Assessment of drone noise impact on urban soundscapes Rory Kerr Nicholls 1

Dr. Antonio J Torija Martinez 2 Nathan Green 3 Dr. Carlos Ramos Romero 4 Acoustics Research Centre University of Salford, The Crescent, Salford, M5 4WT

ABSTRACT

Drones have the potential to be implemented in the infrastructure of metropolitan areas to provide services to the public such as delivery, maintenance, and even blue light services. The scope for drone capability is large, but this comes with the risk of introducing a dominant, unpleasant noise source above urban areas, with potentially adverse health effects. This paper describes research which in- vestigates whether, and to what extent, different urban soundscapes mask noise from drone opera- tions, and using statistical analysis methods such as principal component analysis interrogates which frequency ranges of drone noise should be appropriately masked to reduce perceived annoyance. A subjective experiment was carried out, where participants gave response values to a comprehensive set of drone sounds embedded into differing urban soundscapes. Response values included perceived annoyance, perceived loudness, drone noise dominance and soundscape pleasantness. Critical-band rate specific Sound Quality Metrics (SQMs) were then calculated for the soundscapes with and with- out each drone sound present, and the metric value differences were used in principle component analysis, with the results suggesting which specific sound quality metrics contribute significantly to the response values.

1. INTRODUCTION

The use of Unmanned Aerial Vehicles (UAVs), also known as drones, has the potential to signif- icantly change how cities operate. Whilst some of the more publicised uses of drones include parcel delivery and photography, several other uses, such as emergency response (first aid), firefighting, road traffic control and monitoring air pollution could greatly benefit society as a whole [1]. As more uses for drones become apparent, their implementation could become more widespread than currently anticipated and could lead to a large proportion of urban areas being exposed to drone noise to some degree.

1 R.K.Nicholls@edu.salford.ac.uk

2 A.J.TorijaMartinez@salford.ac.uk

3 N.Green7@edu.salford.ac.uk

4 C.A.RamosRomero@salford.ac. u k

worm 2022

The sound signature of drones is quite distinctive when compared against traditional, ubiquitous noise sources, such as road, rail or aircraft noise, as it generally contains a greater amount of high frequency content (2kHz and above) [2]. The way multi-rotor vehicles balance and propel themselves, i.e., by changing the thrust produced by each of the rotors independently, is also significant as it results in an almost constantly fluctuating sound with different tonal and broadband characteristics [3]. As such, these specific characteristics can make drones perceptible against the rest of the ambient sound climate, even at relatively low overall noise levels.

This paper is part of an on-going research project aimed to understand the human response to drone noise, as a function of acoustic (e.g., prominence over existing ambient sound) and non-acoustic con- text (e.g., purpose of drone operation, attitude towards the noise source, visual aspects, etc.). Specif- ically, the goal of this paper is to investigate how Sound Quality Metrics (SQMs) can aid the assess- ment of drone noise, and how their values can be used to link noise perception with specific drone noise characteristics.

Critical-band rate specific SQMs were calculated for a series of 120 soundscapes with and without a drone operation present. These soundscapes include combinations of 20 drone sounds embedded in 6 different sound environments (as shown in Tables 1 and 2). The value differences for each critical-band rate specific SQM (i.e., soundscape with drone present – soundscape without drone) were used as independent variables in Linear Regression with Principal Component Analysis (PCA). The responses of participants of a subjective experiment were used as dependent variables in the Linear Regression with PCA. The specific subjective variables (or dependent variables) are perceived annoyance, perceived loudness, drone noise dominance and soundscape pleasantness.

2. METHODOLOGY

The methodology described in this paper is an extension of that used in [4], which investigated direct relationships between drone noise characteristics and subjective human response. This meth- odology incorporated a similar subjective testing style, where participants assessed drone stimuli in the context of differing urban soundscapes. The drone stimuli to be included in this subjective exper- iment were chosen based on their varying values of SQMs and operational characteristics.

