Data Augmentation to Improve the Soundscape Ranking Index
Prediction
ROBERTO BENOCCI1, ANDREA POTENZA1, GIOVANNI ZAMBON1,
ANDREA AFIFY2,3, H. EDUARDO ROMAN2
1Department of Earth and Environmental Sciences (DISAT)
University of Milano-Bicocca
Piazza della Scienza 1, 20126 Milano
2Department of Physics
University of Milano-Bicocca
Piazza della Scienza 3, 20126 Milano
3NEXiD Edge
NEXiD
Via Fabio Filzi 27, 20124 Milano
Abstract: Predicting the sound quality of an environment represents an important task especially in urban parks
where the coexistence of sources of anthropic and biophonic nature produces complex sound patterns. To this
end, an index has been defined by us, denoted as soundscape ranking index (SRI), which assigns a positive weight
to natural sounds (biophony) and a negative one to anthropogenic sounds. A numerical strategy to optimize the
weight values has been implemented by training two machine learning algorithms, the random forest (RF) and
the perceptron (PPN), over an augmented data-set. Due to the availability of a relatively small fraction of labelled
recorded sounds, we employed Monte Carlo simulations to mimic the distribution of the original data-set while
keeping the original balance among the classes. The results show an increase in the classification performance.
We discuss the issues that special care needs to be addressed when the augmented data are based on a too small
original data-set.
Key-Words: - soundscape; data augmentation; machine learning; ecoacoustic indices; soundscape ranking index;
urban parks;
Received: April 9, 2023. Revised: July 2, 2023. Accepted: September 2, 2023. Published: September 20, 2023.
1 Introduction
Soundscape analysis has grown in importance dur-
ing the last decades as a tool for a rapid evalua-
tion of environmental sounds. The analysis relies on
non-invasive techniques for ecological surveillance
thanks to the use of a widespread monitoring tech-
nology known as passive acoustic monitoring (PAM)
techniques. The latter can provide large amount of
data for long periods of time even in hardly acces-
sible geographical locations. In this respect, ecoa-
coustics has emerged as a new branch of acous-
tics, which is based on the idea that acoustic ambi-
ence measurements can provide a significant infor-
mation on the sound sources present within a given
landscape. Thus, the contributions from different
sound sources, identified as geophony (physical), bio-
phony (biological) and anthropogenic (anthrophony
or technophony) sources, [1], [2], [3], [4], may pro-
vide an estimation of the health of a habitat across
space and time. The complexity of the assemblage of
sounds in an environment is usually summarized in
terms of ecoacoustics indices, which can be evaluated
over specific intervals of time, which map the acous-
tic dynamics of an ecosystem, [5], into time series
describing the observed status changes taking place
in a given habitat, [6]. The information carried by
the calculation of ecoacoustic indices are usually val-
idated by listening to hours of recordings in order to
WSEAS TRANSACTIONS on ENVIRONMENT and DEVELOPMENT
DOI: 10.37394/232015.2023.19.85
Roberto Benocci, Andrea Potenza,
Giovanni Zambon, Andrea Afify,
H. Eduardo Roman
E-ISSN: 2224-3496
891
Volume 19, 2023
ITALY
ITALY
ITALY
identify known sound categories. This activity is gen-
erally referred to as sound-truthing.
The manual procedure to identify the different
sound sources is typically very time consuming, re-
quiring specific technical preparationa and long-time
training. For this reason, this methodology is usu-
ally applied to small data-sets, [7], [8]. A summa-
tive approach has also been proposed by providing
qualitative information of the sound characteristics in
each audio recording, [9]. In this sense, the consid-
eration of just few or many bird sounds, few or many
bird species, and low or high traffic noise, etc., may
simplify and speed up the validation process. How-
ever, such information alone cannot provide an ac-
curate assessment of the acoustic quality characteriz-
ing a complex habitat. In order to estimate a sound-
scape index, an attempt has been recently made by
empirically accounting for the contribution of differ-
ent sound sources, which assigns a positive weight to
natural sounds (biophony) and a negative one to an-
thropogenic sounds. The results can be summarized
in terms of an index denoted as soundscape ranking
index (SRI), which has proved to disentangle com-
plex environmental sound blends such as those found
in urban parks, [10]. The analysis outlined in this pa-
per refers to the sound recordings taken at Parco Nord
(Northern Park) of the city Milan (Italy).
It is well established that machine learning (ML)
algorithms can be trained to learn generic features
extracted from a database, [11], [12], [13], which
can then be used for dealing with many useful and
complex tasks. For example, the application of ML
techniques for studying the dynamics of a sound-
scape has provided accurate identifications of differ-
ent bird species, [14], [15], and also the separation
of the different audio sources when they are mixed
in a set containing an unknown combination of their
components, [16]. In urban areas, ML models have
been used to predict long-term acoustic patterns from
short-term sound measurements, [17], and for filter-
ing out anomalous noise events before computing the
actual traffic noise maps, [18]. ML techniques have
been also applied for soundscape classification, [19],
[20], species recognition, [21], [22], and the identifi-
cation of mixed noise sources using a two-stage clas-
sifier which is able to discriminate different urban
acoustic events in real time, by relying on the nor-
mally present redundant information from a network
of acoustic sensors in the city of Barcelona, [23].
