*SPL* = *C-SEL* − 10 log10(183 d × 24 h/d × 60 min/h × 60 s/min/s) = *C-SEL* − 72 dB re 1 s

### *2.8. Wind Map*

The wind map was produced by converting the hourly root-mean square sound pressure levels to linear mean-square sound pressures, then integrating over time, and converting back to decibel. 10 log10(3600) was added to account for the number of seconds per hour, yielding cumulative sound exposure levels from wind at each cell over the 6-month winter period.

#### *2.9. Comparison between Ship and Wind Noise*

For comparison, the cumulative sound exposure levels from wind were subtracted from those of ships (summed over all classes) in every grid cell, and plotted, to show in which geographic regions one dominated over the other. We also added the modelled sound exposures from ships and wind, then converted to decibel, to plot the combined ambient noise exposure levels over the 6-month period.

#### *2.10. Validation*

Archival underwater acoustic recordings from the northwest, west, south, and southeast of Australia were used in an attempt to validate the modelled noise maps. These data were collected by autonomous recorders [62] deployed over winter between 2006 and 2017. All recorders had been moored on the seafloor, and sampled at 6 kHz, 5 min every 15 min. Most of these datasets were collected while the passive acoustic observatories of Australia's Integrated Marine Observing System (IMOS) were operational and are thus available from the Australian Ocean Data Network (AODN) (https://acoustic.aodn.org.au/acoustic/ accessed on: 15 March 2021).

Long-term spectral averages (LTSA) were computed in 5 min windows and integrated over frequency (10–2000 Hz) and time (1 April–30 September) to yield cumulative sound

exposure. LTSAs were visualised in the software CHORUS [63] to provide an overview of the soundscape and its contributors over multiple weeks to months at a time. Spectrograms with a resolution of 1 s (50% overlap) were computed to zoom into any 5 min sample when the sound sources were not immediately identifiable in the LTSAs. Power spectral density percentile plots (in which the nth percentile gives the power spectral density level exceeded n% of the time, at each frequency) helped identify the dominant contributors to the winter soundscapes [64].

#### **3. Results**

Australia-wide maps of ship noise C-SEL (by class) over the period 1 April 2015– 30 September 2015 are shown in Figure 6. Cumulative sound exposure levels over all classes are also plotted. Wind noise C-SEL, the level difference between ship noise C-SEL and wind noise C-SEL, and the combined C-SEL of ship and wind noise are shown in Figure 7. Hyperlinks to the data can be found in the Data Availability section.

#### *Validation*

Modelled C-SELs of ship and wind noise are compared to measured C-SELs from the validation datasets in Table 1. There is good agreement (to within 3 dB) between model and measurement noise levels at sites 9 (NW Shelf, WA, Australia), 16 (Bremer Canyon, WA, Australia), and 25 (Tuncurry, NSW, Australia). The former two were dominated by strong wind, the latter by ships. Figure 8A shows almost continuous strong wind at the Bremer Canyon site, a faint Antarctic blue whale (*Balaenoptera musculus intermedia*) chorus throughout winter, peaking in May, and distant passes of ships. The cumulative energy from wind dominates and is the reason for the good agreement between model and measurement. Figure 8B shows briefer periods of strong wind off Tuncurry and a distant Antarctic blue whale chorus. The dominant feature of this soundscape were numerous passes of ships at close range, and this is the reason for the good agreement between model and measurement at this site.

Disagreement between model and measurement noise levels at other sites was due to unaccounted, additional, non-targeted noise contributions to the soundscape: marine animals and industrial operations. Figure 9 provides an overview of the biological contributors to the soundscape. The stereotypical sounds of Omura's whales (*Balaenoptera omurai*); Antarctic blue whales; pygmy blue whales (*Balaenoptera musculus brevicauda*); the unidentified source of the spot call, fin whales (*Balaenoptera physalus*); dwarf minke whales (*Balaenoptera acutorostrata*), and humpback whales (*Megaptera novaeangliae*) have been well described in the literature; as have Australian fish choruses [65–67]. These animals dominated the winter soundscapes near islands and reefs (sites 1, 4, 6, 9, 12, 14, 15, 17 and 18). Examples of soundscapes almost free from ships but noisy with animals are shown in Figures 10 and 11. Examples of soundscapes affected by anthropogenic noise are shown in Figure 12. At the time of recording, seismic surveying was the most common anthropogenic source that we did not model.

