Odor, air quality, and well-being: understanding the urban smellscape using crowd-sourced science

First, we read all reports manually and removed inappropriate reports from the analysis. A summary of the problematic components of such reports is available in the supplementary data file (section S2). Second, to test the assumption of independence of individual odor reports, we conducted a sensitivity test to check the validity of spatiotemporal variability of reported odors in our dataset and found that the spatiotemporal patterns presented in the manuscript remain consistent (section 2.5.4). This test also partially accounts for selection bias since the data is anonymously collected and, for example, if one user repeatedly reports to the app within the same hour. Third, to address the effect of selection bias on the identification of odor report spatiotemporal patterns, OSAC clusters, and spatial hotspots and outliers, we have conducted sensitivity analysis based on age, gender, and racialized/ethnic minority status (section 2.5.4), and results remain consistent with the key findings of the manuscript. While the results show some variability for different demographic groups, which is not surprising, they are consistent with the key findings of the manuscript in terms of the diversity of OSAC types and clusters reported, and general spatial and temporal patterns of reports. We have also conducted robustness checks for different analytical pipelines as listed in table 1. Finally, we reiterate that the analyses presented reflect patterns of odor reports, not necessarily odors themselves, and that this data is therefore meant to complement other sources of data, such as air quality monitoring.

We conducted exploratory data analysis of values/categories of each variable and the demographics within the odor reports (section S3). To quantify the spatial patterns of odor reporting, we aggregated odor counts in MetroVan at the Canadian census tract (CT) level [68]. We used these odor data to calculate spatial metrics such as Global Moran's I, Local Moran's I, Getis Ord I, and Getis Ord Gi* [69–71]. These metrics allow us to map hotspots and coldspots and spatial clusters and outliers based on the Local Moran's I (and the Getis Ord Gi*) metric. As a robustness check, we also analysed area-normalized and population-normalized odor counts, to account for area-based bias (larger areas are expected to have more odor sources) and population-based bias (larger populations are expected to report more odors). Due to consistency in the key findings and for brevity, we only present spatial analysis results based on the Local Moran's I metric for population-normalized odor-counts in the main manuscript. Additional details on the methods and results of spatial analysis are available in the supplementary data file (section S5, supplementary figures S3(a)–(j)).

We also report temporal patterns of ORCs and OSAC categories. To investigate the possible relationship between ORCs (dependent) and meteorology, air quality, and odor-related and app-related counts of news stories (independent), we conducted exploratory multiple linear regression (MLR) modeling of daily-averaged ORC at the regional scale. This analysis was carried out for separate months within the study period. Several quality assurance (QA) and quality control (QC) steps were employed to identify key variables for conducting MLR. Finally, we estimated the relative importance of the independent variables based on the fraction of variance explained in the linear model [72]. We conducted bootstrap analysis to quantify 95% confidence intervals for variance explained, which are reported in brackets with its observed value. Additional details on the meteorological variables, air quality monitoring indicators, news reports, QA/QC, and uncertainty analysis are included in the supplementary data file (section S6, tables S6–S9).

We conducted thematic (free text) analysis [73, 74] on the free text associated with OSAC. Briefly, we coded descriptive text for odors reported, inductively generating high-level odor categories. These high-level categories account simultaneously for the drop-down fields and the free text descriptions of OSAC. This practice of characterising odor perception using reference vocabulary has been widely employed for drinking water, wastewater and compost, urban odors, and even wines [75–78]. The categories for symptoms and actions were refined based on a review of the public health literature on odor [29]. Odors and causes were categorised based on local knowledge of important odors and odor sources in the region [51, 79]. A detailed list of the categories and the related descriptors is available in the supplementary data file (supplementary table S5). For this categorical data (e.g. odor categories), we conducted textual pairwise correlation analysis, presented in terms of the correlation coefficient, phi [80, 81] and hierarchical (divisive) clustering analysis for trinary and quarternary associations [82, 83]. We also conducted bootstrap analysis to quantify 95% confidence intervals for this coefficient phi, which are reported in brackets with its observed value. To better understand the large symptom category of emotional and mood disturbance, we conducted sentiment and emotion analysis using three sentiment scales: the Finn årup Nielsen (AFINN), the National Research Council Canada (NRC) Emotion lexicon, and the NRC Valence, Arousal, and Dominance (VAD) lexicons [84–86]. Finally, to better understand the relationships of OSAC categories, we represent them visually as a 2D network of vertices (OSAC categories) and linear edges (binary relationships) based on the Kamada-Kawai algorithm [87]. Details of the bootstrapping, sentiment analysis, and textual associations and visualizations conducted for OSAC are in the supplementary data file (section S4).

