Application of the SVM algorithm for the development of a model classification of the visual and sound landscape

Samantha Di Loreto 1

Università Politecnica delle Marche Via Brecce Bianche 12, Ancona, Italy

Fabio Serpilli 2

Università Politecnica delle Marche Via Brecce Bianche 12, Ancona, Italy

Valter Lori 3

Università Politecnica delle Marche Via Brecce Bianche 12, Ancona, Italy

ABSTRACT To ensure adequate management of the soundscape in urban environments, urban planning authorities need a range of tools that enable them to perform this condition. Analyzing and classifying a soundscape is necessary to adapt it to the expectations of the people who inhabit it. The term “soundscape” is associated with three di ff erent research areas: ecology / anthropology, music / sound design and architecture / urbanism. In particular, in this paper, the third research area will be investigated, finding a correlation model between auditory and visual sensations of the urban landscape of the port of Ancona. The classification model that is used is the Support Vector Machines (SVM) which is proposed as a tool for a global assessment of the urban sound landscape. In this case study the algorithm is intended for the automatic classification of the sound landscape of the port of Ancona to understand how the sound perception a ff ects the visual one.

1. INTRODUCTION

The landscape represents the whole of a view that contains sound, olfactory and visual elements which simultaneously come into contact with the observer. It is known that various limits are identified in the definition of landscape, since the theme of landscape is tackled by considering it a strictly objective whole, without considering the subjectivity of those in front of it. This limit was overcome in M. Schafer’s World Soundscape Project [1], where the concept of landscape is approached on di ff erent levels.

1 s.diloreto@pm.univpm.it

2 f.serpilli@univpm.it

3 v.lori@univpm.it

a slaty. inter.noise 21-24 AUGUST SCOTTISH EVENT CAMPUS O ¥, ? GLASGOW

Soundscapes defined by the International Organization for Standardization (ISO 12913 series) [2,3], have been idealized as the acoustic environments perceived by one or more people in the considered urban context. Many theories, methods and applications are available in the field of landscape perception research. Kang et all. [4] have demonstrated that human evaluation of subjective loudness and acoustic comfort depends on a large-scale subjective survey. A survey has been undertaken by the same authors on underground shopping streets in Harbin, China, to determine how individual sound sources influence subjective loudness and acoustic comfort evaluation [5]. Oberman et all [6] has discusses the soundscape assessment approaches to soundscape interventions with musical features introduced to in public spaces as permanent sound art. In the virtual soundwalk is it possible to combine the benefits of the on-site and laboratory settings. With this approach it is possible to combine more cognitive experiences. The development of models for environmental problems is becoming increasingly relevant to environmental engineers and scientists. The application of mathematical models for environmental modeling is necessary to solve complex problems. Torija et all. [7] presented a classification model that is proposed to be used as a tool for a comprehensive urban soundscape evaluation. The application of machine learning methods for environmental modeling is extensive and less discusses respect the polynomial regression models [8]. In [9] machine learning methods have shown their superior performances; the work compared the performance of 23 methods, including RF, support vector machine (SVM), ordinary kriging (OK), inverse distance squared (IDS), and their combinations (i.e., RFOK, RFIDS, SVMOK and SVMIDS), to evaluate the landscape of the Australian coast. This paper explores the sound and temporal characteristics and compares them with the physicality of spaces. The final goal is to highlight in particular how the concept of soundscape is closely related to the visual impression. The aim of the work was to highlight the a ff ecting role of sound with the visual setting by investigating the subjective sensations of the human being with a listening test created specifically for the case study. In particular, for the aforementioned purpose, machine learning modeling techniques were exploited. The paper is organized as follows: Section 2 presents the case study, measurements and subjective investigation. In Section 3 a brief discussion of machine learning method has presented and the results of simulation has discussed. Conclusion are reported and discussed in Section 4. The methodology developed in this research is applicable to other open air environments after a specific analysis of cluster and design of the place.

2. MATERIALS AND METHODS

The study was carried out in the seaport area of the city of Ancona. The area is frequented daily by a large number of people, it is geographically delimited and characterized by the presence of a great variety of sound sources. The whole area was subjected to an important investigation of the characteristics of the sound landscape. The present work concerns one of the clusters into which the area has been divided: the Clementino Pier. The aim of the study was to evaluate how much human visual perception a ff ects the sound sensation in the area. The fig.[1] shows the aerial view of the port area with indication of the measurement point P. The sources, to facilitate their evaluation, have been classified, during the research, with a common framework according to the standard [3]. The taxonomy is built on three levels: types of places, types of sound sources and sound sources. The categories of place considered are internal or external. The external categories are divided in urban, rural and wild conditions. The seaport of Ancona has been classified as an urban area. In this study, the identification of the sound sources was fundamental to compare all the noise sources that coexist in the seaport.

