Health effects of multidimensional spatiotemporal environmental exposure

High-Spatiotemporal-Resolution Environmental Mapping based on Stationary-mobile Sensing

Abstract

Personal exposure assessment is crucial for public health research, which requires multi-dimensional environmental data with high spatiotemporal resolution. Previous studies mainly relied on either stationary or mobile sensing methods to measure environmental indicators. However, these approaches faced limitations in spatial coverage or temporal continuity, which may lead to inadequate accuracy for high-resolution mapping. To address this gap, this paper proposes a "Stationary-mobile Sensing" paradigm that integrates stationary and mobile sensing through dynamic calibration and spatiotemporal fusion. Experiments were conducted within a 3-hectare dormitory area, utilizing six multi-sensor devices to collect various environmental data including temperature, noise, and PM_2.5 data. By combining the temporal continuity of stationary sensing with the spatial granularity of mobile sampling, this integrated approach enabled 15m-resolution environmental mapping with temporal variation for each indicator. Results demonstrated that estimation error reductions of the proposed method reduced over 50% compared to single-mode approaches, achieving RMSE values of 0.61°C for temperature, 1.48 μg/m³ for PM_2.5, and 0.90 dB for noise. This optimized sensing and mapping method can help to enhance the precision in individual exposure assessment, supporting targeted interventions to mitigate urban health risks.

Introduction

Background

According to a 2016 report by the World Health Organization, approximately 23% of global deaths and 24% of diseases can be attributed to environmental factors (Prüss-Üstün, Wolf et al., 2016), of which air pollution, extreme temperatures, and ambient noise are essential elements (World Health Organization, 2011), and numerous empirical studies have revealed their significant negative impact on public health (Yin and Wang, 2017). Hence, it is critical for public health to identify and prevent these risks in the outdoor air, thermal, and acoustic environments in cities, which rely on multidimensional environment monitoring and map construction.

As urban research focuses increasingly on problems at finer scales, environment maps with high spatial and temporal resolution grow ever more vital. On one hand, it can provide rich information to accurately assess personal exposure, which serves as a crucial intermediary to investigate the health effects of microscopic built environment factors. On the other hand, it can delineate nuanced environmental profiles across complex cityscapes and enable administrators and planners to pinpoint hotspots where risks often occur. This can thereby guide health-oriented planning, design, and management interventions to create salubrious, equitable, and sustainable spaces that promote public health and wellbeing.

Literature Review

Scholars from various fields have investigated methods for environment monitoring and environment mapping in different dimensions.

Environmental indicators monitoring and environment map inferring based on remote sensing data is a mainstream and relatively mature method, commonly applied in the construction of thermal environment maps (Hanani, Riniwati et al., 2024; Zhang, X., Zhou et al., 2021) and air pollution maps (Li, J., Garshick et al., 2021; Zhan, Gao et al., 2020). Its advantages include the ability to measure surface temperature and pollutant concentrations over a wide spatial range while achieving high accuracy. However, it also has limitations including coarser spatial and temporal resolutions for the resulting environment maps, which are typically only at a daily- and kilometer-level.

Another common approach is to utilize stationary sensing data from weather stations or other sources, applicable to the thermal (Lam, Ong et al., 2021; Yang, Xiaoshan, Peng et al., 2020), outdoor air (Alli, Clark et al., 2023; Wang, Alli et al., 2022) and acoustic environment (Can, Van Renterghem et al., 2011; Geraghty and O’Mahony, 2016) map construction. This method enables very high temporal resolution data capture over extensive timespans, while the sparse distribution density of fixed sensors provides limited spatial coverage of sampling. Constructing environment maps thus requires spatial interpolation with the application of Land-use Regression (LUR) modeling and machine learning (ML). However, achieving accurate maps with high spatiotemporal resolution through this method often requires extensive historical data or the incorporation of various additional datasets (such as built environment attributes) for training. This imposes strict requirements on the data, making it challenging to implement.

Mobile sensing is an emerging environment monitoring approach (Kousis, Manni et al., 2022), using people or vehicles to carry lightweight sensing devices along fixed or adaptative routes. It can help obtain higher spatial coverage of sampling, which compensates for the limitations of stationary sensing. This "snapshot" approach has also been proven to achieve high measurement accuracy, and is therefore widely applied in the monitoring and mapping of noise (Boumchich, Picaut et al., 2024; Guillaume, Aumond et al., 2019; Pődör, Szabó et al., 2021), urban heat island (Clay, Guan et al., 2016; Kawakubo, Arata et al., 2023; Kim, Kim et al., 2019), and outdoor air pollution (Anjomshoaa, Duarte et al., 2018; Chambliss, Preble et al., 2020; Ganji, Youssefi et al., 2023; Ji, Han et al., 2023). However, mobile sensing also has obvious disadvantages in terms of temporal coverage: the data acquired are discrete, sparse, and ephemeral, typically limited to instantaneous value measurements while lacking temporal continuity, which poses a challenge to accurately establishing high temporal resolution environment maps for high dynamic indicators such as noise.

