The Surface Is Not Superficial: Utilizing Hyper-Local Thermal Photogrammetry for Pedestrian Thermal Comfort Inquiry

Highlights

What are the main findings?

• Landsat 8 and 9 Satellite-Derived Land Surface Temperatures (SD-LSTs) were, on average, 10.7 °C hotter than measurements captured with Forward Looking Infrared (FLIR) thermal imagery.
• FLIR measurements were strongly correlated with biometeorological metrics, with models explaining 50–66% of the variance.

What are the implications of the main findings?

• Landsat SD-LST is too coarse for pedestrian thermal comfort analysis.
• Low-cost, FLIR thermography offers practical, fine-scale heat data for public transport research and heat-resilient design.

Abstract

The scale and magnitude of urban heating are often assessed using Satellite-Derived Land Surface Temperature (SD-LST). Yet, discrepancies in spatial resolution limit SD-LST’s ability to reflect pedestrian thermal experience, potentially leading to ineffective mitigation strategies. Hyper-local measurements of urban heat, defined as surface temperatures (T_S) at the scale of pedestrian activity (e.g., bus stops or street segments), may provide more accurate insights into thermal comfort. This study compares hyper-local ~0.01 m resolution T_S collected via consumer-grade Forward-Looking Infrared (FLIR) thermography with resampled 30 m resolution SD-LST from Landsat 8 and 9 images to evaluate their utility in predicting thermal comfort indices across 60 bus stops in Denver, Colorado. During the summer of 2023, 270 FLIR measurements were collected over 19 dates, with a four-day subset (n = 33) coinciding with Landsat imagery. FLIR T_S averaged 25.12 ± 5.39 °C, while SD-LST averaged 35.90 ± 12.56 °C, a significant 10.77 °C difference (95% CI: 6.81–14.73; p < 0.001). FLIR T_S strongly correlated with biometeorological metrics such as air temperature and mean radiant temperature (r > 0.8; p < 0.001), while SD-LST correlations were weak (r < 0.3). Linear mixed-effects models using FLIR T_S explained 50–66% of the variance in thermal comfort indices and met ISO 7726 standards. Each 1 °C increase in FLIR TS predicted a 0.75 °C rise in mean radiant temperature. These results highlight hyper-local thermography as a reliable, low-cost tool for urban heat resilience planning.

1. Introduction

The Urban Heat Island (UHI) is commonly evaluated using Satellite Derived Land Surface Temperature (SD-LST), which estimates Earth’s surface temperature from thermal infrared satellite bands [1,2]. The satellite’s image resolution, land cover type, albedo, climate, time of image capture, and other factors drive SD-LST estimation [3]. However, SD-LST quantification and Surface UHI classification face several challenges. Cloud cover can obscure satellite-based SD-LST estimation, and there can be unavoidable trade-offs between spatial and temporal resolution of the satellites used in their capture [4,5,6]. Moreover, SD-LST is a two-dimensional representation of surface temperature (T_S), which oversimplifies the complexity of three-dimensional thermal environments [7]. For all of these reasons, SD-LST provides a limited portrait of the thermal environment, especially when used for urban heat mitigation inquiry [8].

Despite these challenges, urban heat studies often rely on quantifying the thermal environment through these coarse SD-LST measurements, largely because these metrics are easy to access and are global in scope. A systematic literature review of land use and land cover’s impacts on SD-LST identified the Landsat satellites as the most commonly used satellite for the computation of both land cover composition and SD-LST, accounting for a majority of studies [9], with the resampled 30 m resolution metric being used in the majority of recent papers investigating urban cooling [10]. Thus, while SD-LST is valuable for assessing macro and meso-scale UHI impacts, such as in the creation of Local Climatic Zones (LCZs) to monitor neighborhood-level conditions [11], its utility diminishes at hyper-local scales that capture site-specific variability within just a few meters. For biometeorologists, this often necessitates moving beyond SD-LST toward hyper-local measurements that more accurately reflect the thermal environment and pedestrian thermal comfort [12,13].