Figure 1: Aerial view of the seaport area with indication of the measurement point P, zoom on Clementino Pier (seaport of Ancona, Italy)

2.1. Acoustic characterization Acoustic measurements related to a soundscape have to consider the way human beings perceive the acoustic environment. For this purpose, calibrated binaural measurements system (head and torso simulator) was chosen for the acoustic measurements of the seaport area. Binaural measurements were be made in accordance to the standard [3]. The experimental measurements allowed the evaluation of the acoustic parameters such as sound levels (SPL) and the main objective parameters of psychoacoustics [10]: Loudness and percentiles, Sharpness, Roughness, Fluctuation strength. The measurement of the sound levels and the acquisition of the audio tracks was carried out using Zoom Hn4 and binaural headphones worn by the operator (see tab.[1]). The choice fell on the Roland CS 10-EM binaural headphones (24-bit / 96kHz), the headphones in-ear monitoring combined binaural recording, improve the field recording.

Table 1: Acoustic and psychoacoustic parameters measured with binaural headphones.

Acoustics and Psychoacoustics parameters Right ear left ear

Leq [dBA] 51,2 53,8

Loudness [Phone] 62,6 67,5

Sharpness [Acum] 1,1 1,3

Fluctation strenght [vacil] 0,3 0,6

Roughness [Asper] 1,4 1,8

Loudness (free 5%) [sone] 2,3 2,0

Loudness (free 50%) [sone] 1,8 1,6

Loudness (free 90%) [sone] 1,6 1,5

2.2. Creation and management of the listening test When collecting human perception data, the investigator must in no way interfere with the participants’direct experience [11]. Such data collection must capture people’s general mood, appreciation, preferences and behavior to create an accurate representation of a specific place [12]. The final assessment must be holistic, covering all auditory sensations and all other varieties of context, such as visual stimuli and personal expectations. For this case study, in order to evaluate how the quality of the visual environment influences the perception of the quality of the soundscape, two ad hoc tests were created. In the fig.[2] are reported the questionnaires examples that were created

on an online platform (Google form) and shared through social media. The test was anonymous. No sensitive informations were shared and the data were processed for academic purposes only.

Express a judgment on satisfaction with some characteristics of the place ‘The participant is asked to make a 1-5 rating ‘Traffic, pollution, noise Dirt / Street cleaning Degradation / state of maintenance of buildings @e000 Providing green areas, spaces for recreational and sports activities _, ©8080 | Difficulty of parking Security for access by bicycle or on foot Cultural and recreational offer @0000 How would you describe the sound environment in this area? The participant is asked to make a 1-5 rating Pleasantness @000O. Characteristic Relaxing Calm @e@000 Boring Chaotic @@000

Figure 2: Example of questionnaires proposed: Test Visual (left) and Test Sound (right)

The test participants were selected among students, professors of the Polytechnic University of Marche and industry professionals living in the province of Ancona. The age varied between 25 and 70 years. In the first section of each of the two tests, the subjects were asked to write down information of a cognitive type, such as the frequency, purpose and mode of use of the place. Fig.[3] shows the results of preliminary analysis.

WHEN OF THE DAY DO YOU VISIT THE PORT? NIGHT (1436) MORNING (23%) AFTERNOON (539%) LUNCH (1196)

Figure 3: Pie chart that indicates the results of cognitive investigation

This type of analysis has given us general informations on the users of the seaport. In particular, the image test has provided with information of an urban and landscape nature which are essential for the Urban’s assessment from the local authorities.

3. RESULTS AND DISCUSSION

For evaluating the performance of the models, in this section are reported the approaches implemented in this work. Before implementing SVM algorithms [13] to develop a landscape- classification model, a statistical analysis was made to ensure the acoustical correspondence between measurements and the response of the listeners. The regression analysis [8] is an essential step in

INDICATE YOUR WORK PROFESSION UNIVERSITY STUDENTS (1296) ARCHITECTS and / or DESIGNERS (209%) ENGINEER (619%)

developing the classification model because is necessary the existence of a correspondence between the acoustical and perceptual criteria. Statistical analysis was used to understand the optimal relationship between the soundscape and the visual composition of the cluster under examination. This was done after normalizing the judgments collected in the tests by calculating the average response of the sample and the dispersion analysis by calculating the STD.

