Geo-tagged environmental noise measurement with smartphones: Accuracy and perspectives of crowdsourced mapping

Abstract

Environmental noise affects life and health within urban environments through interfering with sleep, rest, study and personal communication. Noise mapping is an important issue of local authorities but due to its requirements (staff, costs and frequency), the available data are limited or outdated. Our aim was to involve people with smartphones in the mapping process and to determine the accuracy of the measurements performed with these devices in a natural environment. The main questions were whether the measured data were dependent on the type of applied software and smartphones. We tested three software (Noise Watch, Noise Meter and Sound Level Meter) and 12 different smartphones. We evaluated the measurements with hypothesis testing and correlation analysis. Although the accuracy of smartphones was below the professional equipment, measurements can be conducted easily due to their availability; thus, a reliability analysis is important. We found that comparison between professional devices and smartphones in a laboratory was misleading as it lacks the environmental factors biasing the measurements. The best method to compare the measurements carried out with smartphones and professional Noise Meters was to use large number of measuring points in a heterogenic outdoor environment where the noise ranged from the low to large values. We revealed that both the applied software and smartphones have relevant effect on the measurements, and, although it is possible to use these devices for noise mapping, one should consider not to apply different software and smartphones. Accordingly, crowdsourcing is not a reliable data collection method because: (1) measurements should be supervised, (2) smartphones’ accuracy should be tested and (3) measurement circumstances should be the same. If any of these requirements are violated, the quality of the resulting maps can be questioned.

Keywords

Citizen science collaborative mapping crowdsourcing environmental impact participatory sensing

Introduction

Environmental noise affects life quality and health within urban environments through interfering with sleep, rest, study and personal communication. Persistent exposure to noise is associated with several risks including heart disease, mental health and hearing impairment (Ballesteros et al., 2014; Basner et al., 2014). Airport and traffic noise are considered the most disturbing (EEA, 2013). The road traffic noise caused by the increase of urban traffic has a serious impact on the quality of urban residential environment (Xia-lin and Ming, 2013).

According to the Environmental Noise Directive of the European Union 2002/49/EG (END), each agglomeration with a population more than 250,000 should create a noise map. The directive regulates the process of the measurements and the kind of improvement that should be proposed by the local government in order to reduce the inconvenience by noise pollution. The Hungarian Government Decree 280/2004 (X.20) in 2004 obliged all settlements with a population more than 100,000 to create a noise map and a plan to reduce noise pollution exceeding the noise levels defined in EU Directive. All these maps should have been prepared by 2012 and there is an order to renew it from 2010 onwards every five years.

Maps can be generated through calculations or measurements. Measurements are usually conducted with special devices and the data are processed with software dedicated to analyse noise propagation. The usage of both devices and software requires special knowledge, and the whole process is very expensive; however, due to the lack of financial resources, local authorities are unable to fund the renewal of these maps. Crowdsourcing may be a good solution to contribute the legislation: volunteers (volunteered geoinformation) can collect data with their smartphones, send them to data centre, and thus, they can be the part of the mapping, and the result may reveal how members of public see their city, and it can help in designing better-managed cities.

In the field of noise mapping, several attempts were performed to involve citizens through developing sensor network and software such as NoiseSpy (Kanjo, 2010), NoiseBattle and NoiseQuest (Garcia-Martí et al., 2014), Ear-Phone (Rana et al., 2010), Noise Watch (EEA), NoiseTube (Maisonneuvre et al., 2009, 2010), WideNoise (Widenoise, 2013; Nold and Francis, 2017); however, only a few paper deals with accuracy issues. From these projects, it seems that the development of NoiseSpy (Kanjo, 2010), Noise Battle, NoiseQuest, Ear-Phone WideNoise, Noise Watch are no longer available, and the development finished; NoiseTube is still available, but users are complaining that the software is not calibrated for newer applications and the maps are not updated. The only scientific project combining crowdsource noise measurement and noise mapping for which development is still ongoing is NoiseCapture (Picaut et al., 2019).

The accuracy of noise measurements is influenced by several components: microphone, software, external factors (such as wind, rain, etc.). Different aspects of accuracy assessment of urban noise mapping can be found in the literature. These are (1) comparisons of maps (comparing maps generated from professional and smartphone measurements; Garcia-Marti et al., 2014; Pődör et al., 2015); (2) calibration aspects, usually in laboratories focusing on the smartphones’ microphones (D’Hondt et al., 2013; Maisonneuvre et al., 2009, 2010; Santini et al., 2009), and under indoor and outdoor circumstances focusing on the accuracy of measurements conducted with smartphones (Garcia-Marti et al., 2013; Kardous and Shaw, 2015; Murphy and King, 2016a, 2016b).

The calibration issue is a crucial part of the measurements. There is software with full calibration options like Ear-Phone (Rana et al., 2010) and NoiseCapture (Picaut et al., 2019). Volunteers have to accomplish several steps for proper calibration to achieve at least a minimal reliability of measurement. The solutions are different from a reference metrological equipment or device to some freely available software playing the reference noise level (preferably between 72 and 102 dB) in a quiet room. Picaut et al. (2019) pointed on that this requirement is almost impossible to be fulfilled, accordingly, approximately 80% of their volunteered participants used uncalibrated smartphones, too.