Some recent studies have combined mobile sensing with stationary sensing and initially explored a collaborative sensing mode, denoted as “Stationary-mobile Sensing” in this paper. In this mode, the mobile sensing data can be calibrated by the nearby stationary devices to enhance credibility. The two compensate for each other's deficiencies in spatiotemporal coverage, enabling inference for more accurate high-resolution thermal environment maps (Li, Y. and Song, 2019; Parison, Chaumont et al., 2023) and air pollution maps (Liu, Chen et al., 2021; Xu, Liu et al., 2025). However, due to the limited number of studies on collaborative sensing, there remains uncertainty about how to optimally integrate these two sensing modes across device deployment, data acquisition, and map fitting processes — either to maximize estimation accuracy or to achieve the most cost-effective solution while maintaining acceptable map accuracy. Besides, when studying the noise level, a factor marked by significant spatial and temporal variability, researchers compromised the temporal resolution to achieve desired levels of accuracy, which led to the specific noise dynamics over shorter intervals remaining inadequately represented (Can, Dekoninck et al., 2014). Recently, a rotating mobile sensing method has been proposed, which can help acquire high spatiotemporal resolution noise maps (Zhang, Y., Zhao et al., 2023). However, these noise maps lack temporal continuity, and the noise predictions at target locations only represent the noise levels near the sampling time.

In general, several critical research gaps are revealed in environmental sensing and mapping as follows:

1. Most studies focus on parameters from a single dimension (e.g., only temperature and humidity from the thermal aspect), rather than capturing multiple factors and their interactions simultaneously.

2. The two conventional sensing modes face inherent limitations. Stationary sensing has limited spatial coverage, while mobile sensing cannot monitor locations continuously. This makes it difficult to create accurate environment maps with both good coverage and frequent updates.

3. Although combining fixed and mobile sensing could address these limitations, research on effective integration strategies remains limited, particularly for different environmental factors.

Research Objectives

To address these research gaps, this study proposes a multi-dimensional environmental mapping method that achieves high accuracy with high spatiotemporal resolution. The primary objectives are threefold:

1. To acquire multi-dimensional environmental data through an integrated multi-sensor system that simultaneously measures typical indicators including temperature, noise level, and PM_2.5 concentrations.

2. To develop a novel “Stationary-mobile Sensing” approach that surpasses existing methods in mapping accuracy (manifested by having a smaller RMSE) while achieving fine-grained spatiotemporal resolution (15-meter spatial resolution and 15-minute temporal resolution).

3. To systematically investigate the optimal implementation strategies of the proposed collaborative sensing framework across different environmental indicators, with consideration of various performance metrics including estimation accuracy and robustness.

Methodology

Experiment Design

Research Site

A 3-hectare student dormitory area on a university campus in Beijing was selected as the study area. This dormitory area offered complexity comparable to real-world urban residential settings, making it ideally suited for analyses of personal exposures. Moreover, the area satisfied the spatial granularity requirements considering the target scale, number and density of stationary sensors, and feasibility of mobile sensing collection (e.g. speed and time). With permission, the experiments were conducted on March 2nd, 2023 under calm, dry weather, which avoided extreme air conditions that could impact equipment accuracy.

Sensing Devices

Six identical sensing devices, the Bat Sound (EVA03) developed by Urbanxyz, were utilized for both stationary and mobile sensing to minimize equipment-inherent measurement errors. These compact (218mm × 147mm × 42mm), 0.65kg lightweight devices were chosen because they can be easily mounted on surfaces or carried, and they integrate multiple sensors to simultaneously capture 10 indicators including temperature, noise levels, and PM_2.5 concentrations, which eliminates the need for multiple separate sensors or additional temporal alignment efforts. Data is recorded at 10-second intervals with location information from the integrated GPS module, enabling continuous geo-tagged data collection for spatiotemporal mapping. It should be noted, however, that the stability of these devices is not comparable to laboratory-grade equipment, necessitating pre-experiment verification and calibration (detailed in Section 2.2.1).

For stationary sensing, devices were affixed to tree trunks using brackets and hoops, orienting towards open areas with ample ventilation (Figure 1 left). Devices were distant from the building surface and 1.6m above ground to prevent interference to precise measurement.

For mobile sensing, a device was attached to the outside of the backpack using magnets, brackets, and safety ropes. The experimenter steadily carried the backpack and rode along a set S-shaped route at a constant speed (5km/h in this study) to achieve equally spaced data points (Figure 1 right). Careful carriage minimized backpack shaking and bicycle-generated noise throughout the acquisition process to ensure stable, accurate data sampling.

Figure 1. Devices in measurement condition: stationary device (left) and mobile device (right)

Roles and Characteristics of Stationary and Mobile Sensing

Stationary and mobile sensing each offer distinct strengths. Stationary sensing achieves greater temporal coverage and captures environment dynamics well. In contrast, mobile sensing obtains greater spatial coverage and effectively presents spatial heterogeneity.

Stationary devices remain fixed in space and receive less environmental interference, ensuring data with greater accuracy and credibility. In comparison, mobile sensing devices are susceptible to wind, vibration, vehicle noise, and other fluctuations, which may result in deviations of the results from the true value. Therefore, for characterizing the environment at a given spatiotemporal location, stationary sensing data can be reasonably taken as ground truth. Accordingly, mobile sensing data can be calibrated by proximal stationary devices during the sampling process.

Device Deployment for Stationary and Mobile Sensing

One device was utilized for mobile sensing. Four stationary sensing points were set, each equipped with one device. An additional stationary device served as a reference for validation. The mobile, stationary, and validation devices were labeled as device0, device1-4, and device5 respectively for clear identification.

For the mobile sensing, an S-shaped traversal path was planned to provide comprehensive, uniform spatial coverage across the study area (Figure 2). Constraints on the 10-second device data feedback interval dictated that the riding speed should not be too rapid, as data points would sparse out, nor too slow, to ensure efficient collection. Balancing both factors, a steady 5 km/h (1.5 m/s) riding speed was maintained. This achieved a 15m data point spacing, with one route lap completed in about 15 minutes, meeting experimental requirements.

