Harnessing Geospatial Data for Urban Climate Resilience: Insights from a Fine Scale Ambient Temperature Analysis in an Urban Heat Island

Abstract

Background:

Cities face increasing risks from the urban heat island effect. Our ability to assess fine-scale differentials in urban heat is limited to the sparse spatial density of temperature monitoring. This study aims to assess ambient temperature variations at high resolution within an urban heat island in the northeastern United States, focusing on spatial disparities in heat exposure and implications for local climate planning and policy.

Methods:

The authors carried out a community-based research project in which 60–80 ambient temperature sensors were deployed in the city of Chelsea, Massachusetts, from 2015 to 2023 and compared to the National Weather Service temperature data from Logan International Airport. Data were analyzed for warm seasons, hot weeks, and heat waves. Warm season, yearly, and day–night ambient heat was analyzed and mapped using geospatial regression and kriging, incorporating natural/built environment variables.

Results:

Local sensors were up to 10°F (5.6°C) higher than National Weather Service readings during hot weeks with heat waves. Within Chelsea, spatial analyses identified approximately 5°F or 2.8° C (average) and 9°F or 5.0°C (maximum) higher temperatures in hot spots compared to cooler areas. Neighbourhoods with higher temperatures were in the more residential, urbanized areas of Chelsea. Day–night mapping further highlighted areas with prolonged heat exposure, crucial for health implications.

Conclusion:

This study highlights the value of local fine spatial ambient temperature data in urban climate planning. Geospatial modelling of outputs can assist policymakers in developing climate interventions. The research supports targeted efforts to mitigate urban heat, underscoring the importance of integrating environmental data with local insights for effective and equitable heat management strategies.

Background

Between 2030 and 2050, global climate change is predicted to cause an additional 250,000 deaths per year (World Health Organization, 2023). We can expect that a portion of these deaths will be the result of an increase in the frequency, intensity, and duration of heat waves (World Health Organization, 2023). In addition to death, warmer temperatures and extreme heat events contribute to a variety of negative health outcomes including heat exhaustion, heat stroke, and the exacerbation of renal, respiratory, and cardiovascular diseases (World Health Organization, 2018). People living in cities have a greater risk of heat-related health impacts, in part due to the urban heat island (UHI) effect, whereby urban areas experience higher temperatures than surrounding rural areas (Hsu et al., 2021; Peng et al., 2012). Characteristics of the urban built environment contribute to the UHI effect, including increased impervious surface area, lack of green space, high building density, and concentrated fossil fuel use (Peng et al., 2012; Zhao et al., 2021).

Within U.S. cities, heat exposure and associated health outcomes are not equitably distributed: Lower income and/or residents of colour face higher UHI intensity than more affluent white residents (Hoffman et al., 2020; Hsu et al., 2021). This disproportionate burden is rooted in racist historical disinvestment practices including redlining, the practice of denying home loans/insurance based on area racial composition, which resulted in more densely populated neighbourhoods of tenant-occupied homes with more impervious surfaces and less greenspace (Hoffman et al., 2020). Across major U.S. cities, land surface temperatures in redlined areas were reported as much as 12.6°F (7.0°C) higher than surrounding non-redlined areas (Hoffman et al., 2020). Redlined and, more generally, environmental justice (EJ) communities disproportionately face extreme heat exposure across U.S. cities (Chakraborty et al., 2019; Hsu et al., 2021). These studies rely on land surface temperature (LST) as a metric for temperature. However, studies have shown that LST measurements may not fully represent human-relevant exposure during extreme heat events (X. Li et al., 2023; White-Newsome et al., 2013). From a public health preparedness perspective, successful local-level cooling interventions require locally tailored data and expertise (Hondula et al., 2015).

According to the latest National Climate Assessment, the United States will face increased extreme heat events and related mortality within major urban areas, disproportionately impacting minority populations (Crimmins et al., 2023). However, most cities lack information to characterize local temperature differences and inform their climate adaptation planning. Temperature data can be collected through on-site monitoring, statistical modelling, or a combination of both. On-site monitoring can utilize low-cost or regulatory monitors: Low-cost temperature monitors, such as Honest Observers by Onset (HOBOs), can be installed at the neighbourhood scale but may require additional data processing to approximate a regulatory grade instrument (Durre et al., 2010; Smoliak et al., 2015), while regulatory monitors record temperature at the city or regional scale with high-grade quality control and assurance (National Oceanic Atmospheric Administration [NOAA], 2024). Satellite data, commonly processed through statistical modelling, are also a well established source used in the scientific literature for describing and analyzing temperature in various settings (Z.-L. Li et al., 2023; Phan & Kappas, 2018). Importantly, regulatory and satellite data are generally lower in spatial resolution, which can mask local temperature trends (Chakraborty et al., 2020; Clement et al., 2023; Harlan et al., 2006; Harrison et al., 2022; Hsu et al., 2021; O’Brien et al., 2023). At the local level, such as within a city, low-cost sensor data may collect and enable analyses of more granular temperature patterns relevant to neighbourhood-level intervention planning.

A need remains for data collection that better characterizes local-level ambient temperature, including community-engaged approaches to facilitate the use of data for locally appropriate and sustainable interventions. Although temperature monitoring efforts have quantified UHI presence and intensity across the United States, data sources such as weather station sensors, satellites, and modelling are limited in their ability to capture variation at the local level for health and policy-relevant applications (White-Newsome et al., 2013). A study in Florida found that local sensors measured maximum ambient temperatures and heat index values that were significantly higher than those from the corresponding NOAA National Weather Service (NWS) airport sensor by an average of 6°F (3.3°C) and 11°F (6.1°C), respectively (Clement et al., 2023). These local measurements were above the regional heat advisory threshold, suggesting a need for locally tailored heat advisory standards (Clement et al., 2023). Mobile ambient temperature monitoring efforts such as NOAA’s UHI mapping campaigns identify local heat islands but only capture data for one or two days, missing key spatial and temporal variability (National Integrated Heat Health Information System, 2023). Satellite temperature data are temporally and spatially resolved but capture land surface temperature, which does not fully or accurately describe human-relevant temperature variability in urban thermal environments (Elmes et al., 2020; Naserikia et al., 2023; Nazarian et al., 2022; Vanos et al., 2020). Additionally, local temperature variation can be missed when using satellite imagery or modelling at larger spatial resolutions (Chakraborty et al., 2020; Clement et al., 2023; Harrison et al., 2022; Hsu et al., 2021; O’Brien et al., 2023).

