatial extent of radar-detected data collected by the M2 (Fig. 1) as it was smaller than that of received AIS data [see Cope et al. (2020)].

We monitored underwater vessel noise using a calibrated omnidirectional hydrophone (High Tech Inc., Long Beach, CA, USA, sensitivity −165 dB re 1 V/μPa from 0.02 Hz to 20 kHz, 3 dB) at the end of a research pier (Fig. 1), approximately 70 m from shore. Total depth at the hydrophone was 10 m mean lower low water (MLLW) with a tidal range of −0.33 to +2.55 m (Table I). The hydrophone was suspended 3 m below surface. This single aspect and shallow inclination angle (angle between surface and hydrophone with respect to source) potentially influence SL and SEL results (discussed in Sec. V A). Data were recorded using a Song Meter SM2 data-logging platform (Wildlife Acoustics, Inc., Maynard, MA, USA), digitizing at 48 kHz, and time-synchronized with M2. We collected 119-min-long recordings in WAV format, every two hours during the study period. Vessel passages with a CPA that occurred in the first or last minute of recording segments were removed from consideration when calculating SLs.

We selected a segment of the pressure time series for use in determining received levels (RL). Previous research has employed different methods for selecting this interval, including a static (Pine et al., 2016; Veirs et al., 2016) or fluctuating window dependent on vessel size or distance from the receiver (McKenna et al., 2012; Gassmann et al., 2017; Joy et al., 2019). Since speed and CPA distance varied across vessel types and passages (Table II), we chose a fluctuating window. Using the vessel's distance to the hydrophone and speed at its CPA, we calculated a time window, centered about the corresponding time, over which the vessel's position changed by a given angle about the hydrophone. The ANSI/ASA (2009) standard requires an angle of ±30°, but we selected a smaller angle of ±0.1 radians (5.7°) to minimize the influence of nearby vessels in a busy urban environment while still providing a signal long enough to average out short-term fluctuations.

For each passage, we used the signal from the full time window to calculate the mean-squared sound pressure at the hydrophone, P, with a reference pressure of 1 μPa² and reported on a relative logarithmic scale in decibels (dB) (Gassmann et al., 2017), where

To calculate RLs in adjacent one-third octave (base 2) bands, the discrete Fourier transform of the full signal was used to calculate the power spectrum, in Pa²/Hz, which was then integrated from the band's lower to upper edge. Standard center frequencies (F_c), referenced by ISO (2017), ranged from 0.02 to 20 kHz.¹

The RLs were considered in two unweighted frequency bands, Band 1 (B1) and Band 2 (B2), in addition to the full broadband spectrum to facilitate SL comparisons with previous research. RLs per band were calculated by summing across one-third octave band levels and converting back to sound pressure levels according to

where L_i was the one-third octave band level, and n was the number of bands. B1 (F_c from 0.1 to 10 kHz) was meant to capture sound potentially perceived by large cetaceans (humpback and gray whales), and B2 (F_c from 5 to 20 kHz) for small cetaceans (bottlenose dolphins and harbor porpoises) in SFB. The upper bound of B2 was limited by the frequency response of the hydrophone. While auditory weighting functions are often applied per species [e.g., Bassett et al. (2012), Williams et al. (2014b), and Ellison et al. (2016)], our chosen frequency bands align well with functional peaks recommended by NMFS (2018) and enable general comparisons across acoustic conditions for cetaceans.

Shipping activity influences background noise levels in the proximity of major ports (Williams et al., 2014a). To estimate varying ambient conditions, the full duration of all acoustic recordings was subdivided into 10-s time windows encompassing the complete study period. For each 10-s sample, one-third octave RLs were determined using Eq. (1), and ambient levels (NL) in B1 and B2 were calculated using Eq. (2). The distribution of these samples per B1 and B2 yielded statistics describing 5th, 50th, and 95th percentiles as a proxy for quietest, median, and noisiest conditions, respectively, in these bands (Williams et al., 2014a; Joy et al., 2019).