To test the assumption of independence of individual odor reports, we checked the spatiotemporal variability of reported odors in our dataset. Specifically, we generated a sensitivity test data subset by only keeping the first report with a reported odor in a particular census tract within a particular hour and excluding subsequent reports. Then, we compared the number of reported odors in the original dataset and the sensitivity test dataset to assess the impact of this assumption on our analysis. Our analysis revealed that the original dataset contained 760 combinations of tract, hour, and reported odor, while the sensitivity test dataset had 732 combinations. This difference of less than 5% of reported odors indicates that the impact of the assumption of independence of odor reporting in the app is mostly limited. To further validate our findings, we replicated the main figures in the manuscript using the sensitivity test dataset (section S7, supplementary figures S4–S6). We found that the figures and the underlying results and conclusions were consistent with those obtained from the original dataset, suggesting that the assumption of independence of odor reporting does not significantly affect our analysis. We have also conducted similar sensitivity tests for 6 and 12 h and have found similarly consistent patterns (section S7, supplementary figures S7–S12). However, we acknowledge that the assumption of independence of reports may not hold true in certain situations [39], such as when there are multiple reports of an odor from the same individual across multiple hours/locations or when there is a systematic bias in the reporting behavior of participants. Additionally, to address the effect of selection bias on the identification of odor report patterns, OSAC clusters, and spatial hotspots and outliers, we have conducted sensitivity analysis based on age, gender, and racialized/ethnic minority status (section S8). While the results show some variability for different demographic groups, which is not surprising, they are consistent with the key findings of the manuscript in terms of the diversity of OSAC types and clusters reported, and general spatial and temporal patterns of reports (figures 1–3, supplementary figures S13–S30).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The descriptions of smells experienced demonstrate the different types and perceptions of odors encountered (supplementary table S10). Users often use rich and evocative language to describe their odor experiences in free text responses. For instance, one user writes, 'rotting waste, garbage cheese, pungent vinegar death, fresh vomit.' 'Rotten' and 'chemical' account for about 65% of submissions with a reported odor (supplementary table S11). Burning is the third most important odor category, accounting for 16% of the reports. While odors can co-occur (supplementary figure S31(a)), the binary associations of such co-occurrences are weak (phi < 0.25). In fact, some odors statistically significantly do not co-occur (e.g. ROT and CHM; $phi = -0.47[-0.54, -0.40]$, ROT and BRN; $phi = -0.38[-0.45, -0.30]$).

Users identify many potential odor sources, including: garbage and compost (29%); chemicals (16%); fire, smoke, and burning (14%); sewage and wastewater treatment (9%); and animal processing (9%) (supplementary table S12). Users mostly report single causes, accounting for about 80% of all reports (supplementary figure S31(b)). Only two pairs of cannabis facilities and smoking ($phi = 0.44[0.24, 0.61]$) and cannabis facilities and farming ($phi = 0.25[0.14, 0.50]$) show relatively strong binary associations.

App users report experiencing several classes of symptoms, such as neurological (e.g. dizziness, headache), respiratory irritation (e.g. cough, difficulty breathing), emotional and mood disturbance (e.g. anxiety, frustration, anger), ophthalmological (e.g. irritated eyes), and dermatological (e.g. hives). Neurological, respiratory symptoms, and emotional and mood disturbance occur most frequently, accounting for 87% of the symptoms reported (supplementary table S13). We observe that while prominent symptoms often co-occur (supplementary figure S31(c)); only two co-occurrence of symptoms is statistically significant—respiratory irritation with ophthalmological symptoms ($phi = 0.30[0.21, 0.38]$) and respiratory irritation with emotional and mood disturbance ($phi = 0.29[0.20, 0.38]$).