Some studies in U.S. cities have conducted geospatial heat exposure analyses at the local level, integrating urban structure and land use variables to better represent temperature variation where on-the-ground sensors are absent (Galletti et al., 2019; Hart & Sailor, 2009; Huang et al., 2011; Oswald et al., 2012; Shandas et al., 2019; Tsin et al., 2020; Venter et al., 2020; Zhou et al., 2014). Many of these researchers used Geographic Information Systems (GIS) technology to map heat exposure within cities, commonly integrating variables such as percent impervious surface, distance to water bodies, and distance to or presence of green space or developed land into final regression models to examine the influence of land use variables on temperature in cities (Hart & Sailor, 2009; Oswald et al., 2012; Shandas et al., 2019; Tsin et al., 2020; Venter et al., 2020; Zhou et al., 2014). Researchers in Boston attempted to better understand land surface temperature variation during heat advisories within census tracts using satellite imagery, geospatial street segment data, and natural/built environment variables (O’Brien et al., 2023). Results showed significant variations in LST between streets within census tracts due to differences in tree canopy cover, impervious surfaces, and albedo (O’Brien et al., 2023). Similar research has tied these natural/built environment variables to heat-related health outcomes: The increased capacity of impervious surfaces to store heat and related persistent nighttime heat exposure during sleep is associated with morbidity and mortality (Kovats & Hajat, 2008; Murage et al., 2017). Overall, however, when characterizing local variation in heat, specific methods, spatial resolution of data, and included natural/built environment variables are inconsistent across studies and may not capture information relevant to climate planning intervention. Research integrating ambient temperature data at the local level is needed to capture extreme heat exposure more holistically.

Tailoring extreme heat research to the populations most impacted is critical to ensuring EJ in policy, planning, and interventions (Van Horne et al., 2023). This emphasis on aligning local data with community-driven climate change policy can be clearly illustrated in the City of Chelsea, Massachusetts, USA. Specifically, amplifying community-member experiences with extreme heat can help tailor appropriate and sustainable climate interventions (McIntyre et al., 2022; Milando et al., 2022).

The current analysis is part of the Chelsea and East Boston Heat Study (C-HEAT), a research and action partnership between investigators at Boston University School of Public Health, in Boston, Massachusetts, and GreenRoots, Inc., a grassroots EJ organization in Chelsea, Massachusetts. C-HEAT operates under an advisory board composed of community leaders and members of city government from both Chelsea and neighbouring Boston, and regional environmental consulting/non-profit organizations. Throughout the study’s conceptualization, we communicated with advisory board members and study partners to discuss C-HEAT sensor placement. Through this multidisciplinary and multisectoral study, we also developed a relationship with the Commonwealth of Massachusetts Department of Conservation and Recreation (DCR) to aid in data collection, preparation, and research. In this study, we aimed to analyze within-city ambient temperature variation using a network of 60 ambient temperature sensors and geospatial natural/built environmental data to inform climate planning. This work is the first city-wide, multiyear temperature monitoring effort completed in Chelsea to date. Our novel, community-relevant results add to the existing literature characterizing extreme heat exposure in vulnerable communities and act as an applicable resource for researchers and policymakers.

Methods

Location

Chelsea is located in the northeastern United States and includes approximately 40,000 residents, with 67% identifying as Hispanic or Latinx (United States Census Bureau, 2019). Approximately 45% of Chelsea residents identify as foreign-born. A significant percentage of residents live below the federal poverty level; 24% of Chelsea’s residents compared to 10.5% in the Commonwealth of Massachusetts overall (United States Census Bureau, 2019). These characteristics meet the demographic criteria used by the Massachusetts Executive Office of Energy and Environmental Affairs (2023) to establish census tracts as EJ communities. The city of Chelsea is actively involved in addressing environmental health-related burdens, including its designation as a UHI (Massachusetts Executive Office of Energy and Environmental Affairs, 2023). Higher temperatures and increased exposure to heat by Chelsea residents can be attributed to factors including historical disinvestment in urban planning and urban greening, exclusion from environmental health decision-making, old housing stock, high tenant-occupied multi-family buildings, high residential instability, and/or economic pressures (Healthy Chelsea, 2023; Healy et al., 2022).

Data

Temperature measurements

We recorded ambient temperature data using HOBO air temperature sensors (ONSET, 2021). HOBO sensors record air temperature between temperatures 0.04°C (0.072°F), and are portable, stable, and relatively low-cost devices with large memory capacity for longer term data logging. Using integrated sensors and ultraviolet radiation shield protection, HOBOs can record outdoor temperature measurements every 10 seconds for eventual download and analysis via an online and mobile application (ONSET, 2021). Every HOBO sensor is serviced yearly (battery replacement, calibration), and the recorded temperature data are downloaded and stored in a data repository.

HOBO data were available from two sources: DCR and C-HEAT. DCR sensors included 60 HOBOs calibrated and deployed in January 2015. Following established protocols, we conducted quality-assurance checks to the downloaded HOBO temperature data: (a) flagging observations that fell outside of all-time temperature record range for Massachusetts (−40°F to 107°F or −40°C to 41.7°C), (b) flagging observations that were more than six standard deviations from the network mean during a particular time, and (c) flagging physically implausible temperature variability (observations that change ±39.2°F/±21.8°C compared to preceding or subsequent observation; Durre et al., 2010; Smoliak et al., 2015). No data were flagged based on checks 2 or 3, and <1% of data were flagged and removed due to physical implausibility.

C-HEAT HOBO sensors were deployed seasonally, from late May through September 2020–2023, and ranged in number from 12 to 20 depending on C-HEAT study objectives each season (Milando et al., 2022). In all seasons, we focused on characterizing temperatures in areas of the city where people live, rather than in areas zoned for industrial uses. We downloaded hourly ambient temperature data from the NWS sensor located at Logan International Airport (approximately 3.2 km, or 2 mi, from Chelsea) for comparison. We restricted our analysis time period to the warm season, defined as June 1–September 30, for all years of study data. A map of sensor locations from 2015 to 2023 is shown in Figure 1.