Figure 2. Distribution of the stationary sensing points and the mobile sensing path

For the stationary sensing, devices were placed near the mobile sensing path to facilitate data integration. This ensured adequate spatiotemporal overlap between data from the stationary and mobile sensors, allowing for effective calibration. Four sensors were evenly mounted at the stationary sensing points (S1-S4) around the periphery of the site to obtain comprehensive spatial sensing coverage. Another sensor, specifically assigned as a reference device to validate the estimation results (which means that only the sensors in S1-S4 are used for estimation and environmental mapping without this sensor), was mounted at the center of the site (S5) in order to make the acquired “ground-truth” more representative.

Data Collection

Device Calibration

A pre-experiment was conducted to guarantee measurement consistency across devices and quantify their relative discrepancies, thereby helping mitigate measuring errors. It was performed from 12:00 to 13:00 on March 2nd, just a few hours prior to the formal experiment. The six devices were co-located on two trees within 2m of point S5, positioned approximately 1.6m above the ground.

All six devices recorded complete and continuous data for temperature, humidity, noise, and air pollution over the pre-experiment duration. After removing outliers, Pearson correlation coefficient (PCC) and intraclass correlation coefficient (ICC) tests revealed consistency between sampling results from those devices, albeit with relatively stable systematic errors. The mean reading across all devices was considered the benchmark. For each indicator, the difference between each device’s reading and the mean value was quantified as the relative systematic error e of that device, which would be used to figure out a more accurate monitoring result in the formal experiment.

Data Acquisition

Due to the sensitivity of data collection and the limitation of the self-powered duration of the equipment, we arranged the pre-experiment and experiment within three representative time periods, namely, lunch time (11:30-13:30), student activity time in the late afternoon (15:30-17:30), and dinner time (17:30-19:30). The intensity of crowd activities is higher and fluctuates greatly during the above hours, and thus the spatiotemporal heterogeneity of the relevant environmental indicators is more obvious.

Within these periods, data collection of the formal experiment was conducted for about 1 hour and 10 minutes each in the afternoon (15:53 to 16:58) and evening (18:20 to 19:28), and around 807 pieces of valid data were gathered by each device concurrently (Figure 3). Cell phone GPS software was additionally utilized to log travel trajectories for subsequent spatial coordinate calibration, which could help preclude occasional GPS coordinate deviations.

Figure 3. Mobile sensing trace and mobile/stationary data points in the spatiotemporal view (Tracks of each color in the mobile sensory data represent one collecting lap)

Data Processing

Data Cleaning

Slight position deviations were revealed for some data points during the data inspection process, probably due to the weaker GPS signals of the devices in areas with dense trees and buildings. Therefore, data points whose coordinates deviated far from the road line were manually corrected by referring to the timestamps and replacing them with cell phone GPS track points.

Additionally, for temperature and PM_2.5 concentration, which typically remain stable over short periods, several outliers deviating from local trends were identified. Specifically, for time-series data, a sliding window approach was employed to calculate local means over 1-minute intervals (Yang, Xing, Zhou et al., 2019). Data points with Z-scores exceeding 3 were identified as outliers and replaced with interpolated values (Torres, Nieto et al., 2011; Wang, Alli et al., 2022). It should be noted that this process was not applied to noise data due to its inherent stronger temporal fluctuations.

Moreover, based on the relative systematic error e calculated in Section 2.2.1, the monitoring result of each device was calibrated by subtracting its corresponding e from its reading.

Data Smoothing for Noise Level with Gaussian-Filter

While temperature and PM_2.5 concentrations remain stable over short time periods, noise shows substantial temporal heterogeneity. The sound pressure level of urban noise at a localized spot can fluctuate dramatically within mere seconds. Under the circumstance, mobile sensing through short-term sampling acquires measurements with high randomness, making it difficult to obtain representative noise level data for a specific location. This manifests in potentially large differences even between neighboring points' values.

To address the issue, we implemented the Gaussian Filter, a common smoothing technique that calculates each point's value using a weighted average of itself and nearby signals. Unlike the uniform weighting of the common Mean Filters, its weights decrease with distance, providing gentler smoothing and better performance than a similarly sized Mean Filter (Aftab and Moghadam, 2022). In this way, the noise value at each of the 807 mobile sensing points was replaced by the weighted acoustic average of surrounding noise levels captured within the same minute, which enhanced the representation of the localized noise conditions. Following parameters from existing literature (Can, Dekoninck et al., 2014), a 50m radius was set, with a 20m sigma decay (weights decaying from 1 to 0.04 from center to periphery). After smoothing, the spatial correlation of the data improved markedly, with each reading attaining greater spatial representativeness (Figure 4).

Figure 4. Comparison of raw data and the smoothing result by Gaussian-Filter

Data Aggregation

To create a unified environment map, 67 fixed points p (including point Si) were set at 15m intervals along the mobile sensing path (Figure 2), with 15m radius buffers built around each as aggregation units. This scale was chosen considering the 15m spacing between data points. A smaller scale might result in some units containing only one or no data points from one collecting lap, while a larger scale would reduce mapping resolution, contradicting our aims. After this procedure, the 807 mobile sensing data points were spatially aggregated into the units via intersection, and remapped to the corresponding points p (Figure 5). The aggregated data in each unit represented the environmental indicator values near the corresponding point p during data collection. Subsequent analyses utilized the points p as the fundamental analytical units.

Figure 5. Remapping original sampling points to fixed point p with 15m buffer

Indicator Categorization

As previously discussed, the three environmental dimensions - thermal environment, outdoor air environment, and acoustic environment - exhibit distinct temporal patterns, corroborated by literature and experimental results. Specifically, the former two show less volatility over time, while the acoustic environment reveals greater fluctuations.