2.1. Soundscape stimuli

A total of 120 soundscape stimuli were used in the subjective experiment. These stimuli were created using 20 differing drone recordings, interpolated into 6 soundscape recordings. The drone recordings included operations such as flyovers, take-offs, hovering, landing, and mid-flight manoeu- vres at distances from 1.2 m to 60 m. The soundscape environments included a variety of typical urban scenes, such as city squares, canal paths, parks and pedestrianised streets. The soundscapes were 10 s long. Table 1 gives more detail on the drone stimuli used in the subjective experiment, and Table 2 gives a description of the environments included, which were recorded in locations around Manchester.

worm 2022

Table 1: Drone stimuli information

Drone Sound No. Drone Model Drone Weight

(kg) Drone Operation Drone Distance

(m) 1 DJI Inspire 2.85 Flyover 15 2 DJI Inspire 2.85 Flyover 7.5 3 DJI Inspire 2.85 Landing 7.5 4 DJI Inspire 2.85 Take-off 2 5 DJI Matrice 600 9.1 Hover 40 6 DJI Matrice 600 9.1 Flyover 40 7 DJI Mavic 0.743 Flyover 15 8 DJI Mavic 0.743 Flyover 30 9 DJI Mavic 0.743 Flyover 60 10 DJI Mavic 0.743 Manoeuvring 7.5 11 DJI Mavic 0.743 Take-off 7.5 12 DJI Matrice 200 4 Flyover 46 13 DJI Matrice 200 4 Flyover 46 14 DJI Matrice 200 4 Take-off 30 15 DJI Matrice 200 4 Landing 30 16 DJI Matrice 200 4 Hover 1.2 17 Gryphon GD28X 11.8 Take-off 30 18 Gryphon GD28X 11.8 Landing 30 19 Gryphon GD28X 11.8 Manoeuvring 30 20 Gryphon GD28X 11.8 Hover 1.2

Table 2: Soundscape description

Environment

Sound No. Environment Description

1 Canal scene with train noise 2 Harehills Ln, road traffic noise 3 Millennium Square, pedestrian noise and adverts 4 Park with distant road traffic noise 5 Peel Park, early morning 6 Market St, very busy with pedestrians

worm 2022

rr ee {Tea 1102 «2 gp) ono eunssong punog, 0 Frequency [Hz]

Figure 1: Frequency spectra (left) and spectrogram (right) for the DJI Matrice 200 hovering at 1.2 m from microphone.

Figure 2: (Top) Spectrogram of Environment 3 without (left) and with (right) DJI Matrice 600 hov- ering at 40 m from the microphone; (Bottom) Spectrogram of Environment 4 without (left) and with (right) DJI Matrice 600 hovering at 40 m from the microphone.

worm 2022

The frequency spectra and spectrogram of the DJI Matrice 200 hovering at 1.2 m from the micro- phone (Drone sound no. 16) can be seen in Figure 1, which shows the fundamental frequencies of both sets of rotors at about 150 Hz, and also illustrates the magnification of temporal fluctuations at their higher harmonics. Figure 2 shows the spectrograms of environment 3 (top) and 4 (bottom), with and without the DJI Matrice 600 hovering at 40 m from the microphone, and illustrates how the ambient noise present can mask the sound produced by the drone.

2.2. Online subjective experiment

Due to the restrictions put in place during the COVID-19 pandemic, it was impossible to carry out a laboratory based subjective experiment. Therefore, an online experiment methodology was adopted to increase the capacity of participants available. Upon acceptance of participation and giving con- sent, the participants were given a personal identification number, which they used to access the experiment via the Web Audio Evaluation Toolkit (WAET) [5]. The WAET meant that a user inter- face could be built for the participants to give their response values after listening to each soundscape stimulus. Prior to accessing the experiment, the participants went through a calibration stage, to at- tempt to mitigate problems that arise when assessing subjective response via an online methodology. This included a level calibration stage, where the participants were instructed to adjust their playback system so that the “quietest” stimulus was just audible, and the “loudest” stimulus was at a comfort- able level, so not to introduce a risk of damage to the participants’ hearing.

The response values available to the participants were perceived loudness, annoyance and drone dominance, which could be rated using a slider for each response metric. The loudness slider had extremes of “very quiet” and “very loud”, the annoyance slider had extremes of “not very annoying” and “very annoying”, and the drone dominance slider had extremes of “not very dominant” to “very dominant”. The participant could listen to each soundscape as many times as they required to make their decision on each response metric, and then continue to the next soundscape. The order of the soundscapes was randomised for each participant. The response values were saved anonymously on a numerical scale from 0 to 1, corresponding to the position of the slider placement for each response metric. The experiment took around 40 minutes in total, and participants had the option to take a 5- minute break halfway through the stimuli, to mitigate the effects of fatigue.