Figure 1:

Map of Chelsea, Massachusetts, showing the location of local ambient temperature sensors, National Weather Station sensor, and zoning designations (industrial, commercial, and residential land)

Natural/built environment variables

Geospatial natural/built environment variables previously associated with higher temperatures were included in our analyses (Arnfield, 2003). Table 1 outlines each variable and data description, source, year, and spatial resolution. Parcel data were obtained from MassGIS in the form of polygon areas. For variables at grid resolution (e.g., the Normalized Difference Vegetation Index [NDVI] and LST), grid centroids were spatially joined to parcels and average NDVI and LST values were calculated. Water bodies were downloaded from MassGIS and included rivers and oceans. Distance in metres was calculated from the edge of the parcel to the nearest water body. Major roads included highways and major arterials/collectors; distance was calculated from the edge of each parcel in metres. Population information (count) was downloaded from the 2020 US Census. All variables were aggregated to parcel (property tax parcel land lot) level (see Figure A2 in the Appendix) and spatially joined to a single data set for analysis with ambient temperature data. Figure A1 in the Appendix shows maps of Table 1 variables.

Table 1:

Natural/built environment geospatial variables

Variable	Description	Source	Year(s)	Spatial resolution
Massachusetts Parcels	Property tax parcel land lot polygons	MassGIS	2018	Parcel
Normalized Difference Vegetation Index	Greenness	Sentinel	2019 (summer)	10 × 10 metre
Land surface temperature	Temperature in °F of land surface	Landsat	2020–2023 (summer)	30 × 30 metre
% impervious surface	Percent impervious surface	C-HEAT	2020	Parcel
Distance to water	Distance (m) from parcel to water body in meters	US Hydrological Survey, MassGIS	2020	Parcel
Distance to major road	Distance (m) from parcel to major road	MassGIS	2020	Parcel
Population characteristics	Population count	U.S. Census	2020	Parcel

GIS = Geographic Information Systems; C-HEAT = Chelsea and East Boston Heat Study.

Data analysis

We calculated the average temperature and the maximum temperature over the entire warm season (June 1–September 30) using R (4.0.2). We defined hot weeks as 7 consecutive days containing a heat wave (3 or more consecutive days with temperatures at or above 90°F or 32.2°C; Robinson, 2001). We created descriptive graphs and tables to understand how ambient temperature sensor data (DCR and C-HEAT HOBOs) compared to the federally managed NWS sensor data during the warm season and hot weeks. We used statistical tests of means (two-sample paired t-test) to quantify these ambient temperature measurements between sensor types.

To describe ambient temperature variation as a function of natural/built environment variables within Chelsea, we conducted land use regression (LUR) in ArcPro, a GIS mapping software. For these analyses, our dependent variable was ambient temperature at each local sensor point, and our explanatory (independent) variables were natural/built environment variables listed in Table 1. For the LUR models, we conducted correlation analyses of Table 1. variables: We excluded land surface temperature due to moderate-high correlation with percent impervious surface (r = −.74) and NDVI (r = .67). All other natural/built environment variables showed weaker correlations (r < .50). We also conducted stepwise regression considering adjusted R² values and Akaike information criterion as guidance for adding or removing variables from the model. We produced LUR models for different temperature scenarios. First, we focused on the 2020–2023 warm seasons (average and maximum recorded temperatures). We then used 2020 (warm season and a hot week) as an example of how to identify hotter areas within Chelsea. Finally, we analyzed the 2020 daytime and nighttime temperatures during the warm season to highlight locations that may not cool down at night. We defined daytime as between 9 a.m. and 9 p.m. (Murage et al., 2017).

The final five independent variables (percent impervious surface, NDVI, distance to water, distance to major roads, and population count) were spatially joined to each sensor point for geographically weighted regression (GWR) analyses. We used the GWR tool in ArcPro to model these spatially varying relationships (ESRI, 2023b). Though our scale of analysis is small (Chelsea parcels), GWR methodology accounts for spatial autocorrelation and non-stationarity by calculating local regression equations for each feature (ESRI, 2023b). We conducted GWRs for different time periods to identify ambient heat variation within Chelsea that parallel each LUR model: 2020–2023 warm seasons, 2020 warm season and hot week, and 2020 daytime and nighttime. For the day–night analysis, we produced a bivariate choropleth map to identify areas in Chelsea where temperatures are both hottest during the day and night. Bivariate choropleth maps effectively and accurately display the relationship between two spatially distributed variables by utilizing a combination of various symbols and colours (ESRI, 2023a). This visualization highlights potential neighbourhoods most in need of heat-related health interventions. We used kriging—spatial interpolation of point data—to visualize the GWR temperature results and further account for spatial autocorrelation (ESRI, 2023b).

Results

Local temperature sensor data

Average ambient temperatures during the entire warm season study period between 2015 and 2023 ranged from 41.4°F to 106.8°F (5.2°C to 41.6°C) for DCR sensors and from 40.7°F to 98.5°F (4.8°C to 36.9°C) for C-HEAT sensors (between 2020 and 2023). NWS temperature data ranged between 44.1°F and 98.9°F (6.7°C to 37.2°C). Detailed yearly ambient temperature averages can be found in the Appendix (Tables A1 and A2).

Warm-season and hot week ambient temperatures during 2020–2023 varied between warm season years, hot weeks, and daytime versus nighttime. Average summer daytime temperatures ranged from 74.0°F to 77.7°F (23.3°C to 25.4°C), and average summer nighttime temperatures ranged from 68.0°F to 69.8°F (20.0°C to 21.0°C). Temperatures for both day and night were consistently higher during hot weeks: average temperatures ranged from 80.6°F to 89.5°F (27.0°C to 31.9°C) during the day and from 71.8°F to 80.2°F (22.1°C to 26.8°C) during the night.