The implementation of a huge manual labelling
(about 60,000 sound recordings) has proved to be
satisfactory for the identification of different sound
sources, [24]. A huge dataset collected over four
years across Sonoma County (California) by citizen
scientists was used to predict soundscape components
with success, [25]. The automatic recognition of
the soundscape quality of urban recordings has been
studied by applying four different support vector ma-
chine (SVM) regressors to a combination of spectral
features, [26]. In [27], a combination of temporal,
spectral, and perceptual features was used to classify
urban sound events belonging to nine different cate-
gories. In [28], different acoustic indicators taken in
the city of Barcelona were used to train several clus-
tering algorithms with recognizing the possibility of
clustering the city area according to the noise levels.
ML has been also used in a preliminary work, to op-
timize the definition of SRI, [29]. Here, the set of
weights that define the SRI were searched by train-
ing four machine learning algorithms, Decision Tree
(DT), Random Forest (RF), Adaptive Boosting (Ad-
aBoost), Support Vector Machine (SVM), over a rela-
tively small number of labelled sound recorded audio
files.
More specifically, it was shown that two classifi-
cation models, DT and AdaBoost, were able to pro-
vide a set of weights characterized by a rather good
classification performance. The obtained results were
in quantitative agreement with two different statis-
tical approaches: a self-consistent estimation of the
mean SRI values at each site, [30], and a cluster anal-
ysis performed, over the extracted features at each
site, [31].
In this work, we studied the possibility of pre-
dicting the SRI at an urban park in the city of Mi-
lan (Italy) from the extracted spectral features of au-
dio recordings in the form of seven ecoacoustic in-
dices. Using these features, in [29], we obtained
quite poor classification scores. In actual facts, clas-
sification performance is strongly affected by unbal-
anced classes (in our case the number of recording
occurring in each sound quality class). To deal with
this issue, we aim here at improving the classification
scores by using Monte Carlo simulations to create a
larger and well-balanced labelled dataset.
2 Materials and Methods
The urban area of study is described, together with
the employed instrumentation. The method used is
briefly reviewed by reminding the reader of the def-
inition of the SRI and the way it is optimized using
machine learning methods.
2.1 Area of Study
Following our previous works, we discuss results on
the urban zone known as the Parco Nord (Northern
Park) in Milan city. The latter covers an area of about
800 hectares, located within a quite developed and ur-
banised zone. The park possesses a tree-covered par-
cel of little more than 20 hectares surrounded by few
agricultural fields, lawns, paths and roads (Figure 1).
The location of the park is delimited to the north by
WSEAS TRANSACTIONS on ENVIRONMENT and DEVELOPMENT
DOI: 10.37394/232015.2023.19.85
Roberto Benocci, Andrea Potenza,
Giovanni Zambon, Andrea Afify,
H. Eduardo Roman
E-ISSN: 2224-3496
892
trafficked roads, including a highway and a main city
street, at about 100 m from the wooded parcel. To the
west side of the park there is also a small city airport,
which is around 500 m from the tree-line edge.
2.2 Audio recorders
In our study, we employed commercial SMT Security
low-cost digital audio recorders. Their working set up
was chosen to be able to measure at a sampling rate of
48 kHz in an almost continuous mode. Devices with
similar characteristics in terms of frequency response
were selected in the measuring campaign. For the
analysis, we scheduled the recording to coincide with
greatest singing avifauna activity, taken on May 25th
2015, during the morning hours from 06:30 a.m. to
10:00 a.m. (CET). They correspond to 3.5 h for each
sensor and recording session. As mentioned in Fig-
ure 1, some sensors did not work properly and were
excluded. As a result, we were left with 16 sites,
while the total recordings data analyzed resulted in
1120 files.
2.3 Aural Survey
In order to quantify the quantify the presence of bio-
phonies, anthrophonies and geophonies, an aural sur-
vey was implemented. In particular, an expert lis-
tened carefully each one of the recordings according
to the following scheme: for every three minutes of
continuous recording, only one-minute was listened
and interpreted. This yielded a total of 70 min of
listening per site. In order to facilitate the task, the
expert considered only five sound categories: birds,
other animals, road traffic noise, other noise sources
(such as airplanes, trains). Then, information about
each type of sound source, such as its occurrence or
not presence, in addition to the intensity, were deter-
mined.
More specifically, four parameters were consid-
ered and evaluated: (1) Individual abundance, re-
ferred to simply as, no–few–many subjects. (2)
Perceived singing activity, represented in terms of
the fraction of time (percent) attributed to avian
vocalizations, (0 100 %). (3) Species richness,
also referred to as, none–one–more than one species.
(4) Vocalization intensity, again referred to as, no–
low–high intensity. Regarding other animals and/or
people, the indicator was considered as an either
present or absent event.
It was found that road traffic was the main anthro-
pogenic noise in the zone. For this source, only two
parameters were considered: (1) Noise intensity, rep-
resented as, no–low–high intensity. (2) Typology of
traffic, that is either continuous or intermittent traffic,
or no traffic at all.