We were able to determine C-SEL variability over time at nearby sites. Winter recordings at sites 17 (2016) and 18 (2017) differed in C-SEL by 1 dB; these sites were only 80 m apart. Similarly, sites 7 (2006) and 8 (2010) were 4 km apart and the C-SEL differed by 1 dB. Moreover, sites 14 (2014) and 15 (2016) were 4 km apart and the C-SEL differed by 1 dB, showing good consistency over 1–4 years at nearby sites. Sites 19–24 were all within 3 km of each other. Recordings were from 2012, 2014, 2015, 2015, 2016, and 2017, respectively. The two simultaneous sets differed by 1 dB in measured C-SEL. The 2014 set had the lowest C-SEL with 179 dB re 1 μPa2s, and one of the 2015 sets had the highest C-SEL at 185 dB re 1 μPa2s, indicating the level of variability that may be expected from such in situ recordings over multiple years. There was no linear trend.

**Figure 6.** Maps of cumulative sound exposure levels (C-SEL) from shipping in the Australian EEZ, by ship class, and cumulatively over all classes. Maximum received C-SEL over the top 200 m of water were picked, representing a 'worst case' for animals that dive within this depth. Sound exposure was accumulated over 183 days (1 April 2015–30 September 2015). Levels can be converted to average mean-square sound pressure levels by subtracting 72 dB re 1 s. Note that the colour bars all start at 80 dB but the highest levels differ, reflective of the peak C-SEL for each class. The final map also shows 200 m and 3 km bathymetry contours.

**Figure 7.** Maps of modelled wind noise within Australia's EEZ during winter 2012 (1 April– 30 September: **top**), ship C-SEL less wind C-SEL (**middle**), and C-SEL from ships and wind combined (**bottom**). The black dots identify underwater recording stations used for validation. To convert to mean sound pressure level, subtract 72 dB re 1 s.





**Figure 8.** LTSAs [dB re 1 μPa2/Hz] near (**A**) the Bremer Canyon, WA, site 16, and (**B**) Tuncurry, NSW, site 25. The contributions from ships, wind, and Antarctic blue whales (*Balaenoptera musculus intermedia*) are marked in red, green, and black, respectively. Only a few ships are marked in (**A**).

**Figure 9.** LTSAs and spectrograms [dB re 1 μPa2/Hz] showing (**A**) an Omura's whale chorus and evening fish chorus at site 1; (**B**) Three Omura's whale calls and a fish chorus at site 5 (this specific example is free from ships and wind); (**C**) An Antarctic blue whale chorus and evening fish chorus at site 22; (**D**) Four Antarctic blue whale Z-calls and a fish chorus at site 23; (**E**) Pygmy blue whale song in front of the Antarctic blue whale chorus and a fish chorus at site 24; (**F**) Three spot calls at site 19; (**G**) Fin whale song at site 14; and (**H**) Humpback whale song at site 6. Note the changing x- and y-scales. All panels but H use a logarithmic y-scale. H uses a linear y-scale to stress the great bandwidth of humpback whale song (100 Hz–>3 kHz) in comparison to the narrow bandwidth of ship noise in this example (<100 Hz), resulting in humpback whales dominating the C-SEL after integration over frequency. An animal (fish?) biting on the hydrophone is marked by the white arrow. Sound from whales, fish, ships, and wind are marked in black, white, red, and green, respectively.

**Figure 10.** Winter soundscape at site 12. (**A**) Power spectral density percentiles showing domination by humpback whales from late June and fishes throughout. The curves follow the shape of the biological spectra 75% of the time (within black ellipse). The characteristic shape of wind is only seen in the absence of whales (lowest two percentiles within green ellipse); (**B**) LTSA of an evening fish chorus (within white box). A distant dwarf minke whale chorus (thin horizontal lines inside black ellipse) occurred in June–July. (**C**) LTSA showing humpback whales (within black box) and fish (within white ellipse).

