Underwater noise pollution from human activities, particularly shipping, has been recognised as a serious threat to marine life. The sound generated by vessels can have various adverse efects on fish and aquatic ecosystems in general. In this setting, the estimation and analysis of the underwater noise produced by vessels is an important challenge for the preservation of the marine environment. In this paper, we propose a model for the spatio-temporal characterisation of the underwater noise generated by the fishing vessels in the Northern Adriatic Sea. The approach is based on the reconstruction of the vessels' trajectories from AIS data. Trajectories are enriched with semantic information like the acoustic characteristics of the vessels' engines or the activity performed by the vessels. This is then used to infer how noise propagates in the area of interest. The conceptual framework has been implemented using MobilityDB, an open source geospatial trajectory data management and analysis platform. We present some preliminary analyses obtained by applying the developed tool.
kinds of seabed floor as well as environmental variables
such as temperature, salinity, waves height, etc., allow
The Northern Adriatic Sea area is one of the most ex- the simulation to be very accurate. Still, covering a wide
ploited areas of the Mediterranean Sea and the underwa- area, such that the one of the Northern Adriatic sea, is
ter noise pollution is certainly among the efects of the challenging and, in any case, the estimated noise values
intensive fishery activity. Underwater noise produced by are not guaranteed to be correct.
vessels has a significant short and long term impact on In this paper we focus on the noise generated by
fishanimal species. Among the adverse efects we can men- ing vessels and we propose an approach to provide a
tion communication interference, behavioural changes, characterisation of underwater noise produced by such
stranding and mortality [
However, determining underwater noise maps is a ated by the vessels spreads in the area of interest. We
complex and resource-intensive process. Gathering data build on our previous work [
Northern Adriatic Sea. Specifically, we associate each the higher the frequency, the faster the disappearance of
ifshing vessel with its estimated Sound Pressure Level the intensity along with increasing distance.
(SPL), computing the intensity of the produced underwa- Paper [
The proposal has been implemented in MobilityDB [
The paper is organised as follows. Section 2 briefly area and time period, or to produce videos illustrating
presents the literature on underwater noise. Section 3 de- the underwater noise dynamics in timelapse.
scribes our model for the underwater sound propagation. Diferently, the aim of [
Summing up, many of the presented works use
mea2. Related works sures obtained via hydrophones to reconstruct
underwater sound. We propose a complementary approach
Underwater noise arising from human activities is known based on AIS data and on the construction of
semanto have a number of adverse efects on aquatic life. These tic trajectories. We use the information on the engine
can range from acute efects such as permanent or tem- power the fishing vessels are equipped with to
deterporary hearing impairment to chronic efects such as de- mine the source noise levels generated by the vessels and
velopmental deficiencies and physiological stress [
The work in [
The estimation is based on real measurements. We The computation of the transmission loss considering
use the dataset delivered by the project SOUNDSCAPE, all these parameters is not a simple task and for this
which carried out an acoustic monitoring in the North reason various models have been introduced. Before
Adriatic Sea from March 2020 to June 2021. Nine monitor- discussing the transmission loss, however, we focus on
ing stations were set up encompassing diferent environ- how to evaluate the source level, that is, in our case, the
mental characteristics. Two datasets have been realised noise generated by the fishing vessels.
composed of 20 and 60 seconds averaged Sound
Pressure Levels (SPLs) data in a wide range of frequencies 3.2. Source level estimation
recorded at the nine stations. These datasets are available
on Zenodo (https://doi.org/10.5281/zenodo.7472152) and
some analyses are in [
The principal sources of underwater noise are machinery, propellers, and cavitation. Our AIS dataset includes some data of the fishing boats, such as the length overall (LOA) of the boat, the horsepower of the engine and also the ifshing gear used. However, the dataset does not include direct measurements of the sound pressure levels of the ifshing vessels. So, we need to infer such values considering the general literature about underwater noise and the measurements provided by the SOUNDSCAPE project.
A first issue is how to evaluate the increase of noise
3.1. Sound propagation model when a trawler is in action. Measurements with
research vessels have 10dB of additional noise when
trawlThe basic objective of noise modelling is to assess how ing, regardless of speed, which can be much lower with
much noise a particular activity will generate in the sur- trawl [18]. We adopt the same increase in our model.
rounding area [
Finally, in order to associate the source levels to all the other vessels, we need to relate the sound pressure level to the engine horsepower, the latter being available in our dataset. If we assume that a constant fraction of engine power gets converted into acoustic power (i.e. acoustic power scales linearly with horsepower), this means that 3dB are added per doubling in engine power. We adopt such a linear progression on logarithmic scale of engine power. For example, for engines between 100Hp and 835Hp we obtain a range between 134dB and 143dB.
elling underwater noise, and its simplicity belies considerable complexity in the task of computing the transmission loss in order to estimate the received noise.