2.4 The soundscape ranking index, SRI
We briefly review the definition of the SRI, intro-
duced as a simple criterion to estimate the sound qual-
ity of a local environment, [10]. Let us consider a
single audio recording labeled by the letter `. In this
case the SRI is defined as,
SRI`=
nc
∑
i=0
cNiNi,`,(1)
where nc+1 is the total number of identified cate-
gories (birds, other animals, road traffic noise, other
noise sources, rain and wind), here nc=4, Ni,r=1
if the ith sound category is present at the record-
ing r, and Ni,r=0 otherwise. The coefficients can
take one of the following values: cNi>0 (cNi→c+,
c++) when the sound category corresponds to a nat-
ural sound, in that case we split the values into two
possible sub-ranges (+,++); and cNi<0 (cNi→c−,
c−−) corresponding to a potential disturbing event,
also split into two sub-ranges. The absence of birds
vocalization is regarded as a neutral event, cN0=0,
setting the separation between natural and anthro-
pogenic events. In Table 1, we summarize the em-
ployed ranges of variability for the coefficients cNi.
Our definition Eq. (1) is expected to yield SRI val-
ues representative of the quality of the environmental
sound. For the sake of simplicity, we choose three
intervals of SRI to define the environmental sound
quality, for a single recording denoted simply as `,
given by
SRI`<0[poor quality],
0≤SRI`≤2[medium quality],(2)
SRI`>2[good quality].
2.5 SRI optimization procedure
To each sound category i, we assign a weight, cNi,
according to the attributes extracted from the audio
recording (see the column ‘Attribute’ in Table 2). For
instance, the singing activity gets a weight depend-
ing on the percentage of singing birds detected in
each recording: the interval (0,25]%, corresponds to
a weight of 0.25 ×c++, while the (25,50]% one be-
comes the weight 0.50 ×c++, etc. As a general con-
straint, we assume that cNican vary within the inter-
vals mentioned in Table 1.
For the purpose of the using the ML techniques,
both the spectral features associated to the ecoacous-
tic indices, and the corresponding SRIs need to be
split into a ‘training’ set, and a ‘test’ one. The former
is used as input for each classification model, whereas
the final performance of the latter is quantified ac-
cording to the classification metrics used, here em-
ployed the F1–score. In the process of classification,
the SRIs become the target variable which is expected
WSEAS TRANSACTIONS on ENVIRONMENT and DEVELOPMENT
DOI: 10.37394/232015.2023.19.85
Roberto Benocci, Andrea Potenza,
Giovanni Zambon, Andrea Afify,
H. Eduardo Roman
E-ISSN: 2224-3496
893
Figure 1: Aerial view of the Parco Nord, on which we have indicated the grid of sensors employed in the record-
ings (red and yellow dots). The red spots correspond to the effectively working sensors, while the fewer yellow
spots those which did not function properly and were excluded from the analysis. On the plot, we have high-
lighted the A4 highway and Padre Turoldo street to the north of the park. The Bresso airport can be seen at the
left side of the picture.
Table 1: Intervals of variation for the coefficient cNiwhich are assigned to each sound category, i=0,...,4, to be
employed in Eq. (1). In the present calculations, we have taken the total range of variation of the coefficients to
be: −5≤cNi≤5.
cNiRange
c++ [2, 5]
c+[0, 2]
cN00
c−[-2,0]
c−− [-5,-2]
to be predicted by the classification model. The pro-
cess is repeated for every combination of weights,
cNi, for i>0, that are varied in the corresponding in-
terval of variability. The optimal set of weights defin-
ing our target SRI is taken from the best classification
score obtained.
2.6 Classification models
It is convenient to briefly describe the classifica-
tion models used for predicting the manual labelling
sound categories and the SRI index from the ecoa-
coustic indices. As a matter of fact, ML learning
methods are suitable to capture the potentially non-
linear relationships among variables, which are not
known a priori. The models we have considered here,
which have been implemented in Python program-
ming language, [32], are the following,
• Random Forest (RF),
• Single-layer neural network or Perceptron
(PPN).
The RF is used to fit a given number of decision tree
classifiers on various sub-samples of the data-set, us-
ing an averaging procedure to improve its predictive
accuracy, and very importantly to also control over-
fitting, [33]. For RF, we use default parameters with
the exception of Max-depth = 3, typically devoted to
control the size of the tree to prevent over-fitting.
WSEAS TRANSACTIONS on ENVIRONMENT and DEVELOPMENT
DOI: 10.37394/232015.2023.19.85
Roberto Benocci, Andrea Potenza,
Giovanni Zambon, Andrea Afify,
H. Eduardo Roman
E-ISSN: 2224-3496
894
Table 2: Coefficient cNiassigned to each sound category to be used in Eq. (1).
Category Attribute cNi
no cN0
Birds singing few c+
many c++
no cN0
Birds species .2c+
>2 c++
0 0.00 ×c++
(0,25] 0.25 ×c++
Singing activity (%) (25,50] 0.50 ×c++
(50,75] 0.75 ×c++
(75,100] 1.00 ×c++
no traffic c+
Traffic type continuous c−
intermittent c−−
zero c+
Traffic intensity low c−
high c−−
Other sound sources absent cN0
present c−−
The PPN was the first and simplest type of artifi-
cial neural network worked out and reported in the
Literature, [34]. In a PPN network, the informa-
tion oscillates back and forth from the initial input
nodes, via the (if any) hidden nodes, toward the out-
put nodes. The simplest kind of neural network is
clearly just a single-layer Perceptron, consisting of
one layer of output nodes. The inputs are then fed di-
rectly to the outputs via a series of weights. The sum
of all the products of weights and inputs are then cal-
culated at each node. If the resulting value is above a
given threshold, the neuron fires and is updated as ac-
tivated value, otherwise it remains at the deactivated
value. For the Perceptron model, we used the follow-
ing settings,
• seven neurons as input layer with activation
function ReLU,
• three neurons as output layer with activation
function Softmax,
• number of epochs 100,
• learning rate 0.001.