Hornikx (2016) found that smartphones’ microphones are not planned for noise measurements and their accuracy differs from the professional devices. Furthermore, a real scenario behaves differently as in lab environment, because of reflections or diffractions (Hornikx, 2016), and other natural phenomena like wind or temperature. Those can influence the accuracy of smartphones measurements as well, so the accuracy should be tested under real circumstances (D’Hondt et al., 2013; Maisonneuvre et al., 2009). Murphy and King (2016a) tested in a lab environment several smartphones. Those equipment were mainly iPhones whereas in Hungary 84% of smartphone owners possess Android phones (HWSW, 2017). Accordingly, a thorough analysis is desirable when all these conditions are ensured with a careful selection of the software and device (i.e. smartphone) based on the comparison with a professional one and, besides the laboratory measurements, phones are tested in a real environment.

Our aim was to reveal whether smartphones can be used for noise mapping as a part of crowd sourcing data collection. We had the following hypothesis: (1) there is a software which can collect reliable data compared to a professional one; (2) indoor and outdoor validation provide different results on reliability; (3) among 12 smartphones there should be models where the collected data do not differ significantly from the reference data. We considered a scenario where we do not force citizens to calibrate their devices, but devices were tested with their original calibration settings to simulate a typical user who may not have access to a calibrated sound source.

Data, materials and methods

Study sites

Initially, we conducted a control measurement at Székesfehérvár (Central Hungary), Óbuda University Campus site in a lab and outdoor with two professional devices (Voltcraft SL 200 and Brüel & Kjær 2250) and a Samsung Galaxy S4 smartphone in 2015 (Supplementary material Fig. 1 Case study A). The laboratory was a perfect silent room directly elaborated for noise measurements. The outdoor environment had a varying level of noise loading: we conducted the comparative measurements with the smartphone and a professional Sound Level Meter at four points, two of the points (#1 and #4) were rather noiseless and the others (#2 and #3) the loudest areas in the city with heavy traffic.

Figure 1.

Range of measurements conducted with smartphones and the reference data (minimum, maximum, interquartile range and median; o: outlier data).

Noise mapping was carried out in the Palotaváros district of Székesfehérvár, around a blockhouse area, where 15% of the inhabitants of Székesfehérvár are living. The central bus station and a shopping centre can be found with heavy traffic throughout the day on the right side of the area. In the centre, a park is situated surrounded by 10-story block-houses (Supplementary material Fig. 1. Case study B).

Applied android applications

We tested three Android OS applications (Sound Level Meter, Noise Meter and Noise Watch developed for Android operation system). The applications used the built-in microphone of the smartphones and were able to measure the noise volume in decibels (dB). It is important to note that built-in microphones are optimized to human voice (300–3400 Hz, 40–60 dB) and none of the applications was meant to be used in the place of scientific instrumentation. Devices employing Automatic Gain Control can have critical measurement errors, which is not the inaccuracy of the applications (Sound Meter User Guide, 2019).

Sound Level Meter

The application is a free software and allows the measurement of environmental noise (SOFTDX, 2015). The amount of noise is expressed in decibel (dB) and the measurement is displayed on the screen indicating a minimum, and a maximum and an average noise level (SOFTDX, 2017).

NoiseWatch

NoiseWatch is developed by the European Environmental Agency with the intent to involve citizens in identifying noisy environments. Measurements can be directly uploaded to the NoiseWatch database; meanwhile, an online map is created from the measurement. The application displays the maximum noise level, and the measurement period is limited to a maximum of 10 s; the application calculates the average noise level, which can be displayed on a map (EEA, 2011).

Noise Meter

Noise Meter is able to define LAeq (A-weighted, equivalent sound level), minimum, maximum and peak and LCpeak (C-weighted, Peak Sound Level) in dB. Weighting is performed by the software with electronic filters: ‘A’ weighting is designed to reflect the response of the human ear to noise, whereas ‘C’ weighting is commonly used for the measurement of Peak Sound Pressure level (Bolin et al., 2014).

It has a complex calibration function but we did not apply it as we intended to reveal the accuracy as normal users upload data to the database with minimal effort (Jynasis, 2017).

Applied devices

Voltcraft SL-200 Sound Level Meter

A professional sound level measurement was also applied with a Voltcraft SL-200 (VC). The measurement range of the instrument was 30–130 dB, with the accuracy of ±1.5 dB. The measurable frequency range was 31.5 Hz to 8 kHz. The apparatus included a sound level calibrator, and it can measure the A or the C filter and it can be adjusted to the sampling speed. We applied the A-filter. The timescale of the measurements can be changed to slow (S) or fast (F). We used Lo settings where the range is between 30 and 100 dB, while in H settings the measuring range is 60–130 dB. The device is mounted by a windscreen foam, with which the accuracy measurement at higher wind is not affected as well. Measurements with this device served as reference data.

Brüel & Kjær 2250 Sound Level Meter device

The accuracy for SL-200 instrument was controlled by Brüel & Kjær 2250 (BK) handheld Sound Level Meter device which can execute Class 1 measurement tasks in environmental, occupational and industrial application areas. The device measures only for a predetermined time period, displays and shows the current noise level during measurement, averaged in short intervals and shows the maximum noise during this interval. The range of the measurements can be set (available ranges are: 30–100 dB, 50–120 dB or 70–140 dB; Brüel & Kjær 2250, 2012). Measurements with this device served as reference data.