Accordingly, temperature and PM_2.5 concentration data were categorized together, while noise data was separated. Diverse approaches were developed per category for data processing, stationary-mobile data collaboration, and map construction. For reference, maps were also constructed from individual sensing mode data using conventional methods. By comparing estimations (through traditional methods versus the method proposed in this paper) against ground truth values at the validation point, the superiority of Stationary-mobile Sensing could be verified.

Results: Spatiotempora Patterns and Accuracy of High-Resolution Environmental Maps

Thermal Environment and Air Pollution: Map Construction and Accuracy Verification

Map Construction Methods of Thermal Environment and Air Pollution

Method 1: Stationary sensing with spatial interpolation

Spatial interpolation based on stationary sensing data is a common method in environment map construction. Since stationary devices sparsely sample limited areas, interpolation estimates indicator values at unmonitored areas to achieve high-resolution spatial coverage.

Inverse Distance Weighted (IDW) interpolation is a classical technique, which assumes the target value equal to the weighted average of proximal observations, with weights negatively correlated to distance. In this study, the indicator value of the target point p at a moment t was determined as:

  

$$\begin{array}{c}\hfill V_{p,t,IDW}=\sum _{}^{}\frac{\frac{V_{i,t,SS}}{{D_{p,i}}^2}}{\sum _{}^{}\frac{1}{{D_{p,i}}^2}},\left(1\right)\hfill \end{array}$$

where $V_{i,t,SS}$ is the indicator value of a stationary sensing point Si at moment t, and $D_{p,i}$ refers to the distance between the target point p and point Si.

Method 2: Mobile sensing with temporal interpolation

Given the thermal environment and air pollution data’s lower temporal volatility, mobile sensing also provides representative snapshots accurately capturing localized environment conditions. However, mobile sensing data for a certain location are often intermittent and discrete, lacking temporal completeness. Existing indoor and outdoor studies usually apply temporal interpolation to estimate indicator values at unsampled times, achieving high- temporal-resolution profiles.

Triple spline interpolation, a typical temporal interpolation approach, was employed in this study. The estimated value of the indicator F(t) at moment t was obtained by constructing a cubic spline function F(x) through known points and extracting the curve’s value at the target moment t.

Method 3: Stationary-Mobile Sensing

The proposed collaborative sensing mode for thermal environments and air pollution consisted of two main aspects. Firstly, the mobile device achieved extensive spatial coverage for data acquisition; Secondly, the stationary devices quantified the average deviations between the two modes near the four stationary sensing points, which were then utilized to calibrate the mobile sensing measurements across all analytical units.

The specific steps were as follows. (1) For the mobile sensing data at each point p, a one-minute temporal threshold was applied for aggregation, and then the average value $V_{p,MS,i}$ was calculated for each pass around point p, with i denoting the number of passes. (2) Cubic spline interpolation was performed on the mobile sensing data located at point p to attain the complete value curve over the full period, where the value at a given moment t denoted $V_{p,t,MS}$ . (3) The average value $V_{s,MS,i}$ of the mobile device each time it passes a stationary point S was calculated as in Step 1, while the average value of the stationary device at point S during the same time period was denoted as $V_{s,SS,i}$ . The deviation $d_i$ between the two is calculated as follow (Figure 6):

  

$$\begin{array}{c}\hfill d_i=V_{s,SS,i}{-V}_{s,MS,i}.\left(2\right)\hfill \end{array}$$

All $d_i$ for all points S were averaged to obtain the average deviation d. (4) The relative deviation was superimposed on the results obtained by spline interpolation, and the final estimated value $V_{p,t}$ was calculated as:

  

$$\begin{array}{c}\hfill V_{p,t}=V_{p,t,MS}+d.\left(3\right)\hfill \end{array}$$

Figure 6. Relative deviation between stationary and mobile sensing data

Estimation and Map Construction Results

Following the calculating method outlined in Section 3.1.1, temperature and PM_2.5 concentration values were estimated for all points p at all moments t throughout the experiment duration. These were visualized to construct thermal environment and air pollution maps. Since street spaces were the primary focus and the points p were used as analytical units, road maps rather than grid maps were formed. The temporal resolution was fine-tuned to 10 seconds, while the 15-meter spacing for point p determined the spatial resolution. The mapping results for four moments are shown here as examples (Figure 7 and Figure 8).

Figure 7. The mapping result for the temperature

Figure 8. The mapping result for the PM_2.5 concentrations

The temperature maps reveal more obvious spatial heterogeneity than temporal variation. Southeast areas with denser vegetation and large lawns consistently show lower temperatures, indicating the cooling effect of green infrastructure. In contrast, southwest areas characterized by numerous booths and crowds congregating, maintained higher temperatures, indicating that increased human activity elevates localized temperatures.

The air pollution maps exhibit greater temporal heterogeneity than spatial variation, with values increasing across the site over time. Peak concentrations were observed near the main crossroads, probably due to higher traffic volumes and emissions.

Accuracy Verification of The Maps

To verify the performance of the proposed method, an accuracy test was conducted at point S5 for thermal environment and air pollution maps. Data from the stationary validation device at S5 served as ground truth, compared with estimations from the three methods at various times. Discrepancies between estimations and ground truth were quantified (Figure 9 and Figure 10) and the RMSE was calculated (Table 1) to assess the precision across methods.

Figure 9. Comparisons between ground truth and estimations from three methods for temperature

Figure 10. Comparisons between ground truth and estimations from three methods for PM_2.5 concentrations

Table 1. RMSE of estimations for temperature and PM_2.5 concentrations from three methods

	Method1	Method2	Method3
Temperature (℃)	1.27	1.16	0.61
PM2.5 (μg/m3)	5.59	3.26	1.48

For both temperature and PM_2.5 concentrations, the estimations based on Stationary-mobile Sensing (Method3) proved closest to ground truth. Overall estimation accuracy of the three methods ranked as: Method3, Method2, then Method1, with Method3 showing optimal precision and stability.