In our calculations for the implementation of the
SRI optimization procedure, the supervised classifi-
cation models have been trained on the 80% of the
data, and tested on the remaining 20%. The split-
ting of the data have been performed using a stratified
procedure in keeping the proportions between the tar-
get variable classes. We chose to implement the RF
model and the PPN for essentially two reasons: The
former yielded a good performance over other classi-
fication models in similar contexts, [29], whereas the
latter represents the simplest neural network, both in
terms of complexity and computing resources.
2.7 Ecoacoustic indices
In this work, we focused on the following set of ecoa-
coustic indices: the acoustic entropy index (H), [35],
the acoustic complexity index (ACI), [36], the nor-
malized difference soundscape index (NDSI), [37],
the bio-acoustic index (BI), [38], the dynamic spec-
tral centroid (DSC), [39], the acoustic diversity index
(ADI), [39] and the acoustic evenness index (AEI),
[39].
The indices were evaluated using the R statisti-
cal package (in particular, the version 3.5.1, [40]).
Specifically, the fast Fourier transform (FFT) was
computed by the function spectro available in the R
package “seewave”, [41]. The calculations were re-
stricted to the frequency interval (0.1-12) kHz based
on 1024 data points, corresponding to a frequency
resolution FR = 46.875 Hz and, therefore, to a time
resolution TR = 1/FR = 0.0213 s. The indices were
calculated using the “soundecology” package, [42].
Finally, a dedicated script running in the “R” envi-
ronment has been written to calculate the DSC index.
For each one-minute recording, obtained as discussed
above, we computed seven cumulative indices. Each
recording was then represented by seven indices or
features.
WSEAS TRANSACTIONS on ENVIRONMENT and DEVELOPMENT
DOI: 10.37394/232015.2023.19.85
Roberto Benocci, Andrea Potenza,
Giovanni Zambon, Andrea Afify,
H. Eduardo Roman
E-ISSN: 2224-3496
895
2.8 Data augmentation by Monte Carlo
simulations
We used Monte Carlo simulations to generate a set
of random numbers with the same distribution as the
original data set, that in our case correspond to the
distribution of the seven ecoacoustic indices in each
sound quality class. For each combination of weights,
cNi, for i>0, varied in the assigned interval, a distri-
bution for each of the seven ecoacoustics indices was
derived. For each combination of weights, a Monte
Carlo data augmentation has been applied in order to
reproduce the original distribution in each class, and
care was taken in order to obtain a balanced distribu-
tion of classes. In our case, we managed to obtain
500 instances for each class. As an example, Figure 2
illustrates the density distribution of ACI of the orig-
inal data-set split into the three classes obtained for a
particular combination of weights (c++=2.7, c+= 1.5,
c−=-1.0, c−−=-2.5).
Operationally, each of the ecoacoustic indices is
first normalized, by subtracting its mean value and
dividing by its standard deviation. Then, the cumu-
lative probability is computed and, finally, its inverse
function is derived (Figure 3). A random number thus
corresponds to a value of ACI belonging to the origi-
nal distribution.
3 Results
We ran RF and PPN models to attempt a prediction of
the soundscape ranking indexes calculated assigning
a set of weights to each sound category. We varied the
different weights according to the intervals reported
in Table 2. Then, the best combination of weights
is the one which has the highest score provided by
each classification model. In our case, we decided to
employ the F1score and the accuracy criterion. The
F1score is usually suited for dealing with unbalanced
classes, whereas the accuracy method for balanced
classes. The results are reported in Table 3.
The best F1–score for RF yields 0.63 and 0.60-
0.61 for DNN (Table 3). This means that the simpler
RF model can provide a slightly higher classification
score.
3.1 Implementation of the Monte Carlo
simulations
For each combination of weights, the occurrences in
each class can substantially vary as it can be observed
in Figure 4, showing the counts distribution for each
class (Class 1, Class 2, Class 3), produced by differ-
ent weight combinations whose range of variability is
defined in Table 2.
The median values and the interquartile ranges
(IQR) for Class 1 and Class 2 show that the major-
ity of counts are unbalanced with respect to Class 3
(Table 4). This result may also be interpreted by con-
sidering that the majority of data files show a preva-
lence of Good (Class 3) SRIs.
This unbalanced distribution of classes may in-
troduce a bias in the classification process. One of
the methods to augment the data-set is to lean on a
Monte Carlo expansion of the data based on the dis-
tribution of the original data-set split into each class.
Monte Carlo data augmentation can effectively repro-
duce the original distribution in each class, thus lead-
ing to a more balanced distribution of classes. In our
case, we managed to obtain 500 instances for each
class.