Smartphones

We applied 12 separate smartphones (Haier w900, Sony Xperia Z1 compact, Samsung Galaxy Fame, Sony Xperia Mini, Sony Xperia Tipo, Asus Zen Pad 7, Sony Xperia Z3, Samsung Galaxy Alpha, Samsung Galaxy S4, Sony Xperia SP, Nokia Lumia 530, LG Optimus 4xHD). These smartphones were typical mid-range ones used by middle-class families in Hungary at that time (The best phones in Hungary, 2016).

Data collection

Software testing

Software testing was performed with a Samsung Galaxy S4 smartphone (installed SLM, NW and NM on it), a Voltcraft SL-200 and a Brüel and Kjær 2250 device in a laboratory and under outdoor conditions at Óbuda University Campus. The Brüel & Kjaer device was calibrated by an official acoustic lab; thus, this device was considered as the reference. The measurements were performed 51 times with 1-min intervals. We repeated the measurements under outdoor conditions on the same day, in the afternoon (15:00–17:00) 51 times with 1-min intervals. Measurements were conducted using the same software parallelly; when all repetitions had been finished, we started the measurements with the next software. Outdoor measurements were performed in constant weather conditions: wind speed was <5 m/s and it was not raining (according to the Hungarian Decree on Noise measurements 284/2007). In each case, we measured A-weighted noise level.

Smartphone testing

We measured the noise level in four points (Supplementary material Fig. 1 – Case study A), for 1 min with the software seemed to be most accurate in the previous phase. Measurements were performed with 12 smartphones and two professional Noise Meters (Voltcraft SL-200, Brüel and Kjær 2250) parallel, in the same time with 20-times repetitions in each measuring point (altogether 4 × 12 × 20 = 960 measurements). We took measurements only in normal situations; thus, in case of unusual events (breaking noise of cars, honking or siren of an emergency car, etc.), we stopped the measurements and excluded them from the experiment. During the measurements, volunteers held the smartphones in their hands in the same position and were not allowed to speak; furthermore, they were lined up at equal distance from the road. The measurement results were recorded manually on a test paper sheet. The collector team consisted of students and a teacher whose task was to give a signal when each measurement should be started and finished, and also to monitor that members of the team should measure at a given distance from the axes of the roads. Each collector team consisted of two persons – one collecting the data with the help of smartphones, the other led the written protocol to record all the data.

Noise mapping

After the testing phase, we performed a high-density noise mapping with the most appropriate software (based on the results of indoor measurements, Section ‘Software testing’) and the most accurate smartphone (based on the experiment with the 12 smartphones, Section ‘Smartphone testing’) at 59 measurement points of the Palotaváros district (Supplemental Figure 1 – Case study B). In each point, we measured the noise level for 5 min and we recorded the calculated average noise level. We executed the measurements twice: once during peak hours (Wednesday afternoon between 15.00 and 17.00) and repeated during off-peak hours (Sunday between 8.00 and 10.00).

Statistical analysis

Measurements followed the normal distribution according to the Shapiro–Wilk test. We performed paired t-test to reveal whether the applications of smartphone and the reference noise measurements had the same mean. Then, we conducted correlation analysis between the smartphone measurements and the professional Noise Meter (Voltcraft): we aimed to find the groups of smartphones having the same error pattern (i.e. the difference of the measurements of the smartphones and the reference data). Results were plotted as a correlation plot (the strength of the correlation was presented as pie charts; i.e. a simple line represents the complete lack of relationship, and a circle with saturated colour means perfect correlation; red indicated negative and blue indicated positive relationship). Variables (i.e. smartphones) having similar correlation pattern were clustered into groups automatically based on Principal Component Analysis (Kabakoff, 2011). PCA monoplot was also determined to provide further visual information on the similar measurement errors of the smartphones. Regression analysis was applied to determine the functional relationship between the smartphones and the reference measurements; furthermore, we performed the Breusch–Pagan test to control heteroscedasticity (if p < 0.05, regression residuals cause heteroscedasticity). Regression models can be biased by influential data which have a large effect on the slope and can distort the model fit. Accordingly, we repeated the regressions with the exclusion of the influential data points applying the Cook’s distances (values of >4/n were excluded). In a standard linear model, the variance of the residuals is assumed to be constant over the values of the response (predicted values) but in case of smartphones, we supposed that the residuals get larger with the increasing loudness. Residual errors were expressed as Standard Error of the Estimate (SEE) as the average difference between the observed and predicted noise values, and we also referred to it as the ratio of SEE and the average noise level (relative error).

Differences between the reference and smartphone data were also plotted in Bland–Altman plots (references against mean differences; Altman and Bland, 1983).

A statistical analysis was conducted with R 3.5.2 (R Core Team, 2018) with the corrgram (Wright, 2018), ggplot2 (Wickham, 2016) and the lmtest (Zeileis and Hothorn, 2002) packages.