Local versus NWS ambient temperature sensor data

Both mean and maximum temperature significantly differed between sensor sources for many of the study period years. Comparing DCR and NWS sensors detailed in Table 3a, we see that mean temperature differences ranged from 0.6°F/0.3°C (95% CI −1.1°F to 2.2°F; −0.6°C to 1.2°C) in 2021 to 3.8°F/2.1°C (95% CI 1.9°F to 5.7°F; 1.1°C to 3.2°C) in 2020. All years showed higher temperatures captured by DCR sensors compared to the NWS sensor, with all years except 2018, 2019, and 2021 being statistically significantly higher. This trend continued when considering maximum temperature differences: DCR sensors were reporting statistically significantly higher temperatures than the NWS sensor for all years, with the maximum temperature difference ranging from 8.4°F/4.7°C (95% CI 6.1°F to 10.7°F; 3.4°C to 5.9°C) in 2021 to 20.3°F/11.3°C (95% CI 12.8°F to 22.9°F; 7.1°C to 12.7°C) in 2017.

Table 3:

Warm season daily temperature (°F) differences between sensortypes

(a) DCR versus NWS
Year	Mean Temp Diff (95% CI)	Max Temp Diff (95% CI)
2015	3.6 (1.8 to 5.5)	19.1 (16.8 to 21.4)
2016	3.1 (1.3 to 4.8)	17.9 (15.5 to 20.4)
2017	3.0 (1.3 to 4.7)	20.3 (17.8 to 22.9)
2018	1.7 (−0.2 to 3.6)	10.7 (8.2 to 13.2)
2019	1.2 (−0.6 to 2.9)	15.4 (13.2 to 17.7)
2020	3.8 (1.9 to 5.7)	12.0 (9.8 to 14.3)
2021	0.6 (−1.7 to 2.2)	8.4 (6.1 to 10.7)
2022	2.7 (0.8 to 4.7)	11.0 (8.7 to 13.4)
2023	2.5 (0.3 to 4.8)	9.9 (7.0 to 12.7)
(b) C-HEAT versus NWS
Year	Mean Temp Diff (95% CI)	Max Temp Diff (95% CI)
2020	1.8 (0.9 to 4.4)	3.5 (0.6 to 6.5)
2021	−0.6 (−2.6 to 1.4)	1.1 (−1.5 to 3.7)
2022	1.4 (−0.5 to 3.3)	5.7 (3.3 to 8.0)
2023	1.2 (−0.6 to 2.9)	4.9 (2.9 to 7.1)
(c) DCR versus C-HEAT
Year	Mean Temp Diff (95% CI)	Max Temp Diff (95% CI)
2020	1.2 (−1.6 to 3.9)	7.2 (3.9 to 10.4)
2021	1.2 (−0.8 to 3.3)	7.3 (4.3 to 10.2)
2022	1.4 (−0.60 to 3.3)	5.4 (2.8 to 7.9)
2023	1.4 (−0.9 to 3.6)	5.1 (2.2 to 8.0)

DCR = Department of Conservation and Recreation; NWS = National Weather Service; C-HEAT = Chelsea and East Boston Heat Study

We see similar trends comparing the C-HEAT sensors and the NWS sensors (Table 3b): The C-HEAT sensor measurements are consistently higher than NWS sensor measurements, except for the mean temperature in 2021. The maximum temperature differences between the C-HEAT and NWS sensors are smaller in magnitude, ranging from 1.1°F/0.6°C (95% CI −1.5°F to 3.7°F; −0.8°C to 2.1°C) to 5.7°F/3.2°C (95% CI 3.3°F to 8.0°F; 1.8°C to 4.4°C). Finally, the DCR sensors consistently reported higher temperature measurements compared to the C-HEAT temperature measurements (Table 3c). Overall, the DCR sensors reported the highest temperature measurements, especially maximum temperature measurements, followed by the C-HEAT sensors and then the NWS sensor.

When exploring specific hot weeks during the study period across data sources, we see differences in both average and maximum temperatures. Table 4 details temperature variation patterns and sensor type differences, respectively. Table 4 outlines hot weeks across 2015–2023, with average ambient temperatures ranging from 73.6°F to 82.7°F (23.1°C to 28.2°C) for NWS sensors and 77.1°F to 85.7°F (25.1°C to 29.8°C) for DCR sensors. Between 2020 and 2023, average ambient temperatures ranged from 75.9°F to 84.1°F (24.4°C to 28.9°C) for C-HEAT sensors. Comparing ambient temperature differences between sensor types (HOBO versus NWS) across years where all sensor sources are present (2020–2023), we see a consistent trend: NWS sensor temperatures tend to be lower than Chelsea-based sensors (on average, 2.5°F/1.4°C lower than DCR and 1.3°F/0.7°C lower than C-HEAT). Based on two-sample pairwise t-tests, the difference between all mean temperature measurements across sensor types was statistically significant (NWS versus DCR, NWS versus C-HEAT, and DCR versus C-HEAT). Additional information on yearly sensor differences can be found in Table A3.

Table 4:

Hot week daily mean (SD) and range temperature (°F) for each sensor type

	NWS (Boston Logan)		DCR (Chelsea)		C-HEAT (Chelsea)
Year (hot week)	Mean (SD)	Range	Mean (SD)	Range	Mean (SD)	Range
2015 (9/3–9/9)	74.2 (8.4)	57.9–93.9	77.6 (10.2)	55.2–107.0	–	–
2016 (7/21–7/27)	80.4 (7.2)	68.0–97.0	83.5 (9.2)	64.3–107.0	–	–
2017 (6/9–6/15)	74.2 (10.6)	55.0–95.0	78.3 (11.7)	54.1–107.0	–	–
2018 (8/24–8/30)	79.6 (8.6)	62.1–97.0	81.4 (9.6)	59.5–106.0	–	–
2019 (7/28–8/3)	80.5 (6.9)	68.0–96.1	82.7 (9.0)	62.1–107.0	–	–
2020 (8/19–8/25)*	73.6 (6.6)	60.98–88.0	77.1 (8.7)	58.0–102.0	75.9 (7.0)	59.6–91.4
2021 (6/4–6/10)	79.4 (8.4)	64.0–95.0	80.3 (8.2)	58.5–107.0	79.6 (6.2)	61.2–93.5
2022 (8/3–8/9)	82.7 (7.4)	69.1–98.1	85.7 (8.8)	68.4–107.0	84.1 (6.9)	70.1–100.3
2023 (7/23–7/29)	77.6 (5.6)	68.0–91.0	80.3 (8.2)	62.9–98.9	78.9 (6.3)	66.5–94.9

^*Data are limited for C-HEAT (monitoring started 8/12/20); hot week heat wave threshold was adjusted to >84°F for 3 consecutive days or more

NWS = National Weather Service; DCR = Department of Conservation and Recreation; C-HEAT = Chelsea and East Boston Heat Study.