At first, we compared the classification by using
the augmented data-set that, as we mentioned, have
been balanced to reach 500 elements for each class
using the same set of weights that are reported in Ta-
ble 3. The results are shown in Table 5, suggesting
that data augmentation and data balancing provide
quite different results than those obtained using the
original unbalanced data.
Since the expanded data represent a balanced as-
set for the classification models, we explored the ef-
fect of data augmentation on other combination of
weights. Thus, we first expanded the data on the ba-
sis of MC calculation, creating new data sets with
balanced classes and then we repeated the classifi-
cation for all the augmented data sets (360,000 files
obtained from the combinations of all the possible
weights in the range reported in Table 1). One of the
major concerns when using MC data augmentation is
the numerousness of the original class. Indeed, data
expansion based on few data (see data belonging to
whiskers of the boxplot of Figure4) would bias the
new augmented class. For this reason, we decided to
set the numerousness of the original class to be ex-
panded above 100 counts. The results of the distribu-
tion for the Accuracy measure, for both models, are
reported in Figure5. The two distributions are similar
showing an accuracy peak value of 0.50 and 0.53 for
PPN and RF, respectively.
In actual situations, we are mainly interested in
the highest scores yielded by the two classification
models. These results are reported in Table 6.
4 Discussion
The seven spectral features, derived by integrating
the ecoacoustic indices over the whole length of the
recording (1 minute), contain condensed informa-
tion about the spectral variability that could be repre-
sented by lower integration times. Therefore, it does
not appear to be enough to represent the complex-
ity of the soundscape in a single summative index.
However, both models yield a significant improve-
ment in classifying the new augmented dataset, in
going from 0.60-0.63 (F1–score) to 0.74-0.75 (Accu-
WSEAS TRANSACTIONS on ENVIRONMENT and DEVELOPMENT
DOI: 10.37394/232015.2023.19.85
Roberto Benocci, Andrea Potenza,
Giovanni Zambon, Andrea Afify,
H. Eduardo Roman
E-ISSN: 2224-3496
896
Figure 2: Density distribution of ACI of the original data-set split into the three classes obtained with the following
combination of weights: c++=2.7, c+= 1.5, c−=-1.0, c−−=-2.5.
Figure 3: (Left panel) Cumulative probability distribution calculated for the normalized ACI, by subtracting the
mean value and dividing by the standard deviation. (Right panel) Inverse function used for the MC simulations.
Table 3: Results for the RF and PPN models, using eco-acoustic indices as extracted spectral features. Clas-
sification measure (F1–Score with its standard deviation), range of weights values and class numerousness are
reported. Class 1, Class 2, Class 3, correspond, respectively, to: Poor, Medium, Good, soundscape quality.
F1–score c++ c+c−c−− Class 1 Class 2 Class 3
RF 0.63±0.12 2.0 2.0 [-1.4, -1.5] [-2.6, -2.7] 228 515 377
DNN 0.61±0.04 2.1 0 0 -4.8 432 364 324
0.60±0.06 [2.2; 2.3] [0.9; 1.3] [-1.9; -1.8] [-2.2; -2.0] [245; 276] [388; 413] [456; 462]
Table 4: Median values and interquartile ranges (IQR) for Class 1, Class2 and Class 3, corresponding to Figure4.
Median IQR
Class1 165 185
Class 2 170 120
Class 2 785 292
racy; F1–score measure is pretty similar). The set of
weights that realizes this new classification presents
some differences from those obtained in the original
unbalanced classification but are similar to each other
(RF and PPN with augmented data).
In Figure 6, we compare the SRI maps calculated
WSEAS TRANSACTIONS on ENVIRONMENT and DEVELOPMENT
DOI: 10.37394/232015.2023.19.85
Roberto Benocci, Andrea Potenza,
Giovanni Zambon, Andrea Afify,
H. Eduardo Roman
E-ISSN: 2224-3496
897
Figure 4: Counts distribution for each class: Class 1, Class 2, Class 3, corresponding to Poor, Medium, Good
soundscape quality, of the original data-set calculated over all the combination of weights.
Table 5: Accuracy measure calculated for RF and PPN models, using seven eco-acoustic indices as extracted
spectral features, 500 augmented data for each class and the same weights reported in Table 3.
Accuracy
RF 0.48
DNN 0.59
Figure 5: Density distribution of accuracy measure calculated from the augmented data. The peak of the distri-
butions is at 0.51 (dashed red line) and 0.53 (dashed black line) for PPN and RF, respectively.
Table 6: Best accuracy scores obtained for RF and PPN models, using 500 augmented spectral features per class.
Weights values and class numerousness of original data are also reported.
Accuracy max F1–score c++ c+c−c−− Class 1 Class 2 Class 3
RF 0.751 0.751 2.1 1.9 -0.0 -3.5 123 128 869
PPN 0.746 0.737 2.0 1.5 -0.1 -5.0 371 354 395
WSEAS TRANSACTIONS on ENVIRONMENT and DEVELOPMENT
DOI: 10.37394/232015.2023.19.85
Roberto Benocci, Andrea Potenza,
Giovanni Zambon, Andrea Afify,
H. Eduardo Roman
E-ISSN: 2224-3496
898
Figure 6: SRI maps obtained for the models: (a) PPN and, (b) RF.
over the study area based on the weights obtained for
the RF and PPN models with the augmented data. For
each of the 16 sites, we considered the median value
of the SRI computed over all the measurements corre-
sponding to the labeled recordings. As expected, the
results are very similar showing a clustering of sites
facing the traffic noise sources and those at the park
interior. These results are in fairly good agreement
with previous analysis obtained using a statistical ap-
proach, [30], [31].