Spatial interpolation

Noise mapping based on the measurement of Case study B was carried out using the Inverse Distance Weighting interpolation method of ArcGIS 10.3 software (ESRI, 2014). Noise is a locationally dependent variable; therefore, the farther we are from the noise source, the more the noise decreases (Watson and Philip, 1985); accordingly, weighting with the distance provides appropriate results. The algorithm calculates an unknown value (z) of the grid network involving the closest measurements (z_i) within a given search radius, dividing all measured values with the distance from point z (d_i; or enhancing the locality of the phenomenon) we can raise the distance to a mathematical power, and then we divide this value with the sum of reciprocal value of the distances (equation (1); Achilleos, 2011).

z = \frac{\sum_{i = 1}^{n} \frac{z_{i}}{d_{i}^{p}}}{\sum_{i = 1}^{n} \frac{1}{d_{i}^{p}}}

(1)

In our case the search radius was 200 m, and we applied a power of 2.

Results

Comparison of the Noise Meter applications

We examined the correlation between the BK and VC devices and found a strong relationship between the measured values (r = 0.94, p < 0.001; N = 218); the mean difference was only 1.3 dB. Therefore, in the following measurements and analyses, we applied only the Voltcraft device.

We then analysed the measurement errors of the noise (values recorded with VC minus the ones of smartphone applications installed on a Samsung Galaxy S4). We found that there were considerable differences and the errors depended on the applied software. Generally, SLM values were consistently under the reference data, while NM and NW provided higher values (see Supplemental Figure 2).

Figure 2.

Correlation plot of the errors (reference – smartphone noise measurements) ordered by principal component analysis to help the visual analysis of the relating values. (The smartphones are namely: Haier w900, Sony Xperia Z1 compact, Samsung Galaxy Fame, Sony Xperia Mini, Sony Xperia Tipo, Asus Zen Pad 7, Sony Xperia Z3, Samsung Galaxy Alpha, Samsung Galaxy S4, Sony Xperia SP, Nokia Lumia 530, LG Optimus 4xHD.)

Wilcoxon test showed significant differences for most of the measurements, the difference between the reference values (VC) and NM under outdoor conditions was non-significant (Table 1). Besides the significance, effect size also indicated large effect for NW and SLM, while in case of NM it was low, around zero, under outdoor circumstances.

Table 1.

Accuracy of the smartphone applications using a Samsung Galaxy S4 smartphone and a Voltcraft SL-200 device.

Software	Conditions	Mean accuracy (dB)	SD of accuracy (dB)	Z-value	Significance/effect size (r)
Noise Meter	Laboratory	9.39	4.18	–6.206	p < 0.001 (0.61)
	Outdoor	–0.98	2.51	–0.154	p = 0.878 (0.01)
NoiseWatch	Laboratory	14.68	5.75	–6.317	p < 0.001 (0.61)
	Outdoor	2.09	4.31	–3.935	p < 0.001 (0.38)
Sound Level Meter	Laboratory	–4.76	4.70	–6.538	p < 0.001 (0.53)
	Outdoor	–15.2	5.23	–6.901	p < 0.001 (0.61)

Analysis of differences and functional relationship between noise measured by smartphones and the reference data

In general, there was no reliable functional relationship between the measurements conducted with smartphones and the reference data; even the highest determination coefficient was only 0.43 (Supplemental Figure 3). It meant that the measurements cannot explain the variance of the reference data properly; i.e. microphones were able to record reliable data and there is no possible way to transform the data of the smartphones to have more realistic values. The relative error of the equations was between 9.3% and 10.5%. Exclusion of influential data points from the regression, R² improved with 0.07–0.23, and the SEE decreased with 0.69–1.44 dB (Table 2). However, our presumption on heteroscedasticity was not confirmed, all the regressions between the reference data and the measurements of the smartphones were homoscedastic according to the Breusch–Pagan test (Table 2).

Figure 3.

Relationship between the reference data and smartphone measurements (a) peak hours and (b) non-peak hours.

Table 2.

Description of linear regressions between noise measurements of Smartphones and the reference data.

Smartphone	R²	Sig.	SEE	Breusch–Pagan statistic	Sig.
Sony Xperia Tipo	4.69 (5.78)	<0.001	0.43 (0.26)	0.188	0.66
Sony Xperia Z3	4.74 (5.81)	<0.001	0.38 (0.25)	1.710	0.19
Sony Xperia Mini	3.92 (5.13)	<0.001	0.59 (0.43)	3.250	0.07
Sony Xperia SP	4.26 (5.62)	<0.001	0.51 (0.3)	0.305	0.58
Sony Xperia Z1 compact	4.08 (5.28)	<0.001	0.53 (0.38)	0.070	0.79
Samsung Galaxy S4	4.07 (5.51)	<0.001	0.54 (0.33)	2.931	0.08
Samsung Galaxy Alpha	4.54 (5.23)	<0.001	0.47 (0.4)	3.380	0.06
Samsung Galaxy Fame	4.64 (5.64)	<0.001	0.45 (0.3)	0.595	0.44
Asus ZenPad 7	3.98 (5.39)	<0.001	0.59 (0.36)	0.486	0.48
LG Optimus 4xHD	4.53 (5.35)	<0.001	0.47 (0.37)	2.293	0.13
Nokia Lumia 530	4.25 (5.47)	<0.001	0.51 (0.34)	1.620	0.20
Haier W900	4.61 (5.73)	<0.001	0.41 (0.27)	3.407	0.06

p: significance, values in brackets show the model fit parameters without influential point exclusion; SEE: standard error of the estimate.