3. ANALYSIS

The analysis aims to understand the effects that urban environments could have on manipulating the perceived response to drone noise. To assess this, critical-band rate specific SQMs were calculated using the HEAD Acoustics ArtemiS Suite for the 6 soundscape environments where a drone stimulus is not present, and then for the 120 soundscape stimuli which include drone noise embedded in the 6 soundscapes. From this, the differences in the specific SQM values which are introduced when the drone noises are present can be calculated. The specific SQMs used in the analysis included specific Loudness (DIN45631/A1 model (1 – 24 Bark)), specific Tonality (Aures/Terhardt tonality model (1 – 24 Bark)), specific Roughness (Sottek’s Hearing Model (1 – 24 Bark)), specific Fluctuation Strength (Sottek’s Hearing Model (1 – 20 Bark)), and specific Impulsiveness (Sottek’s Hearing Model (1 – 23 Bark)). The differences in the specific SQMs at each critical band-rate ( ∆𝑆𝑄𝑀 𝑠,𝑒,𝑏 )

worm 2022

were then calculated by subtracting the SQM values of the soundscape including the drone noise from the corresponding specific SQMs of the environment used in that soundscape. This is demonstrated in Equation 1.

∆𝑆𝑄𝑀 𝑠,𝑒,𝑏 = 𝑆𝑄𝑀 𝑠,𝑏 −𝑆𝑄𝑀 𝑒,𝑏 ሺ1ሻ

Where 𝑆𝑄𝑀 𝑠,𝑏 is one of the specific SQMs for a soundscape (i.e., environment + drone), 𝑠 , at critical-band rate, 𝑏 , and 𝑆𝑄𝑀 𝑒,𝑏 is the specific SQM for the corresponding environment, 𝑒 , used in that soundscape. From this, an individual dataset was built for differences of each individual spe- cific SQM.

A PCA was implemented using MATLAB to build principal components that reduce the dimen- sionality of the specific ΔSQM datasets. PCA is a technique used to reduce the dimensionality of data, which is especially useful in datasets which include many variables, or dimensions. PCA takes the covariance between each of the variables in a dataset and uses this matrix to find eigenvectors that correspond to principal components. The eigenvector with the largest eigenvalue for the dataset is the first principal component, and the eigenvector with the smallest eigenvalue is the last principal component. As you increase the number of eigenvectors and corresponding eigenvalues calculated for a dataset, variance in the dataset explained by the eigenvectors increases, meaning that much of the variance between variables could be explained with a smaller number of eigenvectors, or principal components, than the number of variables included in the dataset. Each variable used by a principal component has a coefficient, or score, which relates the variable to that component, and can be used as an indicator of the variable’s significance.

PCA was applied to each individual specific ΔSQM dataset, using the ΔSQM value at each critical- band rate as variables. The application of PCA allows the identification of the features (i.e., critical bands) that are most informative for each SQM. This approach also identifies the key features that conform a reduced number of principal components while retaining most of the information for each SQM. Assessing the percentage of the ΔSQM dataset variance explained by each principal compo- nent also gives insight into which critical-band rates are most controlling for the correlation between each SQM and the subjective variables investigated. The principal components were used in Linear Regression to reduce the dimensionality of the predictor datasets. The Linear Regression models each included 3 principal components of the ΔSQM datasets as predictor variables, which were regressed against each subjective response metric of perceived annoyance, perceived loudness and drone dom- inance.

4. RESULTS

After all iterations of Linear Regression with PCA were completed, key results were found be- tween various critical-band rate specific SQMs and the response variables assessed. The results pre- sented in this section show the definition of principal components for the critical-band values of each SQM, and the contribution of each principal component to the participants’ responses of perceived annoyance, perceived loudness and drone dominance.

Loudness was found to correlate strongly with all response metrics strongly, yielding adjusted R 2 values of 0.77, 0.74 and 0.74 with perceived annoyance, perceived loudness and drone dominance respectively. This is consistent with previous research [4,6], pointing out loudness as the main factor explaining changes in drone noise perception. Upon assessment of the principal components built

worm 2022

for specific Loudness, the first principal component, which explains 84.1% of the variance in the specific Loudness dataset, includes similar contributions from each critical-band rate, with an in- crease in the scores of the critical-band rates with centre frequencies between 2 kHz and 6.4 kHz (highest sensitivity for human hearing). It should be noted that the scores showed in the PCA results indicate the degree of association between each critical-band value and the specific principal compo- nent. The second principal component explained 6.5% of the variance, and had a particularly strong score for the critical-band rate with centre frequency of 200 Hz, as well as large, negative scores for critical-band rates with centre frequencies between 6.4 kHz and 9.4 kHz. The third principal compo- nent explained 3.4% of the variance, and yielded less affirming scores. The scores for the specific Loudness principal components are shown in Figure 3. These results suggest that the whole range of critical-band specific Loudness is highly associated to the principal component explaining the major- ity of the variance in the subjective responses perceived annoyance, perceived loudness and drone dominance.