Recognizing this scale mismatch, urban climatologists and remote sensing scientists have developed methods for downscaling SD-LST to finer spatial resolutions using statistical, machine-learning, and multi-sensor fusion techniques. These efforts underscore that the gap between coarse satellite-derived and fine-scale thermal environments is known [14]. Yet even when these methods successfully refine the spatial resolution of SD-LST, they still yield a metric that represents only one component of the thermal environment and does not directly capture the radiative load or convective conditions experienced by pedestrians. Thus, even when downscaled, SD-LST alone cannot represent the drivers of pedestrian thermal comfort and must ultimately be validated against micrometeorological conditions [15].

The American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) defines thermal comfort as “the condition of the mind that expresses satisfaction with the thermal environment.” [16]. It is often derived from micrometeorological data, such as T_Air, and Mean Radiant Temperature (T_MRT) [17]. T_MRT is defined as the radiant heat exchange between two surfaces, specifically a person and their environment [18]. Despite the challenges of its measurement, including variations in time, instruments, and settings of capture [19], T_MRT is often considered the most important measurement for assessing human thermal comfort within biometeorological studies [20]. These micrometeorological measurements are further used in the calculation of thermal comfort indices, such as the Wet Bulb Globe Temperature (WBGT), the Physiological Equivalent Temperature (PET) [21], and the Universal Thermal Climate Index (UTCI) [22].

While SD-LST provides useful information at broad spatial scales, it is a poor proxy for thermal comfort indices such as UTCI or PET [23,24]. Nevertheless, Landsat SD-LST continues to be widely used as a coarse proxy for outdoor thermal comfort or pedestrian heat exposure in urban heat studies [25,26,27,28], largely because physiologically meaningful variables often cannot be obtained remotely or are not acquired at hyper-local scales [29]. This emphasis on remotely sensed and/or macro-scale measurements often produces unintended consequences. For instance, increasing surface albedo through reflective building materials can reduce SD-LST and even T_Air [30]. Yet, in some instances, these materials often intensify shortwave radiation and raise T_MRT, ultimately exacerbating heat stress and thermal discomfort [31,32].

Therefore, outdoor thermal comfort is largely determined by T_MRT, which in turn is directly influenced by the hyper-local thermal environment, including surface temperatures (Ts), the infrared energy of surrounding materials [33,34,35], and shade, often the most effective way to reduce pedestrian thermal stress [36]. Advances in thermal imaging cameras, whose costs have declined in recent decades, have made it possible to obtain finer-scale resolution and hyper-local estimates of T_S [37], potentially overcoming the challenges still posed by using SD-LST as a measure of thermal comfort. One widely used instrument is the Forward-looking Infrared (FLIR), which measures the longwave infrared radiation that objects self-emit [38]. While this measurement remains a brightness temperature, FLIRs provide an approximate T_S, which has allowed for higher fidelity radiometric measurements of urban heat fluxes and provided a better understanding of T_S on specific patterns of neighborhood-scale urban morphology, such as differences in cooling rates between roofs and walls, as well as underneath tree canopy and shade structures [39,40]. While shade has also been found to be a better predictor than SD-LST in models estimating T_MRT [12], studies that examine individual thermal comfort alongside T_S often occur indoors in climate-controlled chambers; there is a clear need for the use of infrared thermal imaging for outdoor thermal comfort studies [41].

Therefore, while T_S serves as a crucial metric for evaluating both increased urban heat and thermal comfort—particularly in relation to the cooling effects of shade—a disconnect remains within measurement methodologies and urban heat mitigation strategies [42]. Moreover, shade from different sources often has differing effects on T_S, with built structures typically providing greater T_S reduction than tree shade in arid systems [43], yet this pattern does not hold for all climates, with differing types of shade structures sometimes demonstrating synergistic effects [44].