5 Conclusions
The soundscape analysis in urban parks can assess the
distance between natural habitats and artificial/recon-
structed green areas. In this work, seven ecoacous-
tic indices, calculated over 16 sites of Parco Nord
of Milan, Italy, have been used to predict a single
index, named the soundscape ranking index, SRI,
which has the advantage of yielding a quick overview
of the quality of environment sound. SRI is com-
puted through the optimization of the weights which
appear in its definition (Eq. 2). Each combination
of weights provides a different numerousness of the
classes (Class 1, Class 2, Class 3 corresponding to
Poor, Medium, Good soundscape quality) of the orig-
inal data-set. These imbalanced classes may repre-
sent one of the factors influencing the classification
models. In actual situations, the use of two very sim-
ple classification models, RF and PPN, with the orig-
inal data-set, yielded a maximum F1-score of 0.60-
0.61. We applied MC calculations to balance the
three classes for each combination of weights and
recomputed the classification algorithms. From the
new computation, we ruled out those combinations
of weights with initial numerousness lower than 100.
This choice was justified to avoid the introduction
of artificial peaked distributions in the original data,
which once augmented would have affected the clas-
sification performance. In this way, we ended up with
the following classification performance expressed in
terms of Accuracy (usually employed for balanced
classes classification, but we also report the F1–score
for comparison with the previous calculations):
• RF : Accuracy = 0.75 (F1–score = 0.751),
• PPN : Accuracy = 0.746 (F1–score = 0.737).
As a future development, we envisage the use of
larger labeled datasets, that is by using additional
recordings with the corresponding aural survey, and
the application of deep neural network to develop
more efficient classifications. The definition of SRI
and the threshold defining the classes of sound qual-
ity will be subject of a further investigation.
References:
[1] Krause, B. The Loss of Natural Soundscapes.
Earth Island Journal, Spring 2002. (www.earth-
island.org/journal/index.php/magazine/archive)
[2] Pijanowski, B. C., Farina, A., Gage, S.
H., Dumyahn, S. L., Krause, B. L. What
is soundscape ecology? An introduction
and overview of an emerging new science.
Landscape Ecology 26, 1213–1232 (2011).
(https://doi.org/10.1007/s10980-011-9600-8).
WSEAS TRANSACTIONS on ENVIRONMENT and DEVELOPMENT
DOI: 10.37394/232015.2023.19.85
Roberto Benocci, Andrea Potenza,
Giovanni Zambon, Andrea Afify,
H. Eduardo Roman
E-ISSN: 2224-3496
899
[3] Pavan, G. Fundamentals of Soundscape
Conservation. In: Farina A., Gage S.
H. (Eds.). Ecoacoustics: The Ecolog-
ical Role of Sounds, 235–258 (2017).
(https://doi.org/10.1002/9781119230724.ch14).
[4] Sethi, S. S., Jones, N. S., Fulcher, B.
D., Picinali, L., Clink, D. J., Klinck, H.,
Orme, C. D. L., Wrege, P. H., Ewers, R.
M. Characterizing soundscapes across diverse
ecosystems using a universal acoustic fea-
ture set. Proceedings of the National Academy
of Sciences 117(29), 17049–17055 (2020).
(https://doi.org/10.1073/pnas.2004702117).
[5] Lellouch, L., Pavoine, S., Jiguet, F., Glotin,
H., Sueur, J. Monitoring temporal change of
bird communities with dissimilarity acoustic in-
dices. Methods in Ecology and Evolution 5(6),
495–505 (2014). (https://doi.org/10.1111/2041-
210X.12178).
[6] Kasten, E. P., Gage, S. H., Fox, J., Joo,
W. The remote environmental assess-
ment laboratory’s acoustic library: An
archive for studying soundscape ecology.
Ecological Informatics 12, 50–67 (2012).
(https://doi.org/10.1016/j.ecoinf.2012.08.001).
[7] Pérez-Granados, C., Traba, J. Estimating
bird density using passive acoustic monitor-
ing: A review of methods and suggestions
for further research. Ibis 163, 1–19 (2021).
(https://doi.org/10.1111/ibi.12944).
[8] Shonfield, J., Bayne, E. M. Autonomous
recording units in avian ecological research:
Current use and future applications. Avian
Conservation and Ecology 12(1), 14 (2017).
(https://doi.org/10.5751/ace-00974-120114).
[9] Benocci, R., Roman, H. E., Bisceglie, A., An-
gelini, F., Brambilla, G., Zambon, G. Eco-
acoustic assessment of an urban park by statisti-
cal analysis. Sustainability 13(14), 7857 (2021).
(https://doi.org/10.3390/su13147857).
[10] Benocci, R., Roman, H. E., Bisceglie, A., et al.
Auto-correlations and long time memory of envi-
ronment sound: The case of an Urban Park in the
city of Milan (Italy). Ecological Indicators 134,
108492 (2022). (https://doi.org/10.1016/j.ecol-
ind.2021.108492).