We back-propagated RLs to calculate vessel passage SLs, normalized to 1 m from an idealized point source (ANSI/ASA, 2009; ISO, 2016), by considering transmission loss (TL) between source and receiver. Losses of a sound wave occur due to geometric spreading, the weakening of sound energy as it radiates from a source, and attenuation, which includes boundary interactions and absorption as a result of water column and seabed properties (Urick, 1983). The influence of these factors on TL can be complex and vary with range which has led previous opportunistic research to adopt simple TL models for estimating SLs more systematically, given that simple models were validated using a more detailed method [e.g., Bassett et al. (2012), Erbe et al. (2012), and McKenna et al. (2012)]. We used parabolic equation (PE) modeling to similarly provide range-dependent TL estimates from which to verify an appropriate loss model.

The range-dependent acoustic model (RAM) in the Acoustic Toolbox User Interface and Post Processor (Duncan and Maggi, 2005) provided PE modeling along four bathymetric profiles, sampled every 100 m, at two key frequencies. Using bathymetry data (MLLW) provided by NOAA (2021), transect profiles were created between the hydrophone and four test points along two vessel passages, which included each vessel's CPA and its last detection within study area boundaries. These passages were representative of common routes through the study area (Cope et al., 2020). A source depth of 10 m (Bassett et al., 2012) was paired with a receiver depth of 3 m (depth of the hydrophone).

Other model inputs included sound speed profiles for the water column and seabed. Since turbulent mixing from tidal fluxes keeps the water column in central SFB relatively well-mixed (Cloern, 1996), we assumed a vertically consistent sound speed profile. We used hour-averaged water quality data at the hydrophone (see Table I for a summary of environmental data), provided by San Francisco State University's water quality reporting system (Coastal Observations and Monitoring Science, 2021), to calculate the water column sound speed (m s⁻¹) (Fonfonoff and Millard, 1983), time-matched with each test point. For the seabed, we used the average sediment grain size in the study area, 4.94 (phi) (Barnard et al., 2013), to calculate the corresponding sound speed ratio (m s⁻¹), density ratio (kg m⁻³), and attenuation rate (dB per wavelength) (APL, 1994) for each test point at 0.1 kHz and 5 kHz (lower F_c bounds of B1 and B2, respectively).

We evaluated two simple TL models according to

where β was the TL coefficient, and r was the distance from source to receiver in meters. Sound waves in water initially spread spherically (where β = 20) until reaching the bounds of surface and bottom where spreading transitions to become cylindrical (where β = 10) (Urick, 1983).

First, we applied a single-value model where 10 < β < 20 as both spherical and cylindrical spreading likely play a role in TL in shallow environments (Bassett et al., 2012; Chion et al., 2017; Pine et al., 2016). For each transect and frequency previously described, we fit a line using least squares regression through the origin where PE modeling estimates of TL were regressed against log₁₀(r). We averaged the resultant TL coefficient β (slope) to obtain a mean value for application across all TL calculations (Veirs et al., 2016; Chion et al., 2017). Because environmental parameters were incorporated in PE modeling, we did not apply a term to account for absorption.

Second, we applied a multi-term model incorporating both geometric spreading and absorption (Erbe et al., 2012),

where we assumed spherical spreading [first term in Eq. (4), where β = 20] from the vessel's location to a distance r₁, where a transition to cylindrical spreading [second term in Eq. (4), where β = 10] occurred to a distance r₂ (known distance between source and receiver). We chose a transition distance r₁ equal to three times the water depth at the source due to the bottom composition of central SFB, primarily mud and sand (Barnard et al., 2013; Pine et al., 2016) which weakens bottom reflection (Jensen et al., 2011). We calculated the absorption coefficient α (Ainslie and McColm, 1998)¹ associated with each one-third octave F_c similar to Chion et al. (2017) using water quality conditions concurrent with a vessel's CPA and summed per B1 and B2 according to Eq. (2). TL curves resulting from Eq. (3) and Eq. (4) were visually compared to PE modeling results.