LUR results: Ambient temperatures and natural/built environment characteristics

LUR results were similar across different modelling scenarios. Table 5a details LUR output for the mean and maximum temperatures during the 2020–2023 warm seasons. Measures of effect for each variable are slightly larger when considering maximum temperature (versus mean temperature), although are small overall. For example, as the distance to a water body increases by 100 metres from a given parcel in Chelsea, the mean ambient temperature increases 0.189°F/0.105°C (95% CI 0.090°F to 0.288°F; 0.050°C to 0.160°C). We see positive associations with ambient temperature for percent impervious surface, distance to major roads, distance to water, and population count. We see a negative association with ambient temperature for NDVI: As NDVI increases by one unit (more greenness), mean ambient temperature decreases by 0.048°F/0.027°C (95% CI −2.029°F to 1.934°F; −1.127°C to 1.074°C). However, only associations between distance to major roads and distance to water are statistically significant. Model fit metrics in Table 5b show significant F-statistics for both models: variable coefficients are significantly different from zero. Adjusted R² values for both models are .09 (mean temperature) and .18 (maximum temperature): The percentage variation in ambient temperature explained by the natural/built environment variables is approximately 9% for the mean temperature model and 18% for the maximum temperature model.

Table 5a:

LUR output for mean and maximum temperature during warm seasons 2020–2023

	Mean Temperature					Maximum Temperature
Variable	Coefficient	SE	95% CI		t-stat	Coefficient	SE	95% CI		t-stat
Percent impervious surface	0.003	0.007	−0.011	0.018	0.423	0.016	0.019	−0.021	0.052	0.842
Distance to Roads (100 m)*	0.220	0.075	0.072	0.368	2.920	0.622	0.190	0.250	0.993	3.276
Distance to Water (100 m)*	0.189	0.051	0.090	0.288	3.725	0.702	0.128	0.452	0.953	5.497
NDVI	−0.048	1.011	−2.029	1.934	−0.047	1.734	2.545	−3.254	6.723	0.681
Population	0.000	0.001	−0.002	0.002	−0.104	0.000	0.002	−0.005	0.005	0.105

Table 5b:

Model fit metrics for mean and maximum temperature LUR models

Metric	Mean temperature model	Maximum temperature model
F-Statistic*	3.563	6.745
Adjusted R2	0.089	0.180

^*significant at p < .05

LUR = land use regression; NDVI = Normalized Difference Vegetative Index

Table 6a details LUR output for the 2020 warm season and a specific hot week. Measures of effect for each variable are larger for the hot week, though are small overall. Interestingly, we see a negative association between ambient temperature and percent impervious surface during the warm season (−0.008°F/−0.004°C, 95% CI −0.576°F to 0.020°F; −0.320°C to 0.011°C) but a positive association during the hot week (0.007°F/0.004°C, 95% CI −0.034°F to 0.048°F; −0.019°C to 0.027°C). However, these relationships are not statistically significant, nor do they have large measures of effect. For the remaining variables, we see a positive association with ambient temperature for distance to major roads, distance to water, and population count. We see a negative association with ambient temperature for NDVI: For example, as NDVI increases by one unit (more greenness), ambient temperature decreases by 1.525°F/0.847°C (95% CI −2.029°F to 1.934°F; −1.127°C to 1.074°C) during the warm season. Again, only associations between distance to major roads and distance to water are statistically significant. Model fit metrics in Table 6b show statistically significant F-statistics for both models. Adjusted R² values for both models are .25 (warm season) and .12 (hot week). The percentage variation in ambient temperature explained by the natural/built environment variables is approximately 25% for the warm season model and 12% for the hot week model.

Table 6a:

LUR output for 2020 warm season and hot week (temperature in °F)

	Warm season					Hot week
Variable	Coefficient	SE	95% CI		t-stat	Coefficient	SE	95% CI		t-stat
Percent impervious surface	−0.008	0.015	−0.037	0.020	−0.576	0.007	0.021	−0.034	0.048	0.329
Distance to roads (100 m)*	0.494	0.156	0.189	0.800	3.172	0.418	0.230	−0.032	0.868	1.819
Distance to water (100 m)*	0.424	0.091	0.245	0.602	4.660	0.392	0.130	0.137	0.648	3.014
NDVI	−1.535	2.030	−5.515	2.444	−0.756	−0.156	2.810	−5.663	5.351	−0.056
Population	0.002	0.002	−0.003	0.006	0.821	0.001	0.003	−0.004	0.007	0.464

Table 6b:

Model fit metrics for 2020 warm season and hot week LUR models

Metric	Warm season model	Hot week model
F-Statistic*	5.520	2.561
Adjusted R2	0.249	0.109

^*significant at p < .05

LUR = land use regression; NDVI = Normalized Difference Vegetative Index

Table 7a details the LUR output for the daytime and nighttime ambient temperatures during the 2020 warm season. The measures of effect for each variable are larger during the day, although they are small overall. Interestingly, we see a negative association between ambient temperature and NDVI during the daytime (−0.077°F/−0.043°C, 95% CI −9.901°F to 9.748°F; −5.501°C to 5.415°C) but a positive association during the nighttime (0.041°F/0.023°C, 95% CI −5.556°F to 5.538°F; −3.087°C to 3.077°C). However, these relationships are not statistically significant, nor do they have large measures of effect. For the remaining variables, we see a positive association with ambient temperature for distance to major roads, distance to water, and population count. Again, only associations between distance to major roads and distance to water are statistically significant. There is a negative relationship between ambient temperature and NDVI. Model fit metrics in Table 7b show statistically significant F-statistics for both models. Adjusted R² values for both models are .20 (daytime) and .08 (nighttime): The percentage variation in ambient temperature explained by the natural/built environment variables is approximately 20% for the daytime model, and 8% for the nighttime model.