3.1. Support Vector Machines for developing a soundscape classification model The Support Vector Machines (SVM), algorithm used to solve multiclass classification problems, it is most e ff ective in binary classification problems [14–16] (see fig.[4]). SVM key concepts: • Hyperplane or linear decision limit : for a classification activity with only two spatial dimensions x1 and x2 (or x-y), a hyperplane is depicted as a line that separates and classifies a set of data. In three dimensions x1, x2 and x3 (or x-y-z), a hyperplane is represented by a plane. • Support Vector : they are the data points closest to the hyperplane. These points depend on the data set being analyzed and if they are removed or modified they alter the position of the dividing hyperplane. For this reason, they can be considered the critical elements of a dataset. • Margin : is defined as the distance between the support vectors of two di ff erent classes closest to the hyperplane. At the middle of this distance, the hyperplane is traced, or a straight line if you are working in two dimensions. The SVM, in general, is based on the idea of finding a hyperplane that best divides a data set into two classes. In fact, the classification algorithm looks for a "linearly separable hyperplane" or a "decision limit" that separates the values of one class from the other. If there is more than one, the algorithm look for the one that has the highest margin with the support vectors, to improve the accuracy of the model.

Figure 4: Support Vector Machines for Binary Classification

3.2. Mathematical model First of all, consider the training set of n points { x i , y i } i = 1 , . . . , n where y i ∈{− 1 , 1 } is the class of the point x i . Here, we see the two classes as y i = 1 and y i = − 1. We want to find a classifier (a hyperplane) which can map x i ’s into higher dimensional space so that the two di ff erent classes of points can be divided. Also we want the hyperplane to have the maximum-margin, which can maximize the distance of the hyperplane and nearest points from both groups. In linear cases, the hyperplane can be written as [17]:

x i w + b = 0

We want to find two parallel hyperplanes that can also separate the data and we want their distance to be as large as possible. Here’s a way to describe them:

x i w + b = + 1

and x i w + b = − 1

It’s not di ffi cult to prove that the distance between the two hyperplanes is 2 || w || :

d + + d − = | 1 − b |

∥ w ∥ + | − 1 − b |

∥ w ∥ 2 ∥ w ∥ (1)

which means we want to minimize ∥ w ∥ since it is always positive. Besides, we want for every i ∈ (1 , n ), x i and y i follows the constraints:

x i w + b ≥ + 1 , y i = + 1 (2) x i w + b ≤− 1 , y i = − 1 (3) ≡ (4) y i ( x i w + b ) − 1 ≥ 0 , ∀ i (5)

However, in most cases it is impossible for all the points in the training data to follow the rules. To solve this problem, people add this constraint as a regularization part in the object function which we want to minimize. The "fitcsvm" function implemented in the Matlab numerical computing environment is able to process di ff erent SVM algorithms to solve classification problems.

3.3. Application of the model to the case study In order to study the soundscape of the Clementino Pier in Ancona, 75 participants were selected. The participant filled out a questionnaire consisting of 20 questions to evaluate the visual environment and 20 questions to evaluate the sound environment, assigning each question an integer value between 1 and 5. The number of observations will therefore be equal to 75 (participants) which multiplied by 20 (questions) will produce 1500 (observations) for each environment considered. The training set (train_data) will be composed of 1500 samples for images and 1500 samples for sounds. The training data is combined with the class labels (negative or positive) yj = ± 1 (train_class). Applying the fitcsvm (SMO) algorithm "Model = fitcsvm (train_data ’, train_class’)" we obtain the value of each parameter used for the implementation algorithm. Then, a random partition of the data was performed for a k-Fold cross-validation. The partition randomly divides the training data into K = 10 subsets. Subsequently, a part (1 / K) is selected in turn to use it as a validation set, while the remaining parts (K - 1 / K) resume composing the training set. Finally, after calculating the hyperplane equation of our specific system, the latter was plotted and the result is represented by the following fig.[5] processed by Matlab. From this analysis, some important aspects should be highlighted. This type of classifier has the advantage of being applicable to high-dimensional vector spaces and, as in our case, it is applicable to the analysis of feelings and emotions. The SVM is not a particularly intuitive and easy to interpret tool. The obtained result certifies a greater presence of the vectors of sounds in the vicinity of the hyperplane as opposed to the vectors of the images which, are all marked by the supporting vectors. The visual experience of the Clementino Pier in Ancona was judged by the 75 participants to be slightly more positive than the auditory experience.