After TL model selection, we calculated the broadband, B1, and B2 SLs associated with each vessel passage according to

We also calculated representative SLs meant to capture vessels transiting under typical operating conditions by selecting a subset of unique passage records in which vessels traveled at the most common tide-corrected speed within each category. From these records, we calculated the median B1 and B2 SLs for use when modeling at the scale of vessel type.

To broadly inform how loud a vessel type may be to a receiver in the study area, given distance to the source, we modeled radiated noise from a representative vessel in each vessel type category. We applied B1 and B2 SLs associated with vessels transiting under typical operating conditions (described in Sec. II F) and the same TL equation selected to calculate SLs, rearranging Eq. (5) to

To evaluate modeled TL error, and thus estimated SL error, we calculated the standard deviation of the modeled from observed RLs at corresponding distances. Direct application of modeled RLs should be approached with caution as factors including fluctuations in background noise, source and receiver orientation, and species-specific frequency sensitivity influence detection of sound by cetaceans (Clark et al., 2009; Gannier and Marty, 2015).

We calculated SEL using a 24-h accumulation period, as recommended by NMFS (2018), as a result of all vessel types for which we determined SL by modeling propagation from vessels as they transited the area. Since non-AIS vessels transiting at night were not included in analysis, the cumulative SEL metric likely underestimated true values for motorized recreational craft. To preserve variation within vessel type categories, as noted by Joy et al. (2019) and MacGillivray (2019), as much as possible, we applied the passage specific B1 and B2 SLs to those vessel passages. To incorporate passages without an associated SL, we applied B1 and B2 SLs under typical operating conditions for each vessel type.

We considered all trajectories within study area boundaries in this step regardless of distance to the hydrophone. First, we interpolated vessel detection points along each trajectory to create a position every 60 s and overlaid a 300-m grid. While the positional accuracy of the radar sensor is smaller than 300 m at the range of most vessels (Cope et al., 2020), we chose a larger cell size to minimize data processing time. For each detection point, we calculated B1 and B2 RLs per grid cell, employing the TL model selected to calculate SLs and Eq. (6). We then computed cumulative SEL per grid cell for each 60-s exposure in B1 and B2 according to ANSI/ASA (2013),

where RL was the modeled RL and t was equal to 60. We calculated daily cumulative SEL per grid cell over each 24-h period according to

where n was the number of 60-s cumulative SEL estimates to be added in a day. The arithmetic mean of the daily cumulative SEL estimates was calculated across the eleven days of the study period (Martin et al., 2019). We estimated 5th, 50th, and 95th percentile ambient daily cumulative SELs using Eq. (7), where B1 and B2 NLs were substituted for RL and t was equal to 5%, 50%, and 95% of the total seconds in a day, respectively.

We estimated a SL for 565 (60% of a total 935) vessel transits across 11 days within the six vessel type categories considered for analysis.¹ For those vessels transmitting AIS data, we estimated SL for 45 unique vessels, 73% of which made at least one return trip, and 33% more than two. We estimated a SL for 58 of 173 transits of motorized recreational craft (19% of all radar-detected and AIS transits combined). The relative presence of vessel types was consistent with presence previously observed across 1 year (Cope et al., 2020).

Both TL models showed poor alignment with PE modeling at 0.1 kHz beyond 1 km when depth was variable along the transect (Fig. 3). As a result, these regression results (P3 and P4 transects at 0.1 kHz) were ignored when averaging. Values of the TL coefficient β across the other transects and frequency ranged from 15.6 to 21.0 dB/decade, with a mean value of 18.4 dB/decade.¹ The TL model incorporating Eq. (3), where β = 18.4, showed better alignment with PE results than the model incorporating Eq. (4) and was selected to calculate SLs (Fig. 3). Due to poor model alignment beyond 1 km at 0.1 kHz, we did not consider broadband or B1 SLs for passages of all vessel types with a CPA greater than 1 km, and we limited modeled RLs in B1 to a radius of 1 km from a source. B1 and B2 SELs were both modeled within 1 km of vessel transits to facilitate comparison across frequency bands of interest.