Emotional and mood disturbance accounts for a substantial fraction (23%) of reported symptoms, pointing to the negative moods induced by unpleasant odors [33]. Analysis of semantic descriptors of symptoms shows the expression of a wide range of emotions, but particularly a large number of negative sentiments and usage of words expressing displeasure (NRC VAD lexicon; 'difficulty', n = 132; 'sore', n = 80) and arousal (NRC VAD lexicon; 'disturbed', ORC = 132; 'irritated', ORC = 79) as well as more specific emotions such as anger, sadness, disgust, and fear, all of which occur at least 250 times (supplementary figure S32). Likewise, using the AFINN lexicon, we observe sentiment scores ranging from about +1 to as low as −10, suggesting a strong bias towards negative sentiments (supplementary figure S33). We document quotes from odor reports rated with an AFINN sentiment score of −7.5 and lower in supplementary table S14. Users employ evocative phrases about symptoms ('Disgust, annoyance, anger, concern about carcinogens and family health'), causes ('uncontrolled—not monitored—disregard to permit'), and broader societal effects ('Listened to my wife scream about ongoing problem as home values go down') as they express their negative sentiments, often identifying details of the odor issues (e.g. going on for 'over 12 years'). Finally, users also employed the free text to express non-verbal cues such as emotional accentuation [88, 89] through the use of capital letters [90–92], and we observe multiple such descriptions ('CLOSE WINDOWS—VERY MAD—TIME FOR ACTION—GOVERNMENTS NEED TO COME SEE THIS ON GOING PROBLEM'). We also find that the reported odor strength and offensiveness are positively associated, consistent with literature (section S9, supplementary figure S34).

Users report a range of actions in response to odors (supplementary table S15). Ventilation and air cleaning (43%), gone inside (26%), making a complaint (15%), and stopped exercising outdoors (10%) are the most reported actions, with other actions, such as smell-masking (e.g. adding a pleasant fragrance) accounting for less than 5% each. Users also report long-term avoidance of odorous areas (1% of the reported actions), such as 'moved away' or going 'to a distant part of the city to go for a walk'—significant life changes to avoid odors. Similar observations of the desire to relocate due to the impacts of malodor have been reported elsewhere as well [93]. We observe 'gone inside' co-occurring repeatedly with other actions (supplementary figure S31(d)); likewise, we observe strong associations ('Gone inside' with 'Stopped exercising outdoors': $phi = 0.35[0.28, 0.42]$). Analysis of semantic descriptors for actions also shows the expression of a wide range of emotions. The sentiments are largely negative, dominated by anger (supplementary figure S35).

In a few reports, we also observe instances of maladaptation, where actions taken by users to avoid odors also negatively affect their well-being due to exposure to other environmental stressors or curtailment of healthy behaviors (supplementary table S16). Users report temporarily stopping or changing breathing patterns (e.g. 'breathe through mouth', 'removed my mask temporarily to air it out'), using smell masking (e.g. 'put Vick's in my nose', 'deoderize the house'), as well as changing local ventilation ('turned off car ventilation', 'have to close all windows on summer's evening'), speeding ('ran home'), and inability to exercise or enjoy outdoors ('Very disturbing to to family ANGER and unable to enjoy outdoors'). Smell masking agents can themselves have substantial health effects [94, 95], hence our inclusion of it as a maladaptive behavior.

We observe hundreds of groupings of odors, symptoms, actions, and possible causes, even when only three of the four parameters are analysed at a time (supplementary table S17). Our clustering analysis reveals OSAC associations that show odor- and cause- connections to actions and symptoms (figure 2). In figure 2, OSAC categories in a higher number of combinations are placed closer together and the strength of binary associations are represented by the thickness of the edge connecting them with other OSAC categories [87] (section S4). For example, we find that suspected causes (fires, smoke, and burning) are often linked to specific symptoms (respiratory irritation) through the experience of specific odors (burning) (figure 2). However, suspected causes (garbage and compost) are also directly linked to symptoms (neurological) and actions (making a complaint) without odor mediation.

Here we briefly report on spatial and temporal patterns in ORC and OSAC.