4. CONCLUSION

This study focuses on the e ff ects of landscape factors on soundscape perception. In this discussion, the correlation between auditory and visual sensations of the "sound landscape" of the

Figure 5: Support Vector Machines for landscape classification

seaport of Ancona (Clementino pier) was studied with the aid of the "Support Vector Machines" (SVM) classifier. This work is based on the hypothesis that from a set of categories of subjective evaluations which were classified and categorized using acoustical as well as perceptual descriptors, the algorithm model could be evaluate the soundscape of the location according to acoustical and visual perceptual criteria. On the basis of this hypothesis, the correspondence that exists between the acoustics and the vision of the urban place considered has been confirmed trough a mathematical evaluation. Based on this approach, the authors proposed a SVM classification model as a good tool for evaluating soundscapes on the basis of acoustical measurements. After this evaluation, important information on the qualification of the urban environment could be acquired to give indications to the authorities on the management of urban landscapes. At the territorial level, the study will support regional planning for the integration of the value of natural capital into governance. Since the assessment of environmental noise is crucial for the construction of new construction settlements, it is also necessary to quantify the impact that this construction will have on the community. This classification is made on the basis of an unsupervised categorization so it can be implemented in an automatic procedure with the selected input. The methodology developed in this research is applicable to other open air environments after a specific analysis of cluster and design of the place.

REFERENCES

[1] R. Murray Schafer. The new soundscape . 1969. [2] International standard ISO. Acoustics — soundscape — part 1: Definition and conceptual framework. Standard ISO 12913-1:2014, International Organization for Standardization, 2014. [3] International standard ISO. Acoustics — soundscape — part 2: Data collection and reporting requirements. Standard ISO / TS 12913-2:2018, International Organization for Standardization, 2018. [4] Tin OBERMAN Mercede ERFANIAN Magdalena KACHLICKA Matteo LIONELLO Andrew MITCHELL Jian KANG, Francesco ALETTA. Towards soundscape indices.

Landscape Classification Regions ®@ @08@0808

PROCEEDINGS of the 23rd International Congress on Acoustics , 2019. [5] Jiang Liu, Jian Kang, Tao Luo, and Holger Behm. Landscape e ff ects on soundscape experience in city parks. Science of The Total Environment , 454-455:474–481, 2013. [6] Tin Oberman, Kristian JambroÅ¡iÄ [U+0087] , Marko Horvat, and Bojana BojaniÄ [U+0087] Obad Å Ä [U+0087] itaroci. Using virtual soundwalk approach for assessing sound art soundscape interventions in public spaces. Applied Sciences , 10(6), 2020. [7] Antonio J. Torija, Diego P. Ruiz, and à [U+0081] ngel F. Ramos-Ridao. A tool for urban soundscape evaluation applying support vector machines for developing a soundscape classification model. Science of The Total Environment , 482-483:440–451, 2014. [8] J Chee. Pearson’s product moment correlation: Sample analysis. Technical report, University of Hawaii at M¯anoa School of Nursing, 2015. [9] Jin Li, Andrew D. Heap, Anna Potter, and James J. Daniell. Application of machine learning methods to spatial interpolation of environmental variables. Environmental Modelling Software , 26(12):1647–1659, 2011. [10] International standard ISO. Acoustics – methods for calculating loudness — part 1: Zwicker method. Standard ISO 532-1:2017, International Organization for Standardization, 2017. [11] Deborah A. Hall, Amy Irwin, Mark Edmondson-Jones, Scott Phillips, and John E.W. Poxon. An exploratory evaluation of perceptual, psychoacoustic and acoustical properties of urban soundscapes. Applied Acoustics , 74(2):248–254, 2013. Applied Soundscapes: Recent Advances in Soundscape Research. [12] Rebecca Cain, Paul Jennings, and John Poxon. The development and application of the emotional dimensions of a soundscape. Applied Acoustics , 74(2):232–239, 2013. Applied Soundscapes: Recent Advances in Soundscape Research. [13] S.K. Shevade, S.S. Keerthi, C. Bhattacharyya, and K.R.K. Murthy. Improvements to the smo algorithm for svm regression. IEEE Transactions on Neural Networks , 11(5):1188–1193, 2000. [14] Rong Xiao, Jicheng Wang, and Fayan Zhang. An approach to incremental svm learning algorithm. In Proceedings 12th IEEE Internationals Conference on Tools with Artificial Intelligence. ICTAI 2000 , pages 268–273, 2000. [15] Zichen Zhang, Shifei Ding, and Yuting Sun. A support vector regression model hybridized with chaotic krill herd algorithm and empirical mode decomposition for regression task. Neurocomputing , 410:185–201, 2020. [16] Giuseppe Ciaburro, Gino Iannace, Virginia Puyana-Romero, and Amelia Trematerra. Machine learning-based tools for wind turbine acoustic monitoring. Applied Sciences , 11(14), 2021. [17] Nello Cristianini and John Shawe-Taylor. An Introduction to Support Vector Machines and Other Kernel-based Learning Methods . Cambridge University Press, 2000.