Median broadband SLs, in dB re 1 μPa² (0.02–20 kHz), were 170.0 for ferries, 168.2 for motorized recreational craft, 177.9 for crude oil tankers, 170.8 for bulk carriers, 179.3 for oil/chemical tankers, and 177.5 for vehicle carriers (Fig. 4). Oil/chemical tankers had the highest median B1 SL (178.1 dB re 1 μPa² 0.1–10 kHz). But crude oil tankers had the highest when transiting under typical operating conditions (178.8 dB re 1 μPa²), and the unique vessel with the highest median SL was also a crude oil tanker (MMSI 303031000) (181.4 dB re 1 μPa²). While ferries were smaller in size than the commercial ships (Table II), the median B1 SL of ferries was only 0.7 dB less than the median B1 SL of bulk carriers (169.0 dB re 1 μPa² compared to 169.7 dB re 1 μPa²).

Bulk carriers had the highest median B2 SL (169.0 dB re 1μPa² 5–20 kHz), and the vessel with the highest median SL was also a bulk carrier (MMSI 373327000) (181.5 dB re 1 μPa²). Ferries had a higher median SL than vehicle carriers (164.2 dB re 1 μPa² compared to 163.0 dB re 1 μPa²), and 11 (of 13) unique ferries observed in B2 had higher median SLs than unique vessels in other vessel type categories. Median SL of motorized recreational craft (160.7 dB re 1 μPa²) was higher than two (of ten) unique bulk carriers and one (of six) oil/chemical tanker observed in B2.

The RLs at a potential receiver, as a function of distance from the source, are shown in Fig. 5. RLs from ferries, motorized recreational craft, and oil/chemical tankers were lower than the 95th percentile ambient NLs at 1 km from the source, and motorized recreational craft were also lower than the 50th percentile. Modeled RLs had standard deviations from observed RLs ranging from 2.74 to 10.44 dB and did not appear strongly dependent on range or frequency, suggesting that the error in SL estimates was approximately between 2 and 11 dB.

The spatial distributions of SELs by vessel type (Fig. 6) were similar to distributions of presence across 1 year (Cope et al., 2020). B1 and B2 SELs were highest across the most area as a result of ferries, with approximately 34% and 19% of the study area, respectively, above 170 dB re 1 μPa² s. While larger ships made less transits (Fig. 2), SELs were similar within the concentrated paths of crude oil tankers and bulk carriers. Despite the lower median SLs of motorized recreational craft compared to commercial vessels, SELs as a result of these vessels had the largest spatial footprint of all vessel types.

While TL can be modeled effectively in deep water using simple spreading equations, more complex, range-dependent TL models can account for bathymetry and characteristics of environmental mediums and create more precise estimates (Farcas et al., 2016). Range-dependent PE modeling, an effective method of modeling TL in shallow environments (Kuperman, 1996), confirmed the influence of varying depth on TL in SFB. Maintenance dredging in Regulated Navigation Areas (RNAs) likely helps sustain somewhat abrupt changes in depth in the study area (USACE and RWQRB, 2015), like a minor shoal approximately 7 m deep between two RNAs (Fig. 1). This minimal depth relative to larger wavelengths associated with low-frequency sound likely restricted propagation of sound waves at 0.1 kHz (Kozaczka and Grelowska, 2017), which was evident in PE modeling across the P3 transect (Fig. 3). Since depth was more spatially consistent in other regions of the study area, our choice to limit SEL modeling to 1 km likely underestimated true cumulative SELs.

To facilitate SL estimation across 565 vessel passages, a simple but accurate TL model was desired. By fitting a single-value TL model to PE modeling results, we aimed to leverage the accuracy of PE modeling in shallow environments with the utility of a simple model. Erbe et al. (2012) used a multi-value TL model similar to the model we tested to effectively predict RLs in deeper environments, but this model primarily underpredicted TL in the shallow SFB study area. Employing the single-value TL model generally showed agreement between modeled and observed RLs (Fig. 5) which supports its implementation in SL estimates. The average TL coefficient we calculated was within 0.2 and 0.9 dB/decade of those determined by Veirs et al. (2016) and Chion et al. (2017), respectively.