An interesting finding is that no correlations have been significant between critical-band specific Tonality and the subjective variables investigated. This is consistent with previous research carried out by Gwak et al., [6]; Torija and Nicholls, [4]. The assumption of the authors of this paper is that either the subjective contribution to tonal noise is already capture by the Loudness metric (as all drone sounds are highly tonal) or the Tonality metrics used (Aures Tonality and Sottek’s Hearing Model Tonality) are not able to account for the perception of tonal drone noise.

Figure 3 : Scores of ΔSQM (Loudness) per critical-band rate for each principal component.

worm 2022

Figure 4 : Scores of ΔSQM (Roughness) per critical-band rate for each principal component.

‘Coraoutions of Specie Rovahness per Gicatband ates to Principal Components rs a

Roughness yielded adjusted R 2 values of 0.38, 0.32 and 0.41 with perceived annoyance, perceived loudness and drone dominance respectively. The first principal component shows a clear domination in the contribution of the scores corresponding with critical-band rates with centre frequencies be- tween 1.3 kHz and 5.3 kHz, with the principal component explaining 70.3% of the variance in the specific roughness dataset. The second principal component explains 13.1% of the variance, with scores of the critical-band rates with centre frequencies between 400 Hz and 910 Hz being most prominent. The third principal component explained 6.1% of the variance, with large scores for the critical-band rates with centre frequencies 100 Hz to 200 Hz. The scores for the specific Roughness principal components are shown in Figure 4. As reported by Alexander et al., [7], ambient weather conditions, such as wind speed, highly influence drone noise emission. Torija et al., [2] explained that under outdoor conditions, with the effect of wind gusts, the flight control system of drones vary the rotor rotational speeds to maintain vehicle stability, which leads to an unsteady acoustic signature. Cabell et al., [3] showed that these variations in rotor rotational speeds lead to a significant variability with time of frequency components. Specifically, relatively small variations with time in the Blade Passing Frequencies (of each rotor) are magnified at higher harmonics. Roughness is a metric able to describe these rapid fluctuations of the sound signature. Figure 4 seems to indicate that the variations with time in the mid to high frequency region (1.3 kHz and 5.3 kHz), which composed the first com- ponent of the Roughness metric, correlates with the subjective responses of perceived annoyance, perceived loudness and drone dominance. Also, as shown in Figure 4, the fluctuation with time in the mid to high frequency region seems to be more important for the subjective variables than the fluc- tuation in the low frequency region.

Fluctuation Strength yielded adjusted R 2 values of 0.42, 0.41 and 0.41 with perceived annoyance, loudness and drone dominance respectively. The first principal component, which explains 58.2% of the variance in the Fluctuation Strength dataset, is clearly dominated by the scores of the critical-band rates with centre frequencies of 200 Hz and 300 Hz. The second principal component explains 28%

worm 2022

of the variance and has two controlling critical-band rate regions, being those with centre frequencies of 300 Hz to 510 Hz, and 3.2 kHz to 5.3 kHz. The third principal component explains 4.9% of the variance, and yielded less affirming scores. The scores for the specific Fluctuation Strength principal components are shown in Figure 5. Torija et al., [8], suggested that Fluctuation Strength is a metric able to account for the beating effects (or low frequency amplitude modulation) due to interaction between rotors. For the specific case of the drones investigated in this paper, it seems that the ampli- tude modulation due to the interaction between rotor BPFs (at about 200 Hz to 400 Hz), and poten- tially their first harmonic (at about 400 Hz to 630 Hz), have an important contribution to the partici- pants’ responses of perceived annoyance, perceived loudness and drone dominance. This is described by the critical-band values of specific Fluctuation Strength composing principal component 1 and 2, as shown in Figure 5.

Figure 5 : Scores of ΔSQM (Fluctuation Strength) at each critical-band rate for each principal com- ponent.

Impulsiveness yielded an adjusted R 2 value of 0.35 with perceived drone dominance, but the cor- relations with perceived annoyance and loudness were less conclusive. The first principal component explains 70.8% of the variance in the Impulsiveness dataset, with two critical-band rate ranges having larger scores, those being critical-band rates with centre frequencies between 2.3 kHz to 3.2 kHz, and 6.4 kHz to 9.4 kHz. The second principal component explains 8.7% of the variance, with a range of dominating scores for critical-band rates with centre frequencies from 3.7 kHz to 6.4 kHz. The third principal component explains 6.5% of the variance, with a large score for the critical-band rate with centre frequency 7.7 kHz. The scores for the specific Impulsiveness principal components are shown in Figure 6.

worm 2022

Contributions of Specie npuivenes per Ciesbnd Vales to Principal Components Pri Comparer

Figure 6 : Scores of ΔSQM (Impulsiveness) at each critical-band rate for each principal component.