**Figure 11.** (**A**) Power spectral density percentiles of the winter soundscape at site 15 dominated by pygmy blue whales (within black box). The characteristic spectral shape of a fish chorus is seen at 2–3 kHz in the 1st and 5th percentiles (dotted); (**B**) LTSA of the pristine soundscape at site 4 exhibiting multiple fish choruses at night (within white box), whose intensities vary with the phase of the moon; (**C**) Spectrogram of the soundscape at site 1 showing at least two simultaneous whale species (Omura's whales at 20–50 Hz and one other at 50–3000 Hz, uncertain) and a fish chorus at 300–500 Hz.

**Figure 12.** (**A**) Power spectral density percentiles of the winter soundscape at site 23 showing that pygmy blue whales (black box) were present the entire 6 months (because the spectral shape of their song is seen even in the 99th percentile, meaning it did not become quieter than this). However, the strongest sound in this soundscape came from ships (identified by the broad and smooth spectral hump between 20 Hz and 200 Hz; red ellipse). The spot call was also strong at this site (marked by the black arrow). The fish chorus at 800–2000 Hz (dotted box) was present the entire time as well; (**B**) Power spectral density percentiles from site 2 being entirely dominated by broadband industrial noise of unknown origin at this time; (**C**) Spectrogram of a strong seismic survey temporarily present at site 3; no other sounds were visible.

#### **4. Discussion**

The aim of our project was to develop a model for underwater ship noise in the Australian EEZ that could be used by industry and government to manage marine zones, their usage, stressors, and potential impacts. To put ship noise into context, we also modelled natural noise from wind under water. The models are based on numerous assumptions and involve a lot of averaging in space and time, leading to uncertainty. We therefore validated the models as a whole by comparing modelled sound exposure levels to measured underwater sound levels from 25 acoustic data sets collected over a 12-year span. Agreement was good when the underwater soundscape mostly contained the two sources modelled: ships and wind. Agreement was poorer when sound sources were missed (i.e., not modelled): seismic surveying, whales, and fishes.

Ship presence and movement were based on AIS data from the winter of 2015. Ships logged their positions at irregular time intervals, requiring that we interpolate between successive logs. We applied criteria for speed and direction continuity before straight-line interpolation, and where these were not met, we accepted holes in tracks, leading to an underestimation of ship time in the corresponding cells. We further had very few vessels in the smallest class (<25 m), as these mostly private recreational vessels do not log AIS positions. We therefore clearly underestimated their underwater noise contribution, in particular to coastal soundscapes. In addition, we did not take into account ships just outside of the EEZ and so underestimated noise levels near the EEZ boundary. Given that most AIS data were available for the larger and noisier vessels, we chose a monopole source depth corresponding to larger vessels (5 m) and applied this to all vessels in the model, for simplicity. The introduced uncertainty in modelled received levels is perhaps greater in winter (which we modelled) than summer, given all of our sound speed profiles exhibited a shallow surface duct of variable depth. Accounting for different source depths for the different vessel classes would require modelling sound propagation over the 64 cluster centroids in each zone multiple times, which we did not do, but could be done to improve accuracy. This might be desirable for more localised applications and modelling over smaller areas than the entire EEZ (e.g., regional seismic surveys or coastal developments). Placing the monopole at deeper depth than the propeller depth of small vessels during sound propagation modelling will likely enhance long-range received levels of the smaller, hence quieter, vessels, which are possibly underrepresented in the AIS data, meaning the errors do not add but work in reverse. Finally, the source levels produced by the RANDI model fall within the broadband quartiles reported recently [68]; however, the spectral shapes might differ. MacGillivray and de Jong [69] very recently showed that the RANDI model overpredicted source power spectral density below ~250 Hz for bulk carriers, vehicle carriers, tankers, container ships, and cruise ships, yet underpredicted source power spectral density above ~250 Hz. This might lead to differential errors in different regions (deep versus shallow water), depending on the efficiency with which sound below and above 250 Hz propagates in each environment. Other studies reported RANDI to overestimate [70,71] or underestimate, particularly above 200 Hz [72]. Underprediction of source levels by the RANDI model might be more common for the smallest vessels, in particular those with powerful motors, such as whale-watching boats and tugs [69,73–76].