Table 7a:

LUR output for day–night analysis (temperature in °F)

	Day				Night
Variable	Coefficient	SE	95% CI	t-stat	Coefficient	SE	95% CI	t-stat
Percent impervious surface	−0.023	0.035	−0.091	0.045	−0.670	−0.019	0.020	−0.058	0.019	−0.980
Distance to Roads (100 m)*	1.073	0.375	0.339	1.807	2.865	0.521	0.212	0.105	0.937	2.455
Distance to Water (100 m)*	0.962	0.223	0.524	1.400	4.304	0.344	0.127	0.095	0.592	2.712
NDVI	−0.077	5.012	−9.901	9.748	−0.015	0.014	2.842	−5.556	5.583	0.005
Population	0.005	0.006	−0.006	0.016	0.972	0.001	0.003	−0.005	0.008	0.439

Table 7b:

Model fit metrics for day–night LUR models

Metric	Day model	Night model
F-Statistic*	4.433	2.166
Adjusted R2	0.202	0.079

^*significant at p < .05

LUR = land use regression; NDVI = Normalized Difference Vegetative Index

Kriged modelling results: Ambient temperature variation across different scenarios

Ambient temperature varied spatially across Chelsea during the warm season. Figures 2a–2d show the average and maximum ambient temperature (°F) variation within Chelsea during the 2020–2023 warm seasons. Maps on the left show kriged observed ambient temperature while maps on the right show kriged GWR-modelled ambient temperature. All maps display hotter areas (darker red) throughout the central vertical third of Chelsea and cooler areas (lighter pink) in the western and eastern areas of Chelsea. Maximum temperature maps show a larger area of high heat in both north and central Chelsea. Of note, observed temperature maps visually display finer scale temperature variation differences, while GWR-modelled temperature maps show wider scale variation patterns.

Figure 2:

Ambient temperature (°F) maps for 2020–2023 warm seasons, showing kriged surfaces of the (a) observed average temperature (b) geographically weighted regression (GWR)-modelled average temperatures and the (c) observed maximum temperatures (d) GWR-modelled maximum temperatures

We use 2020 data to showcase an example of mapping ambient temperature variation for one summer and one hot week to identify the hottest areas within Chelsea (Figure 3). The warm season overall versus hot week predicted temperature maps show few differences; however, hotter temperatures do expand farther into southwest Chelsea during the hot week. Again, we see finer variation in the observed hot week temperature map (Figure 3c).

Figure 3:

Maps of (a) GWR-modelled warm season temperature (°F), (b) GWR-modelled hot week temperature (°F), and (c) observed hot week temperature (°F) using all low-cost sensors for 2020 during the warm season

We extend the mapping example from Figure 3 to Figure 4, mapping spatially varying day and nighttime maximum temperatures. In Figure 4a, the hottest areas tend to concentrate in the central vertical third of Chelsea (darker red), with cooler areas on the western and eastern edges (lighter pink). This pattern stays consistent moving from day to night, however, part of the western edge of Chelsea appears to cool down more during the nighttime compared to the daytime (Figure 4b). The bivariate choropleth highlights areas that are both hottest during the day and night: most consistently concentrated in the central vertical third of the city (dark purple; Figure 4c). We show green space (green) and parcel delineation (outlines) for ease of interpretation in this map: Cooler areas appear to be closer to green spaces, while darker areas tend to be in more residentially populated, impervious parcels (see Figure A4).

Figure 4:

(a) Day- and (b) night-time maximum temperatures (°F) with (c) bivariate choropleth symbology during 2020 in Chelsea, Massachusetts

Discussion

In this study, we analyzed ambient temperature variation across recent warm seasons and hot weeks in the highly urbanized city of Chelsea, Massachusetts, leveraging a local network of sensors at a fine spatial scale. Our findings reveal notable variations in temporal, spatial, and sensor-type temperature measurements.

Our sensor networks’ variability in ambient temperature measurements across different sensor sources poses implications for accuracy in environmental monitoring and climate planning. Although the average temperature trends during the warm season generally remained relatively consistent between sensors, more divergent trends emerged when exploring maximum temperatures during hot weeks: The NWS sensor values consistently underestimated local temperatures as measured by C-HEAT and DCR sensors. This observed trend could be impacted at the sensor location by factors such as local weather conditions, physical structures, and/or distance to water, tree cover, and shade; failing to account for these aspects may result in heat exposure misclassification. Other studies have demonstrated discrepancies between local monitoring networks and NWS sensors during high-temperature time periods, with some citing the potential impact of physical installation location (Clement et al., 2023; Feichtinger et al., 2020; Meier et al., 2017). Although limitations of sensor placement may exist, we emphasize the importance of identifying locations consistently hotter compared to the NWS sensor measurements because protections like heat advisories, which rely solely on defined temperature thresholds, may fail to alert residents living in locations hotter than described by the closest NWS sensor. Climate planning for public health must incorporate local data, if available, to ensure heat advisories and other means of environmental health communications reach the most vulnerable (Hondula et al., 2015).

Identifying influential natural/built environment contributors to heat can help policymakers and public health officials prioritize at-risk locations for potential cooling interventions. Consistent patterns emerged through our LUR analyses, indicating that physical geography and human activities may contribute to local ambient temperature variation. Notably, proximity to water bodies and major roads carried statistically significant relationships with ambient temperature. The variability in ambient temperature explained by natural/built environment factors, as denoted by the adjusted R² values, spans from 9% for mean temperature to 18% for maximum temperature models, extending to 25% for the warm season and 12% for a particularly hot week. These low R² values in our LUR models suggest the need for the development of updated, spatially resolved neighborhood-level natural/built environment variables that can better explain small intra-urban temperature variability. Daytime–nighttime differences in LUR models presented inconsistent (not statistically significant) results between NDVI and temperature: a negative relationship during the day and a positive relationship at night. Future climate research and planning should aim to integrate updated high resolution (<5 m) Light Detection and Ranging– or National Agriculture Image Program–derived greenness data to potentially improve analyses, although paywalls and computing expertise requirements exist (Davis et al., 2016).