In summarizing opportunistic studies of vessel SLs, Chion et al. (2019) reported median broadband SLs of commercial ships between 170 to 210 dB, a range under which our estimates fell, although few unique passage SLs were higher than 180 dB re 1 μPa² (0.02–20 kHz) (Fig. 4). Except for bulk carriers, median broadband SLs of commercial ships in this study show agreement within 5 dB compared to reported SLs in similar previous studies, which is less than the SL error we estimated as a result of TL modeling.

Broadband SLs reported by Bassett et al. (2012) and average SLs of unique vessels reported by McKenna et al. (2012) were higher than median SLs in this study by <3 and <2 dB for tankers and vehicle carriers, respectively. Because the bandwidth of this study was narrower than that of Bassett et al. (2012) (0.02–30 kHz) but wider than that of McKenna et al. (2012) (0.02–1 kHz), it is unlikely that the discrepancies in SLs result from the frequency ranges considered. One potential explanation is that commercial ships traveled at lower speeds in SFB than those reported by Bassett et al. (2012) in Puget Sound, Washington, USA and McKenna et al. (2012) in the Santa Barbara Channel, California, USA. Tankers traveled 3.7 and 2.7 kn faster, and vehicle carriers traveled 5.1 and 2.8 kn faster on average in Bassett et al. (2012) and McKenna et al. (2012), respectively, than these vessel types traveled in SFB. In contrast, median SLs reported by Veirs et al. (2016) in Haro Strait, Washington, USA for tankers and vehicle carriers were lower by <5 and <2, respectively, than in SFB, despite wider broadband bandwidth (0.02–40 kHz) and faster average speeds.

Our resulting median broadband SL for bulk carriers was within the inter-quartile range of those reported by Veirs et al. (2016) but below the range of estimated SLs reported by Bassett et al. (2012), McKenna et al. (2012), and Arveson and Vendittis (2000). While lower speeds in SFB offer a potential explanation for tankers and vehicle carriers, Arveson and Vendittis (2000) estimated SLs at speeds from 8 to 16 kn, which encompassed our average bulk carrier speed (11.1 kn). Smaller size offers a potential explanation for the low SLs we estimated (Chion et al., 2019). Bulk carriers had the highest standard deviation in vessel length (65 m) of all vessel types, and the three passages retained for broadband SL estimation had the shortest lengths of all bulk carriers observed ranging from 170 to 175 m.

Less systematic estimations of high-speed ferry SLs have been conducted than larger ships, but median broadband SL of ferries in SFB was within 2 dB of two unique high-speed ferries observed by Bassett et al. (2012) and within 1 dB of the average SL associated with one unique high-speed ferry observed by Hatch et al. (2008). Allen et al. (2012) reported a broadband SL associated with one high-speed ferry observation that was 40 dB greater than the median SL we estimated, but the vessel was 68% longer than the longest ferry observed in SFB.

Motorized recreational craft had the highest standard deviation in CPA speed (Table II) and the greatest estimated SL error of all vessel types (Fig. 5), which could be attributed to variation in vessels observed under broad classification. Because the RAM method of PE modeling is sensitive to defined source depth (Duncan and Maggi, 2005), the depth employed in PE modeling (10 m), which artificially extends the hull depth beyond expected depths for these vessels, may have resulted in higher TL error. Despite these considerations, median broadband SL estimated for motorized recreational craft fell within the inter-quartile range of broadband SLs associated with pleasure craft reported by Veirs et al. (2016). Wladichuk et al. (2019) report an average broadband SL (0.05–64 kHz) for monohulled vessels traveling faster than 15 kn (average speed of motorized recreational craft in this study was 16.3 kn) that was within 4 dB of the broadband SL we estimated.