Sound propagation is profoundly afected by some
factors such as the conditions of the surface and bottom
boundaries of the sea as well as by the variation of sound
speed within the ocean volume [
Sound propagation speed is also influenced by various chemical-physical factors such as temperature, salinity
a spherical spreading law, of the following form, where
is the distance from the noise source in meters [
If we consider again the measurements for trip 1001 equally-powerful sound sources, the two contributions described in the previous section, a simple linear regres- would add up in magnitude and phase. However, dission on log2() gives, with multiple R-squared equal to tinct and independent sources, such as two boats, can 0.77, the estimated slope of -6.15dB, which is close to be treated as incoherent sources. Even in a narrow frethe theoretical -6dB per doubling of distance of spherical quency band, there will be a random phase diference spreading [19] (inverse square law). In practice, given a between the two sources. Therefore, the noise in a 1/3 reference reliable measurement in realistic condition, we octave band around 63Hz (or in any other band) gets can justify the adoption of simple spherical propagation. increased by 3dB if there are two equal contributions, by
Simple geometric spreading does not take into account 6dB if there are four equal contribution, etc. [22]. More the environmental characteristics that are needed for a precisely, what does add are the intensities, after invermore accurate estimation of transmission loss. There- sion of the logarithmic function that defines the decibel. fore, it can only be used in uncomplicated propagation Generally and precisely, if we have sources reaching a scenarios, or at frequencies that are barely afected by en- cell with diferent values of RL, the total noise level is: vironmental features. In more realistic models, what we want to consider are all the environmental aspects that = 10× 10(101/10 +. . .+10/10) (5) influence the sound propagation underwater, by adding a term proportional to distance from the source [19]: = + × (3)
(2) The received noise level (RL) at a given point is computed starting from Equation 1. However, the formula does not consider the ambient (or background) noise, which is present in the marine environment. In fact, when the noise generated by a source is the same as that of the background, this noise is no longer distinguishable and, consequently, it is no longer perceived. The formula for calculating then becomes the following:
= − − where is the sound source level, the transmission loss and the ambient noise.
We use the SOUNDSCAPE measurements [
In the literature there are several models for predicting
the absorption of sound in sea water which retain the We develop a framework for the spatio-temporal
characessential dependence on temperature, pressure, salinity, terisation of underwater noise. We partition the Northern
acidity and other environmental features. In the Francois Adriatic Sea into a regular grid composed of square
spaand Garrison model [20] the general equation for the tial cells (1km× 1km) and we estimate the noise generated
absorption of sound in sea water, at a given frequency by the fishing vessels in any cell at regular time intervals
, is given as the sum of contributions from boric acid, (every 20 seconds).
magnesium sulfate, and pure water. At a frequency below First, as described in Section 4.1, starting from the AIS
100Hz only the first contribution is relevant, although data, we create a set of semantic trajectories representing
is approximately of the order of 10− 6/ [19]. the behaviour of the fishing vessels. For this task, we
build on our previous work [
The first step consists in reconstructing the trajectories
of the fishing vessels starting from terrestrial Automatic
(4) Identification System (AIS) data, i.e., the AIS data sent by
ships and received by ground stations on the Italian coast
of Northern Adriatic Sea. AIS data contains the
identiifer of the vessel, called MMSI, its position and the time
instant of the bearing, together with other information,
like speed and course. Since boat positions are recorded
every 10-20 seconds, that correspond to a small spatial
displacement of the boat, trajectories are reconstructed
by linear interpolation of the AIS data. Next, in order
to organise the data into distinct trajectories followed
by the fishing vessels, also called trips, the continuous of the spatio-temporal point , i.e., its coordinates, and
movement of a vessel is split according to several criteria d(1, 2) for the Euclidean distance between two spatial
(see [
A reconstructed trajectory consists of a sequence of segments obtained by connecting consecutive AIS points. Algorithm 1 Given ℛ, T, G and , the algorithm The next step is to enrich such trajectories with diferent computes the total received noise level for each ∈ G kinds of semantic information, called aspects, following 1: Let : map⟨cell , float⟩ the MASTER model [23]. The model distinguishes among 2: for each ∈ ℛ do long-term aspects, (associated with the full trajectory), 3: for each ∈ T do volatile aspects (associated with the segments) and perma- 4: p = ((), ) nent aspects (associated with the fishing vessel, derived 5: = ..