[11] Cavallari, G. B., Ribeiro, L. S., Ponti, M. A.
Unsupervised representation learning using con-
volutional and stacked auto-encoders: A do-
main and cross-domain feature space analy-
sis. In: 31st SIBGRAPI Conference on Graph-
ics, Patterns and Images (SIBGRAPI). IEEE,
440–446 (2018). (https://doi.org/10.1109/SIB-
GRAPI.2018.00063).
[12] Ponti, M. A., Ribeiro, L. S. F., Nazare, T. S.,
Bui, T., Collomosse, J. Everything you wanted
to know about deep learning for computer vi-
sion but were afraid to ask. In: 30th SIBGRAPI
Conference on Graphics, Patterns and Images
Tutorials (SIBGRAPI-T). IEEE, 17–41 (2017).
(https://doi.org/10.1109/SIBGRAPI-T.2017.12).
[13] Nunes C., Solteiro Pires E. J., Reis A. Ma-
chine Learning and Deep Learning applied
to End-of-Line Systems: A review. WSEAS
Transactions on Systems 21, 147–156 (2022).
(https://doi.org/10.37394/23202.2022.21.16)
[14] Christin, S., Hervet, É., Lecomte, N. Applica-
tions for deep learning in ecology. Methods in
Ecology and Evolution 10(10), 1632–1644
(2019). (https://doi.org/10.1111/2041-
210X.13256).
[15] Fairbrass, A. J., Firman, M., Williams, C.,
Brostow, G. J., Titheridge, H., Jones, K. E.
CityNet–Deep learning tools for urban ecoacous-
tic assessment. Methods in Ecology and Evo-
lution 10(10), 1632–1644 (2019). Methods in
Ecology and Evolution 10(2), 186–197 (2019).
(https://doi.org/10.1111/2041-210X.13114).
[16] Lin, T. H., Tsao, Y. Source separation in ecoa-
coustics: A roadmap towards versatile sound-
scape information retrieval. Remote Sensing in
Ecology and Conservation 6(3), 236–247 (2020).
(https://doi.org/10.1002/rse2.141).
[17] Navarro, J. M., Pita, A. Machine Learn-
ing Prediction of the Long-Term Environmen-
tal Acoustic Pattern of a City Location Us-
ing Short-Term Sound Pressure Level Measure-
ments. Applied Sciences 13(3), 1613 (2023).
(https://doi.org/10.3390/app13031613).
[18] Orga F., Socoró J. C., Alías F., Alsina-
Pagès R. M., Zambon G., Benocci R., Bis-
ceglie A. Anomalous noise events considerations
for the computation of road traffic noise lev-
els: The DYNAMAP’s Milan case study (2017)
24th International Congress on Sound and Vi-
bration, ICSV 2017. (pdf at: http://hdl.han-
dle.net/2072/376268).
[19] Piczak, K. J. Environmental sound classification
with convolutional neural networks. In: IEEE
25th international workshop on machine learning
for signal processing (MLSP). IEEE, 1–6 (2015).
(https://doi.org/10.1109/MLSP.2015.7324337).
WSEAS TRANSACTIONS on ENVIRONMENT and DEVELOPMENT
DOI: 10.37394/232015.2023.19.85
Roberto Benocci, Andrea Potenza,
Giovanni Zambon, Andrea Afify,
H. Eduardo Roman
E-ISSN: 2224-3496
900
[20] Salamon, J., Bello, J. P. Deep convolutional
neural networks and data augmentation for
environmental sound classification. IEEE Sig-
nal processing letters 24(3), 279–283 (2017).
(https://doi.org/10.1109/LSP.2017.2657381).
[21] Ward, J. H. Hierarchical grouping to optimize
an objective function. Journal of the American
Statistical Association 58(301), 236–244 (1963).
(https://doi.org/10.1080/01621459.1963.10500845).
[22] Ruff, Z. J., Lesmeister, D. B., Appel, C.
L., Sullivan, C. M. Workflow and convolu-
tional neural network for automated identifica-
tion of animal sounds. Ecological Indicators 124,
107419 (2021). (https://doi.org/10.1016/j.ecol-
ind.2021.107419).
[23] Vidaña-Vila, E., Navarro, J., Stowell, D.,
Alsina-Pagès, R. M. Multilabel Acous-
tic Event Classification Using Real-World
Urban Data and Physical Redundancy
of Sensors. Sensors 21(22), 7470 (2021).
(https://doi.org/10.3390/s21227470).
[24] Mullet, T. C., Gage, S. H., Morton, J. M.,
Huettmann, F. Temporal and spatial variation
of a winter soundscape in south-central Alaska.
Landscape Ecology 31(5), 1117–1137 (2016).
(https://doi.org/10.1007/s10980-015-0323-0).
[25] Quinn, C. A., Burns, P., Gill, G., Bali-
gar, S., Snyder, R. L., Salas, L., Goetz,
S. J., Clark, M. L. Soundscape classification
with convolutional neural networks reveals tem-
poral and geographic patterns in ecoacoustic
data. Ecological Indicators 138, 108831 (2022).
(https://doi.org/10.1016/j.ecolind.2022.108831).