In terms of the underwater sound propagation model used, the most common source of uncertainty is a lack of data on the seafloor composition and thus, acoustic properties. We used typical values from [51], but geoacoustic properties may vary from place to place. Hydroacoustic data (i.e., temperature, salinity, and sound speed profiles) were missing in some coastal zones and thus required spatial extrapolation. The equivalent fluid model applied is only approximate up to grazing angles of 50◦ and thus, more accurate for long-range propagation modelling. Modelling sound propagation only along bathymetry cluster centroids, instead of every source-receiver transect, introduced additional uncertainty. However, with a median water depth of 1809 m for all source cells in the entire EEZ, deviations of individual bathymetries from centroid bathymetries are likely to affect modelled received levels more in shallow and coastal rather than offshore waters. While

deviations in bathymetry from cluster centroids may change the pattern of constructive and destructive interference and thus yield rather variable received levels at specific locations in space and depth, there will not be a consistent bias in modelled received levels. Modelling along centroids will lead to both over- and underprediction, depending on range, depth, and frequency. These effects will be important on a fine spatial scale, but disappear on a coarse grid. Finally, the received level depends greatly on receiver depth. We chose to plot maximum received levels over the top 200 m, corresponding to the water layer in which most baleen whales travel. Any receiver depth (or depth range) may, of course, be picked from the model results, corresponding to specific animal depths.

The wind model we used was based on the classic review done by Wenz [43]. Other models, such as the Cato model [77] extend to lower frequencies and thus, yield higher levels (up to 2 dB) in high sea states. The Cato model would reduce the model-versusmeasurement difference (i.e., improve the wind noise prediction) at the wind-dominated sites (9, 16).

The map of underwater ship noise was based on AIS data from the year 2015, the map of underwater wind noise was based on wind data from 2012, and the in situ measurements were from various years (2006–2017). For a fine-scale model (i.e., small grid size), the exact positions and types of vessels would matter and therefore, validation with measurements from different years might be less successful. However, on a coarse grid, fine-scale variability averages out. For the ship noise map to differ by 3 dB, twice the number of ships (i.e., twice the power) would be needed. We showed close agreement in measured levels over consecutive years at the same sites, except when strong temporary sources occurred in some sets (e.g., industrial exploration) or when more variable, biological sources dominated in some years.

The geographic grid size chosen for the model might affect the received levels in some cells and change the ship-to-wind noise ratio. We modelled on a 5 km × 5 km grid, and so the source cells were assigned a received level at 2.6 km range. A 2.5 km × 2.5 km grid would have a mean receiver range of 1.3 km. If ships are evenly distributed within a 5 km × 5 km cell, then halving the grid size will increase received levels within the source cells by 20 log10(2) = 6 dB. The time spent in the source cell, however, will decrease by a factor 4, or, 10 log10(4) = 6 dB, making up for the increase in received level (i.e., decrease in propagation range and thus, propagation loss). If ships are unevenly distributed within the larger grid cell, then changing to a finer grid will yield a net increase in modelled received noise levels within source cells. In comparison, the modelled noise levels from wind, being a sheet (rather than monopole) source, will not vary with grid size as wind speed changes on a much larger spatial scale offshore. Therefore, in areas where shipping lanes are well-defined and narrow (<5 km wide), ship noise levels may exceed wind noise levels by more than modelled in this article.

Based on our model and its 25-point validation, the Australian EEZ has a higher proportion of natural underwater noise from wind over ship noise than the North Sea and likely other northern hemisphere oceans [18,32]. Part of the Australian marine soundscape appears pristine, if pristine is defined as an absence of anthropogenic noise and a richness of biological noise (see also [78]). We have shown that accurate models of the Australian marine soundscape must include biological sources (i.e., primarily whales and fishes). Natural biological and physical noise ought to be considered in management frameworks to provide context (e.g., for noise management in the Southern Ocean [79]).