There has been considerable effort to quantify SLs for individual vessels (Leaper, 2019) with some studies incorporating >1000 unique records. This growing collection of research helps inform future studies and contextualize results, while also expanding sample size for determining functional relationships between SL and vessel characteristics, environmental conditions, and operations. Overall, higher vessel speeds have been associated with higher broadband SLs most consistently (Leaper, 2019), supporting the power law relationship between propeller cavitation and SL first suggested by Ross (1976).

An analysis of variables compared to SL was not within the scope of this research, but we provide a broad comparison of estimated SLs with normalized vessel and passage variables (Fig. 7). B2 SLs of crude oil tanker passages were lower at higher speeds, and the relationship was significant (p < 0.05). This result is unexpected given the established positive relationship of tanker SL and speed at higher frequencies, due to influence of propeller cavitation noise at frequencies above 1 kHz (MacGillivray, 2019). Distribution of CPA distances showed two peaks (0.9 and 2.3 km) which correspond to distances from the hydrophone to the North Ship Channel and Richmond Harbor RNAs, respectively. The CPA depth differential across these two RNAs was approximately 10 m (Fig. 1), and ships traveled at higher speeds when CPAs were in the North Ship Channel compared to Richmond Harbor. The negative relationship of crude oil tanker passage SL and depth at the source, which was also significant, suggests confounded relationships between SL and speed. The significant positive relationship of B1 SLs and length of bulk carriers supports the explanation of our low SL estimates for smaller bulk carriers.

Research on the extent to which vessel traffic masks acoustic signals has largely focused on reduction of communication space among conspecifics [e.g., Williams et al. (2014a), Cholewiak et al. (2018), and Gabriele et al. (2018)], with less attention on masking of acoustic cues used for collision avoidance. Studies draw on comparisons of frequency and SL of vocalizations with noise to determine the extent of signal masking. Results of modeled RLs compared to NLs here should be approached with caution as ambient estimations include vessel presence. Collision avoidance requires detection of oncoming vessels and appropriate maneuvering to clear the area (Gannier and Marty, 2015). Since many ferries and motorized recreational craft in this research traveled at speeds >32.5 kn (16.7 m s⁻¹) and >16.3 kn (8.4 m s⁻¹), covering more than 1 km and 0.5 km in a minute, respectively, individuals have little time to appropriately maneuver out of the vessel's path. While there have been efforts to understand and improve detection of cetaceans from vessels [e.g., Gende et al. (2019)], this likely becomes more difficult at high speeds. Vessel detection and collision avoidance tactics used by cetaceans are not fully understood [e.g., Szesciorka et al. (2019)], and behavioral responses vary based on activity (Williams et al., 2014b). Future studies of acoustic masking in SFB should further investigate ambient conditions with no vessels present, consider ambient conditions outside SFB, similar to Pine et al. (2016), as cetaceans regularly travel in and out of the bay (Stern et al., 2017), and consider behavioral responses specific to unique populations present.

While specific conclusions drawn based on cumulative SEL should be considered with caution, as values were integrated over 24 h and assume a fixed receiver, maps of SELs (Fig. 6) inform relative comparisons between vessel types. Large ships had some of the highest SLs and contributed similar top SELs as ferries and motorized recreational craft, but the highest SELs from these vessels covered less physical space (Fig. 6), likely due to infrequency and minimal spatial extent of transits (Cope et al., 2020). Large ships made 11 transits through the study area per day on average compared to 58 transits per day by ferries. Because of the consistent paths of commercial vessels, it is not surprising that SEL maps were similar to distributions of presence. The highest SELs as a result of large ships occurred where the paths of vessels transiting the two RNAs likely overlapped near the southwest boundary of the study area (Fig. 6). Relative SELs were different from the energy budget calculated by Bassett et al. (2012). Tankers and bulk carriers contributed similar SELs per area in SFB (Fig. 6), but Bassett et al. (2012) found relative contribution to the energy budget to be 2% and 16%, respectively. Tankers had a higher relative presence in SFB potentially due to oil refineries in Richmond, adjacent to the study area (Fig. 1).