For temperature estimations, the RMSE of Method 3 equaled 0.6℃, almost half that of the other two methods. The line chart indicates that estimates from Method 3 deviated from ground truth within merely 1℃. In contrast, estimates from the other two methods showed greater instability: Method 2 was less accurate in the afternoon, and Method 1 was less accurate in the evening.

For PM_2.5 concentration estimations, the RMSE of Method 3 was just 1.5 μg/m³, about half that of Method 2 and a quarter that of Method 1. The line chart demonstrates the estimating performance of Method 3 and Method 2 exhibited greater robustness, while there was a great deviation between estimations from Method 1 and ground truth in the evening period.

Acoustic Environment: Map Construction, Accuracy Verification and Time Window Selection

Map Construction Methods

It should be noted that noise shows considerable transient fluctuations, rendering instantaneous values less representative. In contrast, the statistical index on equivalent average noise level over a sampling period is a reliable indicator of the acoustic environment (Geraghty and O’Mahony, 2016). Thus, for acoustic environment map construction, it is more reasonable to characterize the noise level surrounding time t by using the average sound pressure level within a symmetric time window T centered on t. Increasing the time window makes the data larger and more stable, but loses the representativeness of the noisy data at localized time periods. Therefore, it is necessary to determine an appropriate T that balances localized acoustic feature capturing and overall representativeness.

Method 1: Stationary sensing with spatial interpolation

First, the mean value of the data at each known stationary sensing point Si was computed to obtain the mean value of the sound pressure level $L_{i,T,t,SS}$ within a time window T around time t. After that, using a similar IDW approach as Section 3.1.1, the average noise level at the target point p around time t is then estimated as a distance-weighted interpolation of the values at Si. The formula for the average noise level $L_{p,T,t,IDW}$ is:

  

$$\begin{array}{c}\hfill L_{p,T,t,IDW}=\sum _{}^{}\frac{\frac{L_{i,T,t,SS}}{{D_{p,i}}^2}}{\sum _{}^{}\frac{1}{{D_{p,i}}^2}},\left(4\right)\hfill \end{array}$$

where $D_{p,i}$ refers to the distance between the target point p and point Si.

Method 2: Mobile sensing data with time interpolation

When constructing the acoustic environment map solely on the mobile sensing data, data points smoothed by Gaussian Filter provided enhanced representativeness. The noise level at point p around time t was estimated as $L_{p,T,t,MS}$ , using the mean values of the mobile sensing data points within the time window T.

Method 3: Stationary-mobile Sensing

A Stationary-mobile Sensing approach for acoustic environment map construction was proposed with two main steps: First, stationary sensing helped establish baseline values of the noise levels around the target times. Second, mobile sensing was used to capture the relative spatial differentials of the noise levels between points within the same timeframes (Figure 11).

Figure 11. Schematic diagram of the general noise level $L_{t,T,SS}$ and the relative spatial differentials $d_{p,t,T}$

Specifically, in the first step, the general noise level $L_{t,T,SS}$ across the site around time t was computed as the baseline, using the mean of measurements from all the stationary devices. In the second step, the mean noise level $L_{p,t,T,MS}$ at p was taken from mobile sensing data aggregated there within the time window T. And then, the average value $L_{\overline{p,t,T,MS}}$ of all the $L_{p,t,T,MS}$ was subtracted from each $L_{p,t,T,MS}$ , in order to obtain the relative spatial differentials of the noise levels $d_{p,t,T}$ at point p:

  

$$\begin{array}{c}\hfill d_{p,t,T}=L_{p,t,T,MS}-L_{\overline{p,t,T,MS}}.\left(5\right)\hfill \end{array}$$

Finally, estimated values of the noise levels $L_{p,t,T}$ at target points p around time t were calculated as:

  

$$\begin{array}{c}\hfill L_{p,t,T}=L_{t,T,SS}+d_{p,t,T}.\left(6\right)\hfill \end{array}$$

In addition, the target estimating time t and the size of the time window T followed a certain pattern. In order to ensure the comprehensibility of the results, the size of T should be constrained to an integral multiple of T₀, the duration for one mobile collecting period, or rather the mobile sampling interval at the same point p. In this way, the time span of the aggregated mobile sensing data at each point p was equal and comparable, and this allowed the calculation of meaningful relative spatial differentials $d_{p,t,T}$ .

Estimation and Mapping Results

Following the calculating method mentioned in Section 3.2.1, the average noise level was estimated surrounding moment t_i (midpoints of the period for each mobile collecting lap), with a time window T equal to 3T₀ (T₀ refers to the duration of each lap). The estimated results were then visualized, and maps of the average noise levels around four moments are shown as examples (Figure 12). The temporal resolution equaled the lap duration T₀, roughly 15 minutes, while the spatial resolution matched the 15m path point spacing.

The noise maps demonstrate apparent spatial and temporal heterogeneity. Peak noise levels were constantly observed in the southwest area where crowds congregate, as well as along the main roads surrounding the site, owing to high human activity and traffic volumes.

Figure 12. The mapping result for the noise level

Accuracy Verification

Following the procedure in Section 3.1.3, an accuracy test for the acoustic environment maps was also carried out at S5. Similarly, the mean measurements from the validation device during each period T served as ground truth. The differences between ground truth and estimations by the three methods were compared around six target moments separately (Figure 13. Deviations between estimations from Method 3 and ground truth remained within 1.5dB, contrasting larger differences observed for the other two methods. Specifically, the accuracy and stability of Method 1 were poorer, with the divergence between its estimations and ground truth exceeding 7dB during the evening periods.