• 1.  
  Spatial distribution of odor reports: Four municipalities (City of Vancouver, Delta, Burnaby, and Richmond) account for 90% of ORC (supplementary table S18). These municipalities show substantial differences in reported OSAC (figure 1(a), supplementary tables S18–S22) and OSAC associations (supplementary figures S36(a)–(e)).
• 2.  
  Urban versus suburban odor sources: We observe distinct differences in odor sources and types between urban and suburban areas (section S10). For example, the City of Vancouver reports a disproportionately large number of reports with the suspected cause of animal processing (95%), while suburbs like Delta report disproportionately more on garbage and compost (58%), cannabis (73%), and farming (60%) causes (supplementary table S20).
• 3.  
  Distinct patterns in OSAC connections: We also observe differences in OSAC connections (section S10). For example, in Vancouver, we observe notable connections in the City with regards to burning odors, which are related to the causes of smoking and fires, smoke, and burning, and also the symptoms of respiratory irritation (supplementary figure S36(a)). The symptom of respiratory irritation itself is related to neurological symptoms and emotional and mood disturbance. There are other notable associations between specific odors and their suspected causes. On the other hand, complex connections in Delta are that of rotten odor and the possible cause of garbage and compost linked to multiple symptoms and actions (supplementary figure S36(b)).
• 4.  
  Temporal variability in odor reporting: Temporal patterns of regional ORC exhibit variability within days and between months (figures 1, supplementary figures S37, S38(a)–(c)). We see spikes in reporting during 8–11 December (35), 28 June–4 July (41), 7–15 July (48), 20 July–1 August (40), 4 August–7 August (33), and October 3 (22) (supplementary figure S38(a)). These spikes together account for about 40% of the reports and are likely related to specific events such as the launch of the app and related news coverage, odor accidents, and extreme weather conditions. These drivers are discussed in sections 3.2 and 4.2. Odor report counts show significant variability by hour of day and across months (supplementary figures S37, S38(b) and (c)). Variation in reported OSAC categories across months is broadly similar to the temporal patterns of all reports, and we observe a similar consistency in terms of the diversity of OSAC connections (figure 2, supplementary figures S37, S39(a)–(d), S40(a)–(l)).

Through Local Moran's I analysis of population-normalized ORC, we find the presence of statistically significant spatial clusters of reports. Figures 3(a) and (b) show four types of spatial distributions of ORC in the region—hotspots: areas with high ORC surrounded by areas with high ORC, high outliers: areas with high ORC surrounded by areas with low ORC, low outliers: areas with low ORC surrounded by areas with high ORC, coldspots: areas with low ORC surrounded by areas with low ORC. We observe that the City of Vancouver and Delta account for a large number of tracts in regional odor hotspots (figure 3(a)).

Hotspots dominate large parts of the City of Vancouver (supplementary table S23, Hotspots and Vancouver: $phi = 0.65[0.56, 0.73]$); however, not all parts of the city are reported to be equally odorous, and there are also low outlier neighbourhoods (figure 3(a); supplementary table S23, Low outliers and Vancouver: $phi = 0.51[0.40, 0.60]$). Further, neighborhoods bordering the inlet in the northeast have a dense mix of industrial and residential zoning, and are hotspots, within the region and within the city (figures 3(a) and (b)).

In contrast to the City of Vancouver, we find more homogeneity in neighboring jurisdictions (figure 3(a)). Most census tracts in Delta emerge as regional odor hotspots. Large parts of the Surrey township are odor coldspots (supplementary table S23, Coldspots and Surrey: $phi = 0.82[0.63, 1.0]$). Future work can test these preliminary spatial connections by conducting proximity and dispersion modeling analysis for odor report locations.

Exploratory MLR modelling suggests that criteria air pollutants and meteorology fail to capture most of the variance in daily ORC (tables 2, S6, S9). The most important air pollutants associated with ORC are PM (3%[1%–4%]), NO and NO₂ (2%[1%–4%]). But, other factors such as accidents (11%[2%–14%]) and prominent events (Launch of SmellVan app, national network coverage, (6%[3%–8%]) and solar radiation and heat fluxes (8%[3%–15%]) may have a stronger influence. However, the relative importance of different variables varies substantially across different months. For example, the contributions of wind speed to explained variance in linear models for ORC ranges from 2%–55% in different months, and we observe mixed behaviors with regards to the association (positive and negative correlation slopes).

Table 2. Key MLR variables explaining variance in daily ORC. Arrows show directionality of relationships—up (blue) is positive, up-down (yellow) refers to split (positive and negative), and down (red) indicates negative. Values inside brackets under absolute variance explained show 95% confidence intervals.