[26] Giannakopoulos, T., Siantikos, G., Perantonis,
S., Votsi, N. E. and Pantis, J. Automatic sound-
scape quality estimation using audio analysis. In:
Proceedings of the 8th ACM International Con-
ference on Pervasive Technologies Related to As-
sistive Environments, Corfu, Greece (2015), pp.
1–9. (https://doi.org/10.1145/2769493.2769501).
[27] Tsalera, E., Papadakis, A., Samarakou,
M. Monitoring, profiling and classifica-
tion of urban environmental noise using
sound characteristics and the KNN algo-
rithm. Energy Reports 6, 223–230 (2020).
(https://doi.org/10.1016/j.egyr.2020.08.045).
[28] Pita, A., Rodriguez, F. J., Navarro, J.
M. Cluster analysis of urban acoustic envi-
ronments on Barcelona sensor network data.
International Journal of Environmental Re-
search and Public Health 18(16), 8271 (2021).
(https://doi.org/10.3390/ijerph18168271).
[29] Benocci, R., Afify, A., Potenza, A., Ro-
man, H.E., Zambon, G. Toward the defini-
tion of a soundscape ranking index (SRI)
in an urban park by machine learning
techniques. Sensors 23(10), 4797 (2023).
(https://doi.org/10.3390/s23104797).
[30] Benocci, R., Afify, A., Potenza, A., Ro-
man, H.E., Zambon, G. Self-Consistent
Soundscape Ranking Index: The Case of
an Urban Park. Sensors 2023, 23, 3401.
(https://doi.org/10.3390/s23073401).
[31] Benocci, R., Potenza, A., Bisceglie, A.,
Roman, H.E., Zambon, G. Mapping of the
Acoustic Environment at an Urban Park in
the City Area of Milan, Italy, Using Very
Low-Cost Sensors. Sensors 2022, 22, 3528.
(https://doi.org/10.3390/s22093528).
[32] Python. Available at https://www.python.org/
(accessed on 05/05/2023).
[33] Tibshirani, R., Friedman, J. The Elements of
Statistical Learning: Data Mining, Inference,
and Prediction. (Trevor Hastie, Second Edition,
2009).
[34] Zell, A. Simulation Neuronaler Netze (Addison-
Wesley, Bonn 1994). (https://doc1.biblio-
thek.li/aal/FLMF007250.pdf).
[35] Sueur, J., Pavoine, S., Hamerlynck, O., Du-
vail, S. Rapid acoustic survey for biodiver-
sity appraisal. PLoS ONE 3(12), e4065 (2008).
(https://doi.org/10.1371/journal.pone.0004065).
[36] Pieretti, N., Farina, A., Morri, D. A new
methodology to infer the singing activity of
an avian community: The Acoustic Complex-
ity Index (ACI). Ecological Indicators 11(3),
868–873 (2011). (https://doi.org/10.1016/j.ecol-
ind.2010.11.005).
[37] Grey, J. M., Gordon, J. W. Perceptual ef-
fects of spectral modifications on musical
timbres. The Journal of the Acoustical So-
ciety of America 63(5), 1493–1500 (1978).
(https://doi.org/10.1121/1.381843).
[38] Boelman, N. T., Asner, G. P., Hart, P. J., Mar-
tin, R. E. Multitrophic invasion resistance in
hawaii: Bioacoustics, field surveys, and airborne
remote sensing. Ecological Applications 17(8),
2137–2144 (2007). (https://doi.org/10.1890/07-
0004.1).
[39] Yang, W., Kang, J. Soundscape and
sound preferences in urban squares:
WSEAS TRANSACTIONS on ENVIRONMENT and DEVELOPMENT
DOI: 10.37394/232015.2023.19.85
Roberto Benocci, Andrea Potenza,
Giovanni Zambon, Andrea Afify,
H. Eduardo Roman
E-ISSN: 2224-3496
901
A case study in Sheffield. Journal of
Urban Design 10(1), 61–80 (2005).
(https://doi.org/10.1080/13574800500062395).
[40] R Core Team. R: A Language and Environ-
ment for Statistical Computing. R Foundation for
Statistical Computing: Vienna, Austria, 2018.
(Available online: https://www.R-project.org/.
(accessed on 05/05/2023).
[41] Seewave: Sound Analysis and Synthesis. Avail-
able online: https://cran.r-project.org/web/pack-
ages/seewave/index.html. (accessed on
05/05/2023).
[42] Soundecology: Soundscape Ecology. Available
online: https://cran.r-project.org/web/pack-
ages/soundecology/index.html. (accessed on
05/05/2023).
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
R.B. conceptualized the study, carried out the M.C.
simulation and wrote the original draft preparation;
A.P. organized the data sets and produced map rep-
resentations; G.Z. supervised the research; A.A. im-
plemented the M.L. software; H.E.R. reviewed and
edited the manuscript.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself:
No funding was received for conducting this study.
Conflicts of Interest
The authors have no conflicts of interest to
declare that are relevant to the content of this
article.
Creative Commons Attribution License 4.0
(Attribution 4.0 International , CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on ENVIRONMENT and DEVELOPMENT
DOI: 10.37394/232015.2023.19.85
Roberto Benocci, Andrea Potenza,
Giovanni Zambon, Andrea Afify,
H. Eduardo Roman
E-ISSN: 2224-3496
902