As discussed by Krishnamurthy et al., [9] and Torija et al., [10], the Impulsiveness metric seems to be able to account for the perceptual effect of impulsive sound due to blade-vortex interaction (BVI). In previous research, the authors of this paper investigated the noise emission of a quadcopter with a varying number of rotors operating. As shown in Figure 7, increasing the number of propellers operating from 2 to 4 leads to a substantial increase in high frequency noise. This increase can there- fore be assumed to be due to rotor-rotor interaction noise, which is unsteady broadband noise in nature. This seems to be an important factor for the perception of drone dominance when embedded in the soundscape, and thus is accounted for by the Impulsiveness metric in the critical-bands in the high frequency region.

Figure 7 : Frequency spectra of a quadcopter with a varying number of propellers operating.

worm 2022

5. CONCLUSIONS

This paper presents the results of an investigation which aimed to understand the contribution of each SQM to the overall perception of drone noise, and the link between SQMs and specific drone noise characteristics. Based on the findings of a Linear Regression with PCA with critical-band spe- cific SQMs as independent variables and the participants’ responses of perceived annoyance, per- ceived loudness and drone dominance as dependent variables, it can be concluded that:

• Loudness across the whole critical-band range is the main driving factor explaining subjective

variables. • Roughness at critical-band rates with centre frequencies between 1.3 kHz and 5.3 kHz seems

to account for the perceptual effect of rapid fluctuations with time of the higher BPF harmon- ics. • Fluctuation Strength at critical-band rates with centre frequencies between 200 Hz and 400

Hz seems to account for the perceptual effect of amplitude modulation due to rotor BPF in- teractions. • Impulsiveness at critical-band rates in the mid-to-high frequency region seems to account for

the perceptual effect of rotor-rotor interaction noise.

6. ACKNOWLEDGEMENTS

Dr Torija Martinez, Mr Green and Dr Ramos Romero would like to acknowledge the funding provided by the UK Engineering and Physical Sciences Research Council for the DroneNoise project (EP/V031848/1).

worm 2022

7. REFERENCES

1. Burchan, A. (2019) Public acceptance of drones: Knowledge, attitudes, and practice. Technology

in Society, 59, 101180 (2019). 2. Torija, A. J., Self, R. H., & Lawrence, J. L. Psychoacoustic characterisation of a small fixed-pitch

quadcopter. In INTER-NOISE and NOISE-CON Congress and Conference Proceedings , Vol. 259(8), pp. 1884-1894 (2019). 3. Cabell, R., Grosveld, F., & McSwain, R. Measured noise from small unmanned aerial vehicles.

In Inter-Noise and Noise-Con Congress and Conference Proceedings , Vol. 252(2), pp. 345-354 (2016). 4. Torija, A. J., & Nicholls, R. K. Investigation of metrics for assessing human response to drone

noise. International Journal of Environmental Research and Public Health , 19(6), 3152 (2022). 5. Jillings, N., De Man, B., Moffat, D. & Reiss, J.D. Web Audio Evaluation Tool: A Browser-Based

Listening Test Environment. Proceedings of 12th Sound and Music Computing Conference , 2015. Maynooth, Kildare, Ireland. 6. Gwak, D. Y., Han, D., & Lee, S. Sound quality factors influencing annoyance from hovering

UAV. Journal of Sound and Vibration , 489, 115651 (2020). 7. Alexander, W. N., Whelchel, J., Intaratep, N., & Trani, A. Predicting community noise of sUAS.

Proceedings of 25th AIAA/CEAS Aeroacoustics Conference , p. 2686 (2019). Delft, The Nether- lands. 8. Torija, A. J., Li, Z., & Chaitanya, P. Psychoacoustic modelling of rotor noise. The Journal of the

Acoustical Society of America, 151(3), 1804-1815 (2022). 9. Krishnamurthy, S., Christian, A., & Rizzi, S. Psychoacoustic test to determine sound quality met-

ric indicators of rotorcraft noise annoyance. In INTER-NOISE and NOISE-CON Congress and Conference Proceedings , Vol. 258, No. 7, pp. 317-328 (2018). 10. Torija, A. J., Chaitanya, P., & Li, Z. Psychoacoustic analysis of contra-rotating propeller noise

for unmanned aerial vehicles. The Journal of the Acoustical Society of America , 149(2), 835-846 (2021).

worm 2022