Model month	December	January	February	March	April	May	June	July	August	September	October	November	Avg.	Wt. Avg.	Variance explained
Variance explained: Absolute (A)/Relative (R)?	R	R	R	R	R	R	R	R	R	R	R	R	R	R	A
Odor-related incidents			98 ${\color{blue}{\uparrow}}$	18 ${\color{blue}{\uparrow}}$					65 ${\color{blue}{\uparrow}}$				15	15	11 (2, 14)
Solar radiation and heat fluxes	4 ${\color{red}{\downarrow}}$	39 ${\color{red}{\downarrow}}$				51 ${\color{blue}{\uparrow}}$	47 ${\color{orange}{\updownarrow}}$	33 ${\color{red}{\downarrow}}$		47 ${\color{blue}{\uparrow}}$		31 ${\color{red}{\downarrow}}$	21	19	8 (3, 15)
SmellVan news	58 ${\color{blue}{\uparrow}}$												5	9	6 (3, 8)
Rainfall								67 ${\color{blue}{\uparrow}}$					6	14	5 (2, 12)
Wind speed	8 ${\color{red}{\downarrow}}$		2 ${\color{blue}{\uparrow}}$				30 ${\color{orange}{\updownarrow}}$		12 ${\color{red}{\downarrow}}$	33 ${\color{blue}{\uparrow}}$	55 ${\color{red}{\downarrow}}$		12	12	5 (2, 8)
Pressure	5 ${\color{blue}{\uparrow}}$			28 ${\color{red}{\downarrow}}$			22 ${\color{red}{\downarrow}}$		6 ${\color{red}{\downarrow}}$			69 ${\color{blue}{\uparrow}}$	11	8	4 (2, 6)
Temperature, Humidity, DPT	25 ${\color{blue}{\uparrow}}$			22 ${\color{red}{\downarrow}}$		49 ${\color{blue}{\uparrow}}$							8	7	4 (2, 6)
PM		61 ${\color{blue}{\uparrow}}$		17 ${\color{blue}{\uparrow}}$					15 ${\color{blue}{\uparrow}}$				8	5	3 (1, 4)
NO and NO2				16 ${\color{red}{\downarrow}}$	41 ${\color{red}{\downarrow}}$				2 ${\color{red}{\downarrow}}$	20 ${\color{blue}{\uparrow}}$	45 ${\color{red}{\downarrow}}$		10	8	2 (1, 4)
CO					59 ${\color{blue}{\uparrow}}$								5	3	1 (0, 1)
Adjusted R2	0.63	0.28	0.86	0.48	0.14	0.23	0.58	0.35	0.66	0.26	0.08	0.16	0.39	0.43
Total variance explained	0.70	0.32	0.87	0.57	0.20	0.28	0.70	0.39	0.75	0.33	0.14	0.22	0.46	0.50
Number of reports	83	19	23	39	24	23	44	113	77	32	50	22

Traditional research on air quality has often neglected the implications of odors with regards to health, well-being, and quality of life, considering them only as a nuisance. Our study addresses this research gap by presenting a new crowd-sourced science framework STOSAC and implementing it in a community science project called Smell Vancouver (SmellVan). This study extensively connects odors with their HWQOL impacts, including behavioral responses to odors, and perceived odor causes, offering a comprehensive view that highlights the limitations of existing odor-monitoring approaches such as FIDOL and CICOP, which fail to capture the nuanced experiences of individuals.

The scope of the article covers both qualitative and quantitative analyses conducted on 12 months of SmellVan odor reports. The use of free text in STOSAC allows app users to communicate their odor experiences and provide information (such as sentiments and emotions) that complements quantitative data collected in the odor reports. Our users report multiple pathways through which odors impact HWQOL, encompassing the physical environment, social support, coping skills, and healthy behaviors. We also observe the interlinking of these OSAC patterns and report multiple complete cases of OSAC associations (supplementary table S29). The identification of behavioral responses (symptoms and actions) and their interlinking to odors and perception of odor causes suggests a substantial drawback of traditional odor documentation approaches such as FIDOL (Frequency, Intensity, Duration, Offensiveness, and Location) and CICOP (Concentration, Intensity, Character, Offensiveness, and Persistence) as well as odor nuisance indices based on these approaches [102], which do not account for such subjective experiences (e.g. relationships between odors and symptoms). The large spatiotemporal coverage of the approach is in much contrast compared to traditional odor monitoring approaches such as dynamic olfactometry [78, 102, 103] that can only measure odor concentrations at a specific location and time. Thus, STOSAC makes it easier to obtain a comprehensive understanding of the spatial and temporal variability of odors and corresponding behavioral responses compared to traditional approaches.