A study conducted a regression analysis in metropolitan Portland, Oregon, to determine intra-urban UHI differences, found that industrial and commercial areas with greater impervious surfaces were related to higher temperatures (Hart & Sailor, 2009). Interestingly, areas with more tree canopy cover were most strongly associated with cooler areas within Portland; this is an important data variable for development in Chelsea to improve future predictive analyses (Hart & Sailor, 2009). Although we used a unique set of variables in our own LUR analysis, we saw similar directions of effect for certain variables in similar studies, such as distance to a water body: Higher temperatures were positively associated with greater distance away from a water body (Tsin et al., 2020; Venter et al., 2020; Zhou et al., 2014). Compared to LUR models in similar studies, our LUR models do not report statistical significance for key variables such as percent impervious surface or NDVI. Improving these variables in terms of spatial scale and updated years of availability may benefit new iterations of LUR modelling in Chelsea. Ultimately, our work aimed to produce results relevant for the city of Chelsea, which experiences a specific climatology and has unique climate action planning goals. There is no universal predictive model for characterizing extreme heat in cities, especially when prioritizing public health and policy-relevant interventions (Hondula et al., 2015).

The spatial variation of ambient temperatures in Chelsea, especially during warm seasons and hot weeks, highlights consistent patterns, with hotter areas primarily concentrated in residentially dense, impervious surface-heavy areas of Chelsea. During an example hot week, we observed higher temperatures both during the day and night compared to the overall summer averages in similar areas. Mapping different scenarios for extreme heat implications can be used for future climate projection planning and policies (de Sherbinin et al., 2019). In this case, we show locations that are hottest in Chelsea throughout the day and night: researchers have shown that persistent nighttime heat exposure can be especially deadly (Kovats & Hajat, 2008). We can identify these locations as a priority for resource deployment: for example, providing air conditioning (AC) is a primary way to quickly reduce heat-related morbidity and mortality in residential settings (Ito et al., 2018). Future efforts can also compare these locations with other social, economic, environmental, and health vulnerability data for different targeted interventions: the publicly available C-HEAT data dashboard provides many of these spatial datasets (C-HEAT, 2023). By offering an understanding of ambient temperature variations within the city of Chelsea, results can complement a variety of data-driven approaches in urban climate decision-making and the creation of impactful geospatial visualizations for policy applications.

Presenting both kriged GWR-modelled and observed ambient temperature maps allowed us to cross-check our analyses and further account for spatial correlation. Overall, our GWR-modelled and observed temperature maps follow similar spatial patterns of heat within Chelsea: There are consistently three main clusters of higher heat within the central third of Chelsea. However, using Figure 4, we observe a difference in temperature distribution in the southwest tip of Chelsea: GWR-modelled measurements, especially during the hot week, show more areas with hotter temperatures. Here, we see the model accounting for the influence of variables associated with increased ambient temperature, for example, the major roadways in that area (Arnfield, 2003). For policymakers, we stress the importance of including geospatial analytical expertise in city climate planning as a best practice for informed decision-making (Kamel Boulos & Wilson, 2023).

It is also important that policymakers and city staff consider the nuances that heat vulnerability maps may not fully capture when considering cooling interventions, especially in EJ communities. Previous qualitative studies showed that a sample of Chelsea residents often refrained from using their portable air conditioner due to associated energy costs (McIntyre et al., 2022; Milando et al., 2022). We also learned from residents that even though the city of Chelsea provided AC units and utility bill assistance through a lottery to 100 eligible residents, hundreds of additional eligible applicants were not selected due to limited resources (Bebinger, 2021b). Ultimately, policy and planning applications of mapping extreme heat must move toward ensuring long-term, community-driven, and systematic solutions that directly reflect resident concerns. According to a recent C-HEAT report (Gross et al., 2023), relevant policy opportunities for Chelsea, the Commonwealth of Massachusetts, and other cities should include more widespread energy cost subsidies, support for building weatherization and/or decarbonization, and passive and active cooling requirements for rental properties.

The city of Chelsea and GreenRoots, Inc., have already begun efforts to cool a specific hot area within Chelsea named the Cool Block Project (Bebinger, 2022; Film, 2022). This residential block was not only hotter based on C-HEAT sensor measurements but was almost entirely an impervious surface area with very little green space or trees. Interventions on this block are still in progress, and include tree planting, re-pavement to lighter impervious surfaces, and the transformation of a vacant lot into a park with grass, plants, and shaded spaces. Through these small-scale interventions, the success of the Cool Block Project has primed community involvement to make climate planning in Chelsea relevant and meaningful to residents (C-HEAT, 2023; McIntyre et al., 2022). Additional hot spots in Chelsea confirmed by our sensor network included the downtown area of Bellingham Square, which is currently in the process of a redesign to increase shading, greenspace, and flooding resilience (Healthy Chelsea, 2022). The C-HEAT team also assisted in identifying Chelsea Public School roofs in hot areas to be painted to a lighter, less heat-absorbing colour instead of black (Bebinger, 2021a). Our updated ambient temperature results will help to inform future climate interventions in Chelsea and must be combined with community engagement to ensure sustainability and support. By producing publicly available reports of our work, we also hope to assist in supporting replicable actions by a range of stakeholders, including policymakers (C-HEAT, 2023).

Our study had several limitations. Sensor placement varied in terms of physical location type, potentially impacting temperature measurement accuracy: The DCR placed sensors in primarily unshaded, residential/commercial areas, while C-HEAT sensors were placed in trees, which may explain the consistently lower temperature measurements compared to the DCR sensors. Additionally, the local NWS monitor is surrounded by an industrial airport close to the Atlantic Ocean, potentially experiencing different overall weather patterns. Consistently low R² values in our LUR models suggest that a large percentage of our models are not fully explained by our data; finer scale, additional natural/built environment factors are needed to better explain trends in ambient temperature within Chelsea. Specifically, we would advocate for future collection of tree canopy cover data (i.e., on-the-ground tree surveys, updated fine-scale satellite data), and other metrics of greenness at the neighbourhood level. Ultimately, these factors were not considered due to a lack of data availability. Due to the spatial scale of our analyses, spatial autocorrelation is a concern in our LUR models and their interpretability. However, subsequent GWR and kriging of these data account for this limitation. GWR results may have underestimated heat in industrial areas, which are predominately impervious surfaces, due to a lack of sensors in those locations. Modelling using GWR is inherently as accurate as the data we include; if additional data were available, we could have added other variables such as boundary layer conditions or weather conditions (humidity, wind speed, etc.) to better estimate the ambient temperature at each sensor location. Additionally, multicollinearity may still exist between predictor variables used in the GWR analysis that could result in inflated standard errors and less reliable model results (ESRI, 2023b). As such, the GWR acted as an exploratory method to understand trends that our study uses to show its utility in assessing ambient temperature patterns. The strengths of our research include using local data at a high spatial and temporal resolution: Our final models and maps included up to 80 sensors with hourly temperature data for summer months across multiple years. We showed specific examples of how to visualize data to identify areas within a city impacted by extreme heat.