Although the relationships cannot be described properly with a functional way, it did not mean that smartphones are inappropriate to use for noise measurements: the other side of the issues was whether the differences were significant or did not reach high magnitude. The paired t-test revealed that the Sony Xperia Z3 provided the best results related to the ones measured with the professional Voltcraft device (Supplemental Figure 3, Figure 1 and Tables 2 and 3). Even the same producer’s phones had varying results and the difference of Sony Xperia Z1 compact and the Haier W900 were the highest related to the reference. The range of the measurements was high in case of Samsung Galaxy Fame and the LG Optimus 4xHD phones. Considering the pairwise differences between the reference data and the smartphone measurements, the range is wide: at some cases the difference was 0 but in case of some phones (Sony Xperia Z3, Sony Xperia Mini, Sony Xperia Z1 compact) the difference was over 40 dB; the maximal difference was about 25 dB on average.

Table 3.

Result of the paired t-test conducted on the measurements with smartphone and the reference data, ordered by the ascending mean differences (p < 0.05).

Smartphone type	Mean difference	SD	t	Sig. (2-tailed)
Sony Xperia Z1 c – reference	–16.3225	9.3203	–15.664	0.000
Samsung Galaxy Fame – reference	–11.4163	9.0534	–11.279	0.000
Sony Xperia Tipo – reference	–6.5425	8.7660	–6.676	0.000
Sony Xperia Mini – reference	–6.4462	6.3588	–8.953	0.000
Sony Xperia P – reference	–3.7350	8.0251	–4.163	0.000
Samsung Galaxy Alpha – reference	–3.4225	7.0340	–4.352	0.000
Asus Zen Pad 7 – reference	–3.2163	8.0520	–3.573	0.001
Sony Xperia Z3 – reference	–1.1975	9.7428	–1.099	0.275
Sony Xperia SP – reference	1.9812	6.5826	2.692	0.009
Nokia Lumia 530 – reference	2.5025	7.6478	2.927	0.004
Samsung Galaxy S4 – reference	2.6275	7.6556	3.070	0.003
LG Optimus 4xHD – reference	4.1950	8.8732	4.229	0.000
Haier W900 – reference	13.1363	8.1711	14.379	0.000

Correlation between the differences of smartphone measurements and the reference data

We determined the correlations among the errors (reference – smartphone data) and found that the smartphones had unique errors, the correlations between the smartphone pairs were low (between –0.31 and 0.57) but the average was 0.14. Based on the similarities we ordered the smartphones into groups on the correlation plot using a PCA-based approach but the pattern does not reflect definite clusters due to the low correlations (Figure 2): the Haier w900, Sony Xperia Z1 compact, Samsung Galaxy Fame, Sony Xperia Mini, Sony Xperia Tipo and the Asus Zen Pad 7 belonged to a group, while the Samsung Galaxy S4, Sony Xperia SP, Nokia Lumia 530 and the LG Optimus 4xHD formed another group. The Sony Xperia Z3 and the Samsung Galaxy S4 did not belong to any groups.

High-density mapping with smartphone case study B

As parallel measurement found NM as the best software in an outdoor environment, and Sony Xperia Z3 provided the best results among the smartphones; therefore, we performed a high-density noise measurement campaign in 2017 using the Sony Xperia Z3 with NM. Results measured in Palotaváros Székesfehérvár (see Figure 3) had a strong correlation with the professional device (VC) both during peak-hours (r = 0.969, p < 0.05) and off-peak hours (r = 0.984, p < 0.05).

Differences usually did not follow a trend according to the Bland–Altman plot (Figure 4); i.e. larger noise level did not have effect on the noise measured by the smartphone. The average difference was 2.36 ± 1.55 dB (4.26% difference from the reference data with 2.69% SD), and only the #7, #8 and #10 measuring points had larger difference than 2×SD with relatively larger values measured with smartphones.

Figure 4.

Bland–Altman plot of peak hour measurements (–: mean difference; –-: 2 × standard deviation of differences; ^…: zero difference).

The map interpolated from the measurements of the smartphone (Supplemental Figure 4), confirming the results of regression analysis, correctly showed the areas of the lowest noise pollution (Park) and the extreme noisy areas such as the bus station. The difference between measurements conducted with the smartphone measurement and professional device was between 1 and 4 dB (measuring points #7, #8 and #10). The highest difference is connected to the noisiest areas where the difference is 4 dB meanwhile in quiet areas such as parks, the difference is less than 1 dB.

Discussion

We revealed that smartphones and applications behave differently under indoor laboratory and outdoor natural circumstances. In laboratory environment only the sensitivity and quality of inbuilt microphone matters, while outdoor environment brings many other factors such as local meteorological factors, topography and interference sound waves caused by reflections from the buildings. Furthermore, as smartphones are not prepared to scientific measurements and have different hardware quality, they react differently for these phenomena. Our results are identical with the findings of Murphy and King (2016b) emphasizing that for smartphones, outdoors measurements are more important related to laboratory because of the real circumstances.