Figure 13. Comparisons between ground truth and estimation results from all the three methods

Accuracy Assessment with Different Time Window Sizes

Beyond comparing different methods under the same time window T, noise level estimation accuracy was also assessed across varying time window sizes through RMSE (Table 2). Results demonstrated that Method 3 obtains the optimal estimation precision when T=3T₀, with an RMSE of merely 0.90dB. In contrast, the highest RMSE occurred when T=T₀, reaching twice the previous value. Meanwhile, the peak performance of Method 2 was achieved when utilizing a larger time window of 5T₀.

Table 2. RMSE of the estimation results from all methods

Time period	Method3(dB)	Method2(dB)
5T0	1.15	1.82
3T0	0.90	2.40
T0	2.04	2.65

Discussion: Comparative Advantages of The Stationary-Mobile Sensing

The Superiority of The Map Construction Method with Stationary-Mobile Sensing Mode

Direct comparison of accuracy metrics (e.g., RMSE) with existing studies is not feasible due to varying site conditions. However, we conducted a comparative analysis within our study context by adapting the essence of traditional methods through our experimental design and controlled variables. Under identical environmental conditions and data sets, we developed three scenarios: Method1 eliminated stationary sensor data, representing mobile-only sensing; Method2 excluded mobile sensor data, representing stationary-only sensing; and Method3 utilized all data, representing the combined Stationary-mobile Sensing approach. Accuracy testing across thermal, outdoor air, and acoustic environment dimensions all demonstrated that the Stationary-mobile Sensing method (Method3) outperformed conventional single-mode methods. Their estimation accuracy ranks as follows: Method1, Method2, and Method3 in descending order.

Spatial interpolation in Method 1 assumes smooth variations of the indicator values over space, overlooking their random spatial distribution and localized volatility. Even with dense sampling at 120-250m intervals, this method failed to achieve satisfactory estimation results. This indicates that stationary sensing with spatial interpolation is inappropriate for high-spatial-resolution map construction.

The mobile sensing in Method 2 vastly expands the spatial coverage and sensitively captures microscale heterogeneity, improving upon the estimation of Method 1. However, disturbance from vehicles or microclimate potentially affects the sampling process, causing the measurement accuracy to deviate from the benchmarks by stationary devices and thus not reach an ideal performance.

Method 3 combines the advantages of the two sensing modes through varied procedures for thermal/air pollution versus acoustic environments. Mobile sensing data can provide comprehensive spatial coverage and capture localized heterogeneity, while stationary sensing data can capture more precise and continuous dynamic features. Therefore, this collaborative method yields superior performance compared to the other two.

The Time Window Size for Acoustic Environment Map Construction

When estimating the noise level and mapping the acoustic environment, the time window size T is crucial as it has a significant impact on the accuracy of estimation and mapping methods based on mobile sensing and Stationary-mobile sensing. However, existing studies tend to take only one window that meets the requirements and lack sufficient exploration of “how the size of the window affects the accuracy of the estimation” (Can, Dekoninck et al., 2014), which is addressed by the results of this study.

After analyzing the results, we speculate that a small T value carries the risk of inadequate mobile sensing data points gathered for assessing relative spatial deviations of the noise level. This situation can potentially result in substantial influence from intrinsic noise fluctuations, diminishing the representativeness and accuracy of the estimation results. As the T value increases, more mobile sensing data points are counted. However, an excessively large time window incorporates samples temporally distant from the target time t, which have weaker correlations with each other and can also misrepresent the acoustic conditions near the time t. Therefore, it is crucial to set a medium range of the time window T. In this study, it is found that the optimal noise level estimation accuracy was achieved with a proper time window equal to 3T0, about 45 minutes long, and containing approximately 8~10 data points for the target point p.

Conclusion

Contributions and Application Scenario

This paper is organized around three research objectives with the following findings:

First, this study demonstrated the efficacy of Stationary-mobile Sensing for simultaneous collecting and mapping of multidimensional environmental data, enhancing cost-effectiveness while improving integration across multiple environmental indicators.

Second, a novel Stationary-mobile Sensing method was developed for accurate high-resolution environmental mapping, demonstrating superior estimation accuracy compared to conventional stationary or mobile sensing modes at equivalent resolutions.

Third, tailored implementation of sensing and estimating were developed for temperature, PM_2.5, and noise respectively. Optimal time window settings were experimentally determined as well to enhance estimation accuracy, particularly for acoustic environment mapping.

The above explorations provide actionable insights for urban design and public health interventions to mitigate urban health risks: Urban planners can utilize thermal environment maps to optimize tree canopy coverage in heat-vulnerable areas, reducing heat-related mortality risks. Traffic managers may adjust signal timing or install noise barriers in high-decibel corridors identified by acoustic maps. Real-time exposure alerts based on PM2.5 dynamics could be integrated into smart city platforms, empowering citizens to avoid polluted routes.

Limitations and Future Work

Several limitations of this study need to be noted and addressed with targeted future research:

First, the study area’s spatial scale and experiment duration were constrained by limited devices and manpower. To ensure estimation accuracy, only one validation device (S5) was reserved, while others were allocated for model training, potentially affecting generalizability. Future efforts should expand monitoring to diverse land-use types (e.g., commercial / industrial zones). Moreover, deploying additional stationary/mobile devices and using unmanned patrol vehicles to carry the equipment to extend temporal coverage would enable multi-point validation and repeated trials, which are essential for integrating sensitivity analyses and confidence interval determination with RMSE values to establish a robust evaluation framework.