ℎ + 10 · . ℎ from the MMSI). A long-term aspect is the length and the 6: = 10(− )/20 duration of the trajectory whereas a permanent aspect, 7: for each = (, ) ∈ G. d(.ctd , ↓ 1) < do defined for this specific work, is the sound level asso- 8: = d(.ctd , ↓ 1) ciated with the engine horsepower of the vessel. This 9: = − 20· log10()− . · − aspect is computed as specified in Section 3.2, and it is 10: [] = [] + 10/10 denoted by .hpSL. A crucial volatile aspect is the 11: for each ∈ G do activity carried out by the fishing vessel. We consider the 12: . = 10 · log10([]) following activities: in port, entering to and exiting from the port, navigation and fishing . The in port, entering to The noise is estimated every 20 seconds, i.e., in the port and exiting from port situations can be deduced from time instants belonging to T. The centroids of the grid the position of the extremes of the segment w.r.t. the port cells are considered as listening points (we have 43, 386 area. If none of the previous cases applies, the fishing of these points), and consequently the noise perceived by or navigation activities are established on the basis of a centroid models the noise in all the points of the cell at the average speed of the boat. This aspect is of funda- a certain time instant. mental importance for the underwater sound propaga- In order to build the noise maps, we get the positions tion model, because when a boat is fishing it produces a of all the fishing vessels at the same time instants, i.e. much more intense sound. Given a spatio-temporal point every 20 seconds. We use these points to calculate the = ((, ), ) belonging to a segment ((, ) ∈ ) in noise generated by the fishing vessels (line 5) obtained a certain time interval ( ∈ ), we set .fishingto 1 by adding to the sound level associated with the boat if the activity associated to the segment during is (mmsi.hpSL) the noise due to the fishing activity ( 10dB ifshing , and to 0 otherwise. as explained in Section 3.2) if it occurs in . In line 6, the sound propagation radius (expressed in meters), 4.2. Construction of the noise map i.e. the distance at which the noise generated by the ifshing vessel is no longer perceptible, is computed. This is obtained by using Equation 4 and setting the received noise () to 0: In this section we describe the procedure for assigning a noise level induced by the fishing vessels to the cells of a regular grid, partitioning the Northern Adriatic Sea, every 20 seconds. We define G = S × T as a set of spatiotemporal cells, where S is a regular grid consisting of 1km× 1km spatial cells, and T is a set of time instants, such that 0 is a fixed time instant and +1 = + 20.
Hence each spatio-temporal cell ∈ G consists of two components, (, ), representing the spatial cell at time instant , and it has three annotations: (i) stores the absorption of sound as defined in Section 3.3; (ii) ctd contains the coordinates of the centroid of ; (iii) rl records the total noise perceived in , i.e., by the centroid of at time instant .
Let ℛ be the set of the trajectories of the fishing vessels, G be the spatio-temporal grid, and be the ambient noise for the Northern Adriatic Sea, equal to 85dB, as reported in Section 3.4. Algorithm 1 computes the total received noise level for every cell ∈ G. We use ↓ 1 to denote the projection on the first component 0 = − − In computing the radius we ignore the coeficient of absorption in Equation 3 for and with some simple mathematical steps we get = 10(− )/20. Note that since is very small (on the order of 10− 7/) ignoring its contribution simplifies the calculation while producing a negligible approximation. Also observe that in this way we overestimate hence the approximation is safe. Then, we propagate the noise in the cells that are within the radius (lines 7-10), summing up the received noise levels. Finally, by using Equation 5, we combine all the received sound levels to obtain the total noise level to be associated with the cell.
Concerning the complexity, let = | ℛ|, = |T|, = |G|, be the area of a grid cell and the largest radius arising in line 7. Then the complexity is ( · · 2/ + ). The factor 2/ is motivated by the fact every 20 seconds in the cell and dividing it by the number of values. Instead, Figure 2 reports the peak maps for each month: each cell is characterised by the maximum noise detected in that cell (the peak of the month in the cell). It is worth recalling that Equation 4 is employed for computing the received sound level in each cell. Thus, the sound depicted in the maps represents the noise exceeding the ambient noise perceived by each cell centroid. that in line 7 we consider the cells in a neighbourhood of radius . Note that depends on the source level, Figure 1 shows that in April 2020 some zones in the which is bounded by the maximum engine power of the central and in the central-southeast area are totally not exmonitored fishing vessels (in our case ≤ 3, 548). plored compared to January 2020. In particular, there are
In order to process all this data and build our model, more cells with a medium-low underwater sound (0-3dB we used a machine that features 32 Intel(R) Xeon(R) CPU on average) in April (3823 more). This perfectly reflects E5-4610 v2 processors running at 2.30GHz, ofering mul- the reduction of boats and trips during the COVID-19 pantithread performance. It is equipped with 256GB of DDR4 demic period. Instead, in April, there is a slight increase ECC RAM and it utilises a 500GB RAID 5 storage con- in cells characterised by a high average underwater sound ifguration. We evaluated the time for constructing the value (>4dB), 499 additional cells with respect to January, model assuming a daily data processing and consider- located mainly near the coasts. This phenomenon could ing 7 days in January 2020. With an average of 400, 499 be explained by the fact that, during COVID-19, vessels AIS data and 1, 050, 651 timestamps (|T|) per day, the reduced the navigation time, preferably staying near the construction of the model requires about 34 minutes. coast, thus limiting the fuel consumption and the related costs. A consequence of this behaviour is the increase of 4.3. Preliminary analyses and results the average sound level in coast areas which have seen a larger concentration of vessels.