Our results join other recent studies highlighting the importance of considering non-AIS vessels in assessment of noise in urbanized coastal areas (Marley et al., 2017; Hermannsen et al., 2019). Anecdotal reports of encounters between recreational craft and humpback and gray whales in Bay waters frequently appear in local SFB news, and there is a growing trend worldwide of injurious collisions between large cetaceans and sailing vessels (Ritter, 2012). While many collisions along the U.S. west coast may be undetected or not officially reported (Rockwood et al., 2017), most stranded cetaceans in the SFB area show evidence of physical injury consistent with collision (Barcenas-De la Cruz et al., 2017). Lower SLs of recreational craft may make them more difficult to detect by cetaceans, but more research is needed that distinguishes auditory disturbance from presence (Erbe et al., 2019). Despite lower SLs, motorized recreational craft contributed a large portion of the radiated vessel noise in the study area and were generally louder at higher speeds (Fig. 7) (similar to Wladichuk et al., 2019), thus making them an important component in the overall acoustic environment in SFB.

It is important to understand how sound from vessels is distributed so that as cetacean researchers learn more about cetacean use of SFB [i.e., Stern et al. (2017)], threats associated with noise can be evaluated spatially [e.g., Joy et al. (2019)]. Because noise can degrade and introduce risk in cetacean habitat when the two overlap, it can be managed in a spatially explicit framework. Spatio-temporal restrictions on noise levels in biologically important areas show promise in minimizing significant impacts (Hatch and Fristrup 2009; Weilgart, 2018). In addition to overlap, species-specific sensitivity to that noise should also be considered (Southall et al., 2019). Some attempts to specify acoustic indicators for assessing marine habitat quality have been focused on lower frequency sound, such as the European Union's suggested maximum annual average ambient noise level within unweighted one-third octave frequency bands centered at 63 and 125 Hz (EU, 2017). Research on vessel SLs, including our results, confirms that vessels produce sound at frequencies above these bands.

Reducing vessel speed is one measure that can reduce vessel noise with the additional benefit of reducing chance of lethal collision (Leaper, 2019); mandatory speed restrictions have successfully reduced collision risk to north Atlantic right whales (Eubalaena glacialis) off the U.S. east coast (van der Hoop et al., 2015). In a recent voluntary vessel slowdown trial conducted by the Port of Vancouver through important killer whale (Orcinus orca) habitat, participating vessels reduced SLs by 5.9 dB for bulk carriers, 9.3 dB for vehicle carriers, 6.1 dB for tankers, and 10.5 dB for passenger ships (MacGillivray et al., 2019). While reducing speeds results in longer transit times, and thus higher SELs, previous research has maintained a net benefit of noise reduction at slower speeds (MacGillivray et al., 2019). Further investigation of SLs and speed could confirm if this would hold true in SFB.

In shipping lanes outside SFB, NMFS recommends a speed restriction of 10 kn (USCG, 2017), although this area does not extend to SFB waters. Within SFB, navigation is highly coordinated for maritime safety (Code of Federal Regulations 33 CFR § 161.50). While vessels greater than 1600 gross tons must transit at or below 15 kn within RNAs (Code of Federal Regulations 33 CFR § 165.1181), the minimum speed at which to transit safely for these larger ships can be influenced by tidal currents in the narrow inland waterway (California Code of Regulations 14 CCR § 851.9). Smaller vessels, like ferries and recreational craft, do not meet the size requirement for the speed limit. Speeds greater than 10 and 30 kn are common for commercial ships and ferries, respectively, within the study area (Cope et al., 2020). Reducing these speeds could potentially minimize collision risk as well as reduce vessel SLs and SELs across the area.