Despite limited participation rates in this study, we document 100s of unique combinations of census tract, hour of reporting, and reported odors. Typically, instrument-based deployments such as PTR-MS in traditional regulatory approaches would at best capture odorous chemicals in a single tract for a few months, and even these deployments are quite nascent in their development [104]. Additionally, given that regulatory monitoring with its hourly or coarser measurements often does not capture concentration peaks or odor exposure that occur at shorter time scales [34–37], STOSAC-based odor reporting could serve as a marker of such pollution exposures and their HWQOL impacts.

In summary, the range of odor experiences, and complex links between odor and HWQOL impacts, documented in SmellVan indicate that more nuanced approaches to odor management may be required to support community HWQOL than fixed separation distances from odorous sources [105]. Community feedback in the OSAC framework could be used as a starting point to design policy actions that prioritize specific odor sources (e.g. facility-specific improvements) and address their co-occurring symptoms (e.g. targeted resident health monitoring and care), and this process of OSAC-based data collection can be conducted on a recurring basis at low cost and at large spatiotemporal scales. The STOSAC approach, which incorporates the behavioral response of humans to environmental exposures, points to new opportunities for community science to inform policy and planning decisions.

We identify potential use cases of crowd-sourced odor report data as a complement to traditional air quality, odor, and health impact data sources:

• 1.  
  Real-time spatiotemporal data on emergent odor and air pollution events: Crowd-sourced data encourages spontaneous reporting from users, which may be particularly useful for the rapid identification of emergent hotspots or real-time tracking of transient odor and air pollution events affecting communities (e.g. industrial facility accidents). In these cases, crowd-sourced spatiotemporal data on odor impact zones and odor experiences can provide unique insight into evolving conditions for environmental and public health authorities, and concerned members of the public—including where it may be useful to systematically collect odor experience data, conduct air quality monitoring, and recommend protective measures.
• 2.  
  Enhancing community right-to-know: In many jurisdictions, community members do not have access to other odor complaints made, without going through onerous data request steps. The utility of publicly accessible crowd-sourced data therefore extends beyond data collection. By addressing data transparency issues, it links to the community's right-to-know about conditions that might pose risks to their health and the environment [106, 107] and fosters a shared sense of responsibility and community engagement among citizens. These aspects are vital in environmental health, as they empower individuals to actively participate in identifying and responding to potential health hazards in their environment [108, 109].
• 3.  
  Exploratory understanding of health, well-being, and quality of life impacts: Self-reported data on odor and odor impacts could support more comprehensive investigation of HWQOL, and complement data traditionally used by health authorities, such as administrative health data (e.g. emergency room admissions). Many of the HWQOL effects reported in Smell Vancouver data were subclinical in nature, and therefore cannot be captured in administrative data sources as individuals may not seek care. Although crowd-sourced self-reported HWQOL impact data has many limitations, knowledge of these subclinical symptoms and experiences may nevertheless have value as early indicators, or provide insight into potentially vulnerable populations [110, 111]. Further, data on OSAC linkages such as the odor-symptom relationships identified in this study can help generate hypotheses about the mechanisms of smell impacts, which could then be further assessed using methods with less selection bias such as controlled trials, or epidemiological research. Finally, future work could utilize the STOSAC approach to build policy-relevant tools and metrics based on OSAC data.

We highlight one illustrative example of these use cases: in early 2024, an incident at a Vancouver refinery involving restart of processing operations after a cold weather related shutdown was accompanied with 'an unplanned release of emissions, experienced across parts of the region as a foul odour' [112]. In response, the Smell Vancouver app received hundreds of reports well ahead of official notifications [113] and air quality statements, and was later used in media reporting [114]. This incident underscores the value of crowd-sourced data in providing a real-time, publicly accessible platform for odor reporting, which can be especially critical in emergency situations, as well as community right-to-know.