For the implementation, we used MobilityDB [
We next present some analyses performed by using ager, it is possible to generate animations which, for a our spatio-temporal characterization of the underwater selection of vessels, visualise the noise propagation deternoise. For our experiments we focus on two months, mined by these vessels moving in the Northern Adriatic January and April 2020, with the aim of investigating Sea. The user can choose the boats according to sevthe efect of the COVID-19 pandemic outbreak on under- eral criteria, such as the range of horsepower, the MMSI, water sound pressure levels due to the fishing activities the length overall, or the activity, and the time window in the Northern Adriatic Sea. January 2020 represents of the analysis. In Figure 3 we can observe a diferent a pre-COVID-19 period with normal fishing activities sound propagation depending on the engine power of while April 2020 is a month where several containment the vessel and its activity. We focus on four vessels: measures were adopted. In particular in Italy a lockdown has engine power 959.4Hp (SL 146), vessel 246.4Hp was imposed from March 9, 2020 until May 18, 2020. Ta- and 335Hp (having the same SL 140) and vessel ble 1 reports the number of vessels, AIS data and trips 679.6Hp (SL 143). Vessels and are navigating but in these months. The first remark is that there is a clear not fishing (red dot) and, as expected, the sound propdrop in the number of boats during the pandemic period, agation is limited and it is greater for vessel which leading to a dramatic reduction in AIS data and trips. has a greater engine power. Vessels and are fishing
We generated two diferent maps to compare the un- (light blue dot): this increases the noise level and thus the derwater sound in these two periods. Figure 1 reports the propagation radius. It is worth noticing the diference average underwater noise value estimated in each cell between and having the same SL but is fishing, for January and April. The average value for a cell is cal- and between and , with having a greater engine culated by summing up the underwater noise computed power and fishing. These comparisons highlight how (a) January.
(b) April. a fishing vessel generates more substantial underwater noise than a boat merely sailing, even when the latter has higher horsepower.
In this work we proposed an approach for the characterisation of underwater noise in the Northern Adriatic Sea, based on the reconstruction of the fishing vessels’ trajectories and on the propagation of their noise along the trajectories. The approach has many advantages. Since we reconstruct trajectories starting from AIS data, we can readily obtain noise maps at diferent time granularities, e.g., on a daily basis, or monthly (as shown in the paper) or seasonally. The framework can be used in absence of hydrophones, which can be expensive to install and maintain and cover a limited area. Besides, it allows for distinguishing contributions to the underwater noise from diferent ship types, not only fishing vessels but also tankers or cruise ships, provided that the AIS data are available. The noise propagation model used in our implementation, although very simple and calibrated on a single measured trajectory, is suficient to demonstrate the advantages of semantic trajectories for a first and prompt characterisation of underwater noise. More complex calculations, relaxing some of the assumptions and exploiting the available geographic as well as boat-related information, may actually be introduced for a more accurate characterisation at diferent frequencies, without altering the algorithmic and information structure.
A drawback of our approach is that the boats without an AIS transceiver cannot be modelled. It is therefore not possible to estimate their contribution to the total noise, which, as a consequence, could be underestimated. This means that areas of the sea showing high values of underwater noise are surely risky for the underwater world, whereas areas that result to be quiet could hide some untracked noise.
Ecosystem Program "Interconnected Northeast Innovation Ecosystem (iNEST)", CUP H43C22000540006 and by the MASTER project funded by the European Union’s Horizon 2020 research and innovation programme under the Marie-Sklodowska Curie grant agreement N. 777695. We thank Fabio Pranovi for providing us the AIS data and his valuable knowledge as domain expert.