Future climate planning efforts should continue to support local monitoring to identify neighbourhoods in need of cooling interventions. Additional modelling methods should include down-scaled (within-city) weather conditions and/or sociodemographic variables, although this may require specific expertise not commonly available at the municipality level. With these down-scaled weather data, researchers and policymakers would be able to target mechanisms for public protection and alerts during extreme heat events at the neighbourhood level rather than rely on city- or region-wide averages from an NWS monitor.

Integrating data sources that connect directly to cooling interventions, such as the need for AC and usage, cooling centre locations and availability, and energy-saving programmes, should be pursued for policy-relevant action. These projects should consider the use of qualitative data (community-driven input) to validate quantitative analysis to resolve questions including Can people afford AC? How effective are current cooling centers, if at all? and Are funding levels for energy cost assistance programmes sufficient? Effective climate planning will depend on the harmonious integration of advanced hazard monitoring, data modelling, community-driven projects, and public policy.

Conclusion

Our analysis in Chelsea, Massachusetts, emphasizes the potential of local ambient temperature data in guiding urban climate policy and planning. Our analyses reiterate the findings of prior studies that urban heat exposure is influenced by the built environment and the current lack of a dense temperature sensor network. Interventions such as the Cool Block Project demonstrate the effectiveness of community-engaged efforts informed by local heat data. Our day–night mapping example offers valuable insights for prioritizing resource allocation in heat-vulnerable areas. While our study acknowledges limitations in sensor placement and data modelling, we ultimately recommend integrating local environmental data with community insights for impactful climate resilience strategies. Overall, this research affirms the need for locally tailored data in developing effective urban heat mitigation policies.

Table 2:

Daily (minimum, mean, maximum) temperatures (°F): day and night (warm season, hot week) from local ambient temperature sensors

		Minimum		Mean		Maximum
	Year	Warm season	Hot week	Warm season	Hot week	Warm season	Hot week
Day	2020	69.3	71.4	77.0	80.6	83.7	88.4
	2021	71.0	76.9	77.4	85.8	82.9	92.7
	2022	70.1	78.4	77.7	89.5	84.0	98.0
	2023	67.1	75.0	74.0	84.0	79.6	91.9
Night	2020	65.0	67.6	68.8	71.8	75.0	79.7
	2021	65.9	69.6	69.4	74.4	74.8	82.3
	2022	66.0	76.1	69.8	80.2	76.1	87.5
	2023	64.3	70.3	68.0	75.1	73.8	81.9

SUMMARY FOR POLICYMAKERS.

• Temperature varies within the city of Chelsea; some areas were notably hotter than others. By visualizing and identifying hotter areas, we can assist in prioritization of cooling intervention efforts.

  • Identifying areas that stay hot into the night can be used to target home-based interventions focused on preventing human health risks.

• GIS is a visualization tool that can be used to highlight within-city areas requiring intervention.

  • Up-to-date local natural/built environment data are necessary to evaluate contributors to extreme heat in cities.

• Research conducted with and by community-based organizations increases the likelihood that the studies’ results reflect the concerns of community members.

• Local temperature monitoring allows for a more detailed picture of disproportionate extreme heat exposure within cities. This can help increase equity in climate action planning, as many areas with disproportionate exposures were impacted by long-standing systematic economic and social disinvestment.

Appendix

Table A1:

Hourly mean (SD) and range for ambient temperature (°F) for stationary sensors in Chelsea across study period years (2015–2023)

	NWS (Boston Logan)		DCR (Chelsea)		C-HEAT (Chelsea)
Year	Mean (SD)	Range	Mean (SD)	Range	Mean (SD)	Range
2015	70.12 (8.77)	46.04–93.92	73.77 (10.41)	42.92–106.84	–	–
2016	71.67 (8.55)	48.02–95.98	74.83 (10.10)	44.03–106.84	–	–
2017	69.83 (8.09)	46.94–95.00	72.89 (10.16)	47.09–106.84	–	–
2018	72.24 (8.79)	48.92–96.98	73.97 (10.14)	47.98–106.84	–	–
2019	71.99 (8.29)	50.00–96.98	73.15 (10.03)	43.65–106.84	–	–
2020	70.50 (8.85)	44.06–95.00	74.34 (10.39)	41.44–106.84	–	–
2021	73.08 (7.93)	51.98–98.06	73.67 (9.45)	48.70–106.84	73.32 (8.31)	52.02–97.79
2022	71.44 (8.99)	48.02–98.96	68.96 (11.25)	46.19–106.84	73.38 (8.94)	47.82–98.53
2023	70.83 (7.54)	46.94–91.04	73.11 (9.81)	46.73–103.31	68.89 (9.45)	40.71–92.78

NWS = National Weather Service; DCR = Department of Conservation and Recreation; Chelsea and East Boston Heat Study

Table A2:

Warm season hourly mean (SD) ambient temperature (°F) for stationary sensors in Chelsea across study period years (2015–2023)

	NWS (Boston Logan)	DCR (Chelsea)	C-HEAT (Chelsea)
Year	Mean (SD)	Mean (SD)	Mean (SD)
2015	70.12 (8.77)	73.77 (10.41)	–
2016	71.67 (8.55)	74.83 (10.10)	–
2017	69.83 (8.09)	72.89 (10.16)	–
2018	72.24 (8.79)	73.97 (10.14)	–
2019	71.99 (8.29)	73.15 (10.03)	–
2020	70.50 (8.85)	74.34 (10.39)	–
2021	73.08 (7.93)	73.67 (9.45)	73.32 (8.31)
2022	71.44 (8.99)	68.96 (11.25)	73.38 (8.94)
2023	70.83 (7.54)	73.11 (9.81)	68.89 (9.45)

NWS = National Weather Service; DCR = Department of Conservation and Recreation; Chelsea and East Boston Heat Study

Figure A1:

Maps of selected predictor variables for geographically weighted regression in Chelsea, Massachusetts. NDVI = Normalized Difference Vegetative Index

Figure A2:

Parcels in Chelsea, Massachusetts