Our recommendations for future work include the establishment of a databank of Australian ship source spectra as started by [80], which will allow replacing the RANDI model with monopole source spectra from actual measurements. We have shown that other anthropogenic noise sources cannot be excluded in areas and years where these dominate and their contribution to the marine noise budget should be assessed. Comparing long-term cumulative sound exposure might not be the quantity most useful to managers. Instead, sound energy could be integrated over much shorter time frames and maps of % time above certain management thresholds be plotted [81], which is likely more relevant to biological

receptors than an annual or seasonal integral or average. The different sound sources have different acoustic features (e.g., ship and wind noise are continuous, while seismic surveying and pile driving are pulsed) and bioacoustic impact is likely driven by different acoustic quantities (e.g., sound exposure versus peak pressure [82]). Therefore, different quantities will have to be mapped for different types of impact. Moreover, these sources exhibit fundamentally different sound radiation fields, where an underwater explosion is a monopole, a ship is a dipole, pile driving a line source, and wind a sheet source, requiring different modelling approaches.

**Supplementary Materials:** The following are available online at https://www.mdpi.com/article/10 .3390/jmse9050472/s1.

**Author Contributions:** Conceptualization, C.E., J.N.S. and D.P.; methodology, C.E.; software, C.E. and R.P.S.; formal analysis, C.E. and R.P.S.; data curation, C.E., D.P., J.N.S. and R.P.S.; writing original draft preparation, C.E., R.P.S., D.P. and J.N.S.; writing—review and editing, C.E., R.P.S., D.P. and J.N.S.; visualization, C.E., D.P., J.N.S. and R.P.S.; project administration, D.P.; funding acquisition, D.P., J.N.S. and C.E. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work was funded as part of the Australian National Environmental Science Programme, Marine Biodiversity Hub, under Project E2.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** A spreadsheet with the acoustic parameters that characterise each zone (i.e., mean winter sound speed profile; mean water depth and slope; sediment thickness; and compressional sound speed, shear sound speed, compressional absorption coefficient, shear absorption coefficient, and mean seafloor density) is available for download, as is a shape file of the spatially separated 28 acoustic zones (see https://tinyurl.com/3webp3pn). A spreadsheet with the sound speed profiles, water density profiles, and geoacoustic properties of the seafloor is available as Supplementary Material. Maps and data of cumulative sound exposure levels (C-SEL) from shipping over all ship classes in the Australian EEZ corresponding to Figure 6 can be found at Seamap Australia (https://seamapaustralia.org/; https://tinyurl.com/ahbs6nwr) and the AODN (https://catalogue.aodn.org.au/geonetwork/srv/eng/metadata.show?uuid=480847b4-b6 92-4112-89ff-0dcef75e3b84). Map and data of modelled wind noise from Figure 7 can also be found at Seamap Australia (https://tinyurl.com/a477j3b9) and the AODN (https://catalogue.aodn.org.au/ geonetwork/srv/eng/metadata.show?uuid=0d3c7edc-463a-4fa0-8039-4d5a779035c3). All sites last accessed on: 30 March 2021.

**Acknowledgments:** We sincerely thank Robert D. McCauley for providing the sea noise data sets used for validation.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **Appendix A**

Step-by-step process of modelling ship noise:

	- a. Find all grid cells that contain ships of any class, cast 36,100 km radials in 10-degree intervals, and extract bathymetry along the radials.
	- b. Cluster all extracted bathymetries (over all radials around all cells with ships) with a neural network and subsequent k-means into 64 clusters.
	- c. Compute sound propagation along each cluster centroid, for the centre frequencies of adjacent octave bands.
	- d. For each ship size class:
	- For each radial:
		- Look up into which cluster this radial went;
		- For each frequency:
			- -Retrieve propagation loss as a function of range and depth.
			- -Add octave band source level for this ship class.
			- - Add cumulative time that a ship of this class spent in this source cell to yield sound exposure level as a function of range and depth.
			- -Regrid from polar to Cartesian coordinates.

#### **References**