Difference between the smallest and the highest median values captured by smartphones was 30.8 dB which call the attention that using different phones in a noise mapping project can be biased relevantly by the phones themselves. The usage of a selected smartphone will have a consistent error over the area, while including more types cause differences due to the sensitivity of the inbuilt microphones.

As Kardous and Shaw (2014) and Murphy and King (2016a) proved that software written for iOS platform were more reliable than those for Androids in a laboratory environment, they had no comparative analyses for that for outdoor circumstances; nevertheless, the GlobalWebindex (2018) reported that Android devices were more widespread globally. Although Murphy and King (2016a) found NoiseMeter as an Android application not reliable, Pődör and Révész (2014) and Pődör et al. (2015) found it more reliable than other freely available application, and in this current study we reached the same result.

We found that there was a difference between environmental noise maps based on the measurement of smartphones and VC. Similarly, Murphy and King (2016b) based on iOS smartphones found that there was a significant difference between traditional and smartphone-based strategic noise maps: smartphones-based techniques generally undermeasured the environmental noise level compared to the professional devices. Concerning simultaneous field measurement with smartphones with iOS app and traditional Sound Level Meter, Murphy and King (2016b) and Aumond et al. (2017) found that smartphones regularly over-measured the noise levels. Our measurements provided similar results as the detected noise levels varied by the applied Android application: in some cases values were consistently under the reference data, while other software (NM and NW) provided larger data. However, there was an application (NM) which was able to produce acceptable measurements (i.e. not significantly differing from the reference).

Similarly to Maisonneuvre et al. (2009, 2010), we recognized that using the same software on different smartphones measurements had a wide range with a maximal difference of 25 dB, which is not surprising due to the lack of the calibration. Even the identical smartphones using the same application can produce different values which corresponds to former findings of Pődör et al. (2015), D’Hondt et al. (2013) and Aumond et al. (2017). Furthermore, errors did not have a pattern over the smartphones, the differences usually did not correlate.

The large variability can be explained as Aumond et al. (2017) stated smartphones are more sensitive to noises that are typical in an urban environment in the streets than professional devices. The other main reason for the differences lies behind the fact that smartphones are using various types of filters (Santini et al., 2009) which can alter the result of the measurement.

Differences of the testing phase and the mapping measurements were risen by the captured data range: testing phase was performed with four points with 20 repetitions and represented lots of similar measurements with low variance, while mapping consisted of 59 measuring points with relatively larger variance. It was important when we compared the results to the reference data: when range of the captured data was larger, results were relevantly better (Supplemental Figures 3 and 4). R² ranged between 0.26 and 0.43 or with removing the influential data the best R² improved with even 0.23, from 0.36 to 0.59 in the testing phase, but model fits were still worse than the mapping phase’s 0.90–0.94 R². Although smartphones performed better for mapping, and it was a good outcome from the aspect of using them for crowdsourced mapping, testing was also important to point on the uncertainty of the smartphones when the noise level distributed only in a limited range. Testing was appropriate to filter out the smartphone with the best sensitivity, which also performed well in the mapping of different noise levels. However, we suggest to use more points with different noise levels for the smartphone selection related to conducting testing with 1–2 points with several repetitions.

The visual interpretation of the maps also support ideas that smartphone measurement can reveal noisy and quiet places in urban environment as explained in D’Hondt et al. (2013), Aumond et al. (2017), Pődör and Zentai (2017) and Miller et al. (2016).

Some authors emphasize the relevance of crowdsourced smartphone-based noise measurements (e.g. Marques and Pitarma, 2019; Zuo et al., 2016) while others also point on limitations (Picaut et al., 2019). We agree with the relevance of the potential of the topic but came to the same conclusion as Picaut et al. (2019): only the supervised data collection can be acceptable for mapping. This is the only way to control the data quality which is crucial when the aim is to provide reliable noise maps. Generally, in spite of the lack of calibration, smartphones can be the part of noise mapping based on the high R²-values (0.94), but only with careful preparation, and not in a crowdsourcing way: the high data density can be gained with well-organized campaigns after parallel measurements with professional devices. Then, a comprehensive statistical data comparison can justify the application of smartphones if the evaluation confirms an acceptable accuracy. Thus, a better and detailed mapping can be carried out due to easier data capture and, consequently, with higher spatial resolution. As we pointed on, there is no general suggestion for the type of smartphone, as the available devices always change, and each device has its own specification, and possibly even two smartphones of the same type can provide different results in parallel measurements.

Conclusions

We aimed to reveal the reliability of smartphones in noise mapping for a possible crowd-sourced mapping campaign. The method raises several issues to be cleared: is there a reliable application for smartphones to measure noise pollution, is there a reliable smartphone, and the measurements should have consistent errors (i.e. should fulfil the prerequisite of homoscedasticity).