Secondly, the 10-second response interval restricted the data density and led to large distances between mobile sampling points. Consequently, the spatial resolution and predictive accuracy of the generated environment maps were hampered. Future studies should explore the use of more advanced sensors with shorter response intervals to further improve spatiotemporal mapping resolution.

Finally, there is room to optimize the data acquisition protocols under Stationary-mobile Sensing mode. For instance, spatial clustering can be carried out in the future according to environmental indicators distributions obtained by mobile sensing, which will guide smarter deployments of stationary devices and reduce their numbers. Exploring the trade-offs between mobile sensing speed, sampling interval for one point (time spent for one lap) and measurement fidelity can also maximize the spatial coverage of sampling with a certain degree of accuracy. Additionally, historical data and deep learning algorithms can be leveraged to improve the interpolation accuracy of sparse data, thereby enhancing the integration of mobile and stationary sensing. Such advances are expected to further improve the data collection efficiency of Stationary-mobile Sensing.

Author Contributions

Conceptualization, Qi Hao; methodology, Qi Hao; investigation, Qi Hao, Qiyuan Hong; data curation, Qi Hao; writing—original draft preparation, Qi Hao; writing—review and editing, Qiyuan Hong, Qi Hao. All authors have read and agreed to the published version of the manuscript.

Ethics Declaration

The authors declare that they have no conflicts of interest regarding the publication of the paper.