We took advantage of available data collected in the same location and time period and so were somewhat limited by existing configurations, including a single measurement aspect between source and receiver. By taking measurements at varying inclination angles, previous research has found significant influence of aspect on SL (Gassmann et al., 2017), and thus systematic estimation of SLs using hydrophones positioned at multiple depths is recommended by ISO (2016), although the standard is exclusive to deep water. By averaging resultant SLs from hydrophones at varying depths, the influence of surface reflections (which can lead to underestimation of SLs) can also be minimized in shallow environments (Brooker and Humphrey, 2016). A single hydrophone at a low inclination angle in this study likely underestimated true SLs (Gassmann et al., 2017; Erbe et al., 2019) in a very shallow environment. Since recreational craft were primarily detected via radar and thus could not be confirmed by photograph at night, daily cumulative SEL values likely underestimate true noise contribution from these vessels and thus overall SEL from all vessel types.

Other considerations could have resulted in overestimations of radiated noise. Recordings showing artifacts of current flow around the hydrophone and wave action against nearby structures is a common issue (Erbe et al., 2019), although various methods for tethering the hydrophone were evaluated to minimize environmental noise. Recordings also likely captured industrial noise as cargo transfer occurs at the Ports of Richmond and Oakland (approximately 10 and 18 km from the hydrophone, respectively). Allen et al. (2012) found that noise radiated asymmetrically from vessels with an acoustic shadow zone at the bow. Because of the hydrophone location and traffic patterns in the area, we only incorporated RLs associated with the starboard or port side of vessels in SL calculations and assumed that radiated noise from vessels was omnidirectional in SEL calculations, thus potentially overestimating SEL at a 0–30° segment in front of vessels.

Complex geography of SFB and primarily manmade shoreline in central SFB likely influence sound propagation in conjunction with sea surface, water column, and seabed properties (Jensen et al., 2011; Gannier and Marty, 2015; Gassmann et al., 2017). Future research would benefit from fine-scale incorporation of environmental variables and additional hydrophone sensors to help rectify differences in SLs of unique vessels. When the spatial and temporal extent of cetacean habitat in SFB is more well understood, studies could potentially be conducted over smaller targeted areas that may alleviate the complexity of modeling sound propagation across varying depths. To minimize error, we employed the same equipment and location for the study period, which was on a temporal scale small enough to eliminate influence of seasonal changes (McKenna et al., 2013; Gannier and Marty, 2015).

Opportunistic research on SLs in areas of high shipping activity along the U.S. west coast has been focused on certain geographic regions, with the SFB region receiving relatively less attention than the Santa Barbara Channel approaching the Ports of Los Angeles and Long Beach [e.g., McKenna et al. (2012), McKenna et al. (2013), and Gassmann et al. (2017)] and the boundary waters between Washington state and British Columbia, Canada where the Ports of Seattle and Vancouver are located [e.g., Bassett et al. (2012), Veirs et al. (2016), and MacGillivray (2019)]. In addition to major industrial shipping, concentrated high-speed ferry and recreational traffic (Cope et al., 2020) likely make SFB unique, while also creating a complex underwater soundscape. The presence of recreational craft makes it imperative that data beyond AIS broadcasts be incorporated into acoustic assessments in SFB. While previous opportunistic studies on vessel noise have benefitted from the systematic collection and availability of AIS data, the M2 system extended systematic data collection to include recreational craft by integrating radar and AIS data in a single platform, enabling a comprehensive analysis of vessels common in SFB.

This research provided spatial information on noise levels and frequency for future consideration of impact on local cetacean species. Evaluations of these impacts from vessels have been similarly concentrated (Erbe et al., 2019), primarily off the coasts of southern California (Redfern et al., 2017) and Washington state (Williams et al., 2014a; Joy et al., 2019) along the U.S. west coast. Further research on the spatiotemporal use of SFB by cetaceans will help determine when and where mitigation of noise and impacts from vessels is necessary and if speed or other restrictions could be beneficial. In doing so, it will be important to consider the appropriate frequency range of sound from vessels for those specific cetacean species being considered.

This work was made possible by the support of the Marine Science program at the Estuary & Ocean Science Center of San Francisco State University. The authors also wish to thank the San Francisco Bay chapter of the American Cetacean Society for their partial funding of this research.