We revealed that indoor measurements are not appropriate to confirm the reliability of measurements conducted with smartphones due to the low variance of the data; i.e. in this noise range smartphones’ microphones are not efficient. Outdoor measurements are biased by the environmental effects, which influence the microphones, and better to perform it in a study area where noise level is varying from lower to higher. Our hypothesis was that there is at least one noise level meter application for smartphones and at least one reliable smartphone. In this case, we found that Noise Meter application has acceptable measurement results under outdoor circumstances (differences were not significant related to the professional devices). We revealed that smartphones usually have large differences (with a median of 30 dB) related to the reference data, but there are models with reliable measurements, in our case it was the Sony Xperia Z3. Test measurements were carried out at four points with 20 repetitions, and this test was more sensitive to homogenous data; i.e. the variances were rather small by data pairs (smartphone and reference). However, when we used the selected smartphone for mapping, the device performed well: when variance was larger, the relationship was better. The R² was >0.9, and the differences between data pairs usually stayed in the range of 2×SD. Consequently, smartphones can effectively help noise mapping, but a careful smartphone selection, with comparing the noise measurements with a calibrated device, is needed to verify the accuracy. Our suggestion is the Noise Meter for the software, but smartphones develop quickly; our selection, the Sony Xperia Z3 gets outdated very soon; thus, our work serves as a methodological guide.

Supplemental Material

sj-pdf-1-epb-10.1177_2399808320987567 - Supplemental material for Geo-tagged environmental noise measurement with smartphones: Accuracy and perspectives of crowdsourced mapping

Supplemental material, sj-pdf-1-epb-10.1177_2399808320987567 for Geo-tagged environmental noise measurement with smartphones: Accuracy and perspectives of crowdsourced mapping by Andrea Pődör and Szilard Szabó in EPB: Urban Analytics and City Science

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The research was funded by the TNN 123457 NKFI project.

ORCID iD

Andrea Pődör

Supplemental material

Supplemental material for this article is available online.

Andrea Pődör, MSc in Cartography, PhD in Earth Sciences, is an associate professor at the Obuda University, Alba Regia Technical Faculty, Institute of Geoinformatics in Székesfehérvár, Hungary. Andrea recently does research in crowdsourcing noise maps and fear of crime.

Szilard Szabó, MSc and PhD in Geography, is a full professor at the Department of Physical Geography and Geoinformatics, University of Debrecen, Hungary. Szilard does research in Geoinformatics (GIS), Geomorphology and Remote Sensing.

References

Achilleos

(2011) The inverse distance weighted interpolation method and error propagation mechanism – Creating a DEM from an analogue topographical map. Journal of Spatial Science 56: 283–304.

Altman

Bland

(1983) Measurement in medicine: The analysis of method comparison studies. The Statistician 32(3): 307.

Aumond

Lavandier

Ribeiro

, et al. (2017) A study of the accuracy of mobile technology for measuring urban noise pollution in large scale participatory sensing campaigns. Applied Acoustics 117: 219–226.

Ballesteros

Fernández

Flindell

, et al. (2014) Estimating leisure noise in Spanish cities. Applied Acoustics 86: 17–24.

Basner

Babisch

Davis

, et al. (2014) Auditory and non-auditory effects of noise on health. The Lancet 383(9925): 1325–1332.

Bolin

Bluhm

Nilsson

(2014) Listening test comparing A-weighted and C-weighted sound pressure level as indicator of wind turbine noise annoyance. Acta Acustica United with Acustica 100(5): 842–847.

Brüel & Kjær 2250 sound level meter equipment manual (2012) Available at: https://www.bksv.com/downloads/2250/2250_User_Manual_-_BE171315.pdf (accessed 2 January 2020).

D’Hondt E, Stevens M and Jacobs A (2013) Participatory noise mapping works! An evaluation of participatory sensing as an alternative to standard techniques for environmental monitoring. Pervasive and Mobile Computing 9(5): 681–694.

European Environment Agency, European Union (EEA) (2013) Environment and human health. Joint EEA-JRC report No5/2013, EUR 25933 EN. ISSN 1725-9177.

10.

ESRI (2014) ArcGIS Desktop: Release 10. Redlands: Environmental Systems Research Institute.

11.

European Environment Agency (2011) Noise levels across Europe now mapped by new EEA application. Available at: http://www.eea.europa.eu/highlights/noise-levels-across-europe-now (accessed 5 January 2019).

12.

Garcia-Martí

Torres-Sospedra

Rodríguez-Pupo

(2014) A comparative study on VGI and professional noise data. In: Huerta

Schade

Granell

(eds) Connecting a Digital Europe through Location and Place. Proceedings of the AGILE'2014 International Conference on Geographic Information Science, Castellón, Spain, 3–6 June. Czhech: AGILE. ISBN: 978-90-816960-4-3.

13.

GlobalWebIndex (2018) The device trends to know. Available at: https://www.globalwebindex.com/reports/device (accessed 5 January 2019).

14.

Hornikx

(2016) Ten questions concerning computational urban acoustics. Building and Environment 106: 409–421.

15.

HWSW (2017) How many smartphones sold at home last year? Available at: https://www.hwsw.hu/hirek/56808/idc-2016q4-okostelefon-tablet-laptop-desktop-pc.html (accessed 5 January 2019).

16.

Jynasis (2017) Available at: https://play.google.com/store/apps/details?id=com.pjw.noisemeter

17.

Kabakoff

(2011) R in action: Data Analysis and Graphics in R. Shelter Islands, New York: Manning Publication.

18.