References

•  Aftab,  S. and  Moghadam , R. H. (2022). "Robust Data Smoothing Algorithms and Wavelet Filter for Denoising Sonic Log Signals". Journal of Applied Geophysics, 206, 104836. doi: https://doi.org/10.1016/j.jappgeo.2022.104836.
•  Alli , A. S.,  Clark , S. N., et al. (2023). "High-Resolution Patterns and Inequalities in Ambient Fine Particle Mass (Pm2. 5) and Black Carbon (Bc) in the Greater Accra Metropolis, Ghana". Science of the Total Environment, 875, 162582. doi: https://doi.org/10.1016/j.scitotenv.2023.162582.
•  Anjomshoaa,  A.,  Duarte,  F., et al. (2018). "City Scanner: Building and Scheduling a Mobile Sensing Platform for Smart City Services". IEEE Internet of things Journal, 5(6), 4567-4579. doi: https://doi.org/10.1109/jiot.2018.2839058.
•  Boumchich,  A.,  Picaut,  J., et al. (2024). "Blind Calibration of Environmental Acoustics Measurements Using Smartphones". Sensors, 24(4), 1255. doi: https://doi.org/10.3390/s24041255.
•  Can,  A.,  Dekoninck,  L., et al. (2014). "Measurement Network for Urban Noise Assessment: Comparison of Mobile Measurements and Spatial Interpolation Approaches". Applied acoustics, 83, 32-39. doi: https://doi.org/10.1016/j.apacoust.2014.03.012.
•  Can,  A.,  Van Renterghem,  T., et al. (2011). "Sampling Approaches to Predict Urban Street Noise Levels Using Fixed and Temporary Microphones". Journal of Environmental Monitoring, 13(10), 2710-2719. doi: https://doi.org/10.1039/c1em10292c.
•  Chambliss , S. E.,  Preble , C. V., et al. (2020). "Comparison of Mobile and Fixed-Site Black Carbon Measurements for High-Resolution Urban Pollution Mapping". Environmental Science and Technology, 54(13), 7848-7857. doi: https://doi.org/10.1021/acs.est.0c01409.
•  Clay,  R.,  Guan,  H., et al. (2016). "Urban Heat Island Traverses in the City of Adelaide, South Australia". Urban Climate, 17, 89-101. doi: https://doi.org/10.1016/j.uclim.2016.06.001.
•  Ganji,  A.,  Youssefi,  O., et al. (2023). "Design, Calibration, and Testing of a Mobile Sensor System for Air Pollution and Built Environment Data Collection: The Urban Scanner Platform". Environmental Pollution, 317, 120720. doi: https://doi.org/10.1016/j.envpol.2022.120720.
•  Geraghty,  D. and O’ Mahony,  M. (2016). "Investigating the Temporal Variability of Noise in an Urban Environment". International journal of sustainable built environment, 5(1), 34-45. doi: https://doi.org/10.1016/j.ijsbe.2016.01.002.
•  Guillaume,  G.,  Aumond,  P., et al. (2019). "Statistical Study of the Relationships between Mobile and Fixed Stations Measurements in Urban Environment". Building and Environment, 149, 404-414. doi: https://doi.org/10.1016/j.buildenv.2018.12.014.
• Hanani, N., Riniwati, H., et al. (2024). "The Impact of Land Use Change on Land Surface Temperature 1990-2020 for Environmental Monitoring in Kapota Island, Wakatobi Regency, Indonesia". International Review for Spatial Planning and Sustainable Development, 12(2), 43. doi: https://doi.org/10.14246/irspsd.12.2_43.
•  Ji, W.,  Han,  K., et al. (2023). "Trip-Based Mobile Sensor Deployment for Drive-by Sensing with Bus Fleets". Transportation Research Part C: Emerging Technologies, 157, 104404. doi: https://doi.org/10.1016/j.trc.2023.104404.
•  Kawakubo,  S.,  Arata,  S., et al. (2023). "Visualization of Urban Roadway Surface Temperature by Applying Deep Learning to Infrared Images from Mobile Measurements". Sustainable Cities and Society, 99, 104991. doi: https://doi.org/10.1016/j.scs.2023.104991.
•  Kim, Y.-J.,  Kim, B.-J., et al. (2019). "A Case Study of Environmental Characteristics on Urban Road-Surface and Air Temperatures During Heat-Wave Days in Seoul". Atmospheric and Oceanic Science Letters, 12(4), 261-269. doi: https://doi.org/10.1080/16742834.2019.1608791.
•  Kousis,  I.,  Manni,  M., et al. (2022). "Environmental Mobile Monitoring of Urban Microclimates: A Review". Renewable and Sustainable Energy Reviews, 169, 112847. doi: https://doi.org/10.1016/j.rser.2022.112847.
•  Lam , Y. F.,  Ong,  C.-W., et al. (2021). "Improvement of Community Monitoring Network Data for Urban Heat Island Investigation in Hong Kong". Urban Climate, 37, 100852. doi: https://doi.org/10.1016/j.uclim.2021.100852.
•  Li, J.,  Garshick,  E., et al. (2021). "Estimation of Ambient Pm2. 5 in Iraq and Kuwait from 2001 to 2018 Using Machine Learning and Remote Sensing". Environment International, 151, 106445. doi: https://doi.org/10.1016/j.envint.2021.106445.
•  Li, Y. and  Song,  Y. (2019). "Optimization of Vegetation Arrangement to Improve Microclimate and Thermal Comfort in an Urban Park". International Review for Spatial Planning and Sustainable Development, 7(1), 18-30. doi: https://doi.org/10.14246/irspsd.7.1_18.
•  Liu,  M.,  Chen,  H., et al. (2021). "Nonlinear Relationship between Urban Form and Street-Level Pm2. 5 and Co Based on Mobile Measurements and Gradient Boosting Decision Tree Models". Building and Environment, 205, 108265. doi: https://doi.org/10.1016/j.buildenv.2021.108265.
•  Parison,  S.,  Chaumont,  M., et al. (2023). "The Effects of Greening a Parking Lot as a Heat Mitigation Strategy on Outdoor Thermal Stress Using Fixed and Mobile Measurements: Case-Study Project “Tertiary Forest”". Sustainable Cities and Society, 98, 104818. doi: https://doi.org/10.1016/j.scs.2023.104818.
•  Pődör,  A.,  Szabó , S. J. E., et al. (2021). "Geo-Tagged Environmental Noise Measurement with Smartphones: Accuracy and Perspectives of Crowdsourced Mapping". Environment and Planning B: Urban Analytics and City Science, 48(9), 2710-2725. doi: https://doi.org/10.1177/2399808320987567.
•  Prüss-Üstün,  A.,  Wolf,  J., et al. (2016). Preventing Disease through Healthy Environments: A Global Assessment of the Burden of Disease from Environmental Risks. World Health Organization.
•  Torres , J. M.,  Nieto , P. G., et al. (2011). "Detection of Outliers in Gas Emissions from Urban Areas Using Functional Data Analysis". Journal of hazardous materials, 186(1), 144-149. doi: https://doi.org/10.1016/j.jhazmat.2010.10.091.
•  Wang,  J.,  Alli , A. S., et al. (2022). "Nitrogen Oxides (No and No2) Pollution in the Accra Metropolis: Spatiotemporal Patterns and the Role of Meteorology". Science of the Total Environment, 803, 149931. doi: https://doi.org/10.1016/j.scitotenv.2021.149931.
• World Health Organization. (2011). "Burden of Disease from Environmental Noise: Quantification of Healthy Life Years Lost in Europe". Burden of Disease from Environmental Noise: Quantification of Healthy Life Years Lost in Europe.
•  Xu, J.,  Liu,  M., et al. (2025). "A Novel Prediction Framework for Estimating High Spatial Resolution near-Ground Pm2. 5 and O3 Concentrations at Street-Level in Urban Areas". Building and Environment, 267, 112141. doi: https://doi.org/10.1016/j.buildenv.2024.112141.
•  Yang,  X.,  Peng , L. L., et al. (2020). "Impact of Urban Heat Island on Energy Demand in Buildings: Local Climate Zones in Nanjing". Applied Energy, 260, 114279. doi: https://doi.org/10.1016/j.apenergy.2019.114279.
•  Yang,  X.,  Zhou,  W., et al. (2019). "A Fast and Efficient Local Outlier Detection in Data Streams". Proceedings of Proceedings of the 2019 International Conference on Image, Video and Signal Processing, 111-116. doi: https://doi.org/10.1145/3317640.3317653.
•  Yin,  Q. and  Wang,  J. (2017). "The Association between Consecutive Days’ Heat Wave and Cardiovascular Disease Mortality in Beijing, China". BMC public health, 17, 1-9. doi: https://doi.org/10.1186/s12889-017-4129-7.
•  Zhan,  Q.,  Gao,  S., et al. (2020). "Impact Mechanism and Improvement Strategy on Urban Ventilation, Urban Heat Island and Urban Pollution Island: A Case Study in Xiangyang, China". International Review for Spatial Planning and Sustainable Development, 8(3), 68-86. doi: https://doi.org/10.14246/irspsd.8.3_68.
•  Zhang,  X.,  Zhou,  J., et al. (2021). "A Practical Reanalysis Data and Thermal Infrared Remote Sensing Data Merging (Rtm) Method for Reconstruction of a 1-Km All-Weather Land Surface Temperature". Remote Sensing of Environment, 260, 112437. doi: https://doi.org/10.1016/j.rse.2021.112437.
•  Zhang,  Y.,  Zhao,  H., et al. (2023). "Predicting Highly Dynamic Traffic Noise Using Rotating Mobile Monitoring and Machine Learning Method". Environmental research, 229, 115896. doi: https://doi.org/10.1016/j.envres.2023.115896.