Kardous

Shaw

(2015) Do sound meter apps measure noise levels accurately? Sound & Vibration. Available at: http://www.sandv.com/downloads/1507kard.pdf (accessed 18 August 2018).

19.

Kanjo E (2010) NoiseSpy: A real-time mobile phone platform for urban noise monitoring and mapping. Mobile Networks and Applications 15(4): 562–574.

20.

Maisonneuve N, Stevens M, Niessen ME and Steels L (2009) Noisetube: Measuring and mapping noise pollution with mobile phones. In Information technologies in environmental engineering, Springer, pp. 215–228.

21.

Maisonneuve

Stevens

Ochab

(2010) Participatory noise pollution monitoring using mobile phones. Information Polity 15(1,2): 51–71.

22.

Marques G and Pitarma R (2019) Noise mapping through mobile crowdsourcing for enhanced living environments. In: Rodrigues, J.M.F., Cardoso, P.J.S., Monteiro, J., Lam, R., Krzhizhanovskaya, V.V., Lees, M.H., Dongarra, J.J., Sloot, P.M.A. (eds.) Computational Science—ICCS 2019, pp. 670–679. Springer International Publishing, Cham 12-14-June 2019 Fro, Portugal, Available at: http://doi-org-443.webvpn.fjmu.edu.cn/10.1007/978-3-030-22744-9_52

23.

Miller

Springthorpe

Murphy

, et al. (2016) Environmental noise mapping with smartphone applications: A participatory noise map of West Hartford, CT. In: INTER-NOISE and NOISE-CON Congress and Conference Proceedings 252, No. 2, pp.445–451. Institute of Noise Control Engineering

24.

Murphy

King

(2016a) Testing the accuracy of smartphones and sound level meter applications for measuring environmental noise. Applied Acoustics 106: 16–22.

25.

Murphy

King

(2016b) Smartphone-based noise mapping: Integrating sound level meter app data into the strategic noise mapping process. The Science of the Total Environment 562: 852–859.

26.

Nold

Francis

(2017) Participatory sensing: Recruiting bipedal platforms or building issue-centred projects? In: Loreto

Haklay

Hotho

, et al. (eds) Participatory Sensing, Opinions and Collective Awareness. Berlin: Springer Verlag, pp.213–237.

27.

Picaut

Fortin

Bocher

, et al. (2019) An open-science crowdsourcing approach for producing community noise maps using smartphones. Building and Environment 148: 20–33.

28.

Pődör

Révész

(2014) Noise map: Professional versus crowdsourced data. In: Huerta

Schade

Granell

(eds) Proceedings of the 17th AGILE International Conference on Geographic Information Science, June, 3-6, Castellón, Spain. AGILE Digital Editions

29.

Pődör

Révész

Ócsai

, et al. (2015) Testing some aspects of usability of crowdsourced smartphone generated noise maps. In: Car

Griesebner

Strobl

, et al. (eds) GI_Forum 2015 – Geospatial Minds for Society. Salzburg: Universität Salzburg, p.354.

30.

Pődör

Zentai

(2017) Educational aspects of crowdsourced noise mapping. In: Peterson

(ed.) Advances in Cartography and GIS Science: Selections from International Cartographic Conference. Omaha: Springer International Publishing, pp.35–46.

31.

Rana RK, Chou CT, Kanhere SS, Bulusu N and Hu W (2010). Ear-phone: an end-to-end participatory urban noise mapping system. In Proceedings of the 9th ACM/IEEE International Conference on Information Processing in Sensor Networks. ACM, pp. 105–116.

32.

R Core Team, 2020. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria; https://www.R-project.org/ (accessed 08 January 2021).

33.

Santini

Ostermaier

Adelmann

(2009) On the use of sensor nodes and mobile phones for the assessment of noise pollution levels in urban environments. In: 2009 Sixth International Conference on Networked Sensing Systems (INSS).Pittsburgh, PA, pp. 1-8, doi: 10.1109/INSS.2009.5409957.

34.

SOFTDX (2015) Sound Level Meter download page Available at: https://play.google.com/store/apps/details?id=com.softdx.soundmeter

35.

Sound Meter User Guide (2019) Available at: https://codecanyon.net/item/sound-meter-android/9801572 (accessed 18 March 2019).

36.

The best phones in Hungary (2016) Available at: https://player.hu/tech-3/2016-legjobb-telefonjai/

37.

Watson

Philip

(1985) A refinement of inverse distance weighted interpolation. Geoprocessing 2: 315–327.

38.

WideNoise. Available at: edutechwiki.unige.ch

39.

Wickham H (2016) ggplot2: Elegant Graphics for Data Analysis. Springer-Verlag New York, 2016.

40.

Wright

(2018) corrgram: Plot a Correlogram. R package version 1.13. Available at: https://CRAN.R-project.org/package=corrgram

41.

Xia-Lin MA and Ming CAI (2013) Rendering of dynamic road traffic noise map based on Paramics. Procedia-Social and Behavioral Sciences 96: 1460–1468.

42.

Zeileis

Hothorn

(2002) Diagnostic checking in regression relationships. R News 2(3): 7–10.

43.

Zuo

Xia

Liu

, et al. (2016) Mapping urban environmental noise using smartphones. Sensors 16(10): 1692.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.88 MB