A pressing need remains to study outdoor spaces through the lens of human thermal experience, as this directly shapes how such spaces are used [45]. Bus stops represent a critical setting in which to examine the intersection of pedestrian thermal comfort and thermal dynamics. Extreme heat events can reduce transit ridership, with only the most transit-dependent individuals continuing to ride [46]. While shade remains a key mitigation strategy for urban heat at bus stops [47], bus stops and associated shelter designs understudied facets of transit user experience [48]. Designing thermally comfortable transit stops is challenging, as several factors influence thermal comfort. Local micrometeorology, surface materials, and shade availability via shelters and/or vegetation all shape transit user comfort [49]. Given the complexities between shade, heterogeneous transit infrastructure, and surface material in shaping thermal environments at bus stops, there is a growing need for hyper-local, multi-scalar methods to assess surface temperature and inform more effective interventions. This study explores whether hyper-local thermography offers a practical and cost-effective approach to meeting that need.

Amid calls from urban climatologists to integrate in situ measurements with SD-LST for more accurate assessment of UHI phenomena [6], and from urban ecologists to examine emerging data collection tools into multi-scale urban systems research [50], we ask the following: do simplified, low-cost hyper-local measurements of T_S provide more meaningful spatial and temporal information of pedestrian thermal comfort at bus stops? Our aim is not merely to show that FLIR is more spatially precise than SD-LST, but to examine whether hyper-local Ts can serve as an adequate predictor for thermal comfort in pedestrian-relevant microenvironments. We sought to (1) examine the differences in T_S as measured from Landsat 8 and 9 and from hyper-local FLIR photogrammetry, (2) test how strongly these T_S measurements were correlated with human biometeorology, including metrics such as air temperature (T_Air), and other commonly used indicators of thermal comfort, including T_MRT, WBGT, PET, and the UTCI, and (3) assessed how well FLIR thermography can predict these indices of thermal comfort and highlighted a simple and cost-effective way for cities to analyze T_S at a hyper-local scale.

2. Materials

2.1. Study Locations and Data Sources

Biometeorological and surface temperature data collection occurred across 60 bus stops in the Denver metropolitan area, located in the Colorado Front Range, USA. Denver is a semi-arid city (Köppen Climate Classification: BSk; cold semi-arid climate) situated ~1600 m above sea level. The Regional Transportation District (RTD) manages the metropolitan public transit system, serving approximately 3.08 million residents and recording over 41 million bus boardings in 2023 alone [51]. To examine T_S at bus stops, sites were selected from bus stops with above-average daily ridership for 2022 within the Denver metropolitan area and a 30 m buffer around each site representing a range of land cover, including impervious surfaces, vegetation, canopy cover, and building structure. Land cover data originated from the 2020 Land Use Land Cover dataset from the Denver Regional Council of Governments [52]. In total, 60 bus stops were identified as study sites in the Denver metropolitan region (Figure 1).

Three primary data sources were collected: (1) SD-LST calculated from the thermal bands of Landsat 8 and 9 satellite imagery, (2) biometeorological measurements of the thermal environment collected with a series of Kestrel 5400 sensors (Nielsen-Kellerman Co., Chester Springs, PA, USA), and (3) hyper-local FLIR thermographic images to measure T_S captured using the FLIR C5 compact camera (Teledyne Inc., Thousand Oaks, CA, USA). Sources for satellite data are detailed in the following section, while field sampling is described in the methods section.

2.2. Satellite Derived Land Surface Temperature from Landsat

SD-LST were derived from the thermal band (Band 10) of the Landsat 8 and 9 images. Images were obtained from the United States Geological Survey (USGS) Earth Explorer Collection 2 Level 2 products. Because Landsat 8 and 9 have an ~eight-day revisit coverage, only four usable images were available for the Denver metropolitan area that coincided with the field campaign. Dates, satellite image titles, and percent cloud cover are documented below (

Table 1. Landsat images taken from the United States Geological Survey’s EarthExplorer. Satellite-Derived Land Surface Temperature was calculated from these images.

Date	Image Name in Database	% Cloud Cover (Scene)
19 July 2023	LC09_L2SP_034033_20230719_20230802_02_T1	36.37
27 July 2023	LC08_L2SP_034032_20230727_20230805_02_T1	24.03
28 July 2023	LC09_L2SP_033033_20230728_20230804_02_T1	10.25
4 August 2023	LC09_L2SP_034032_20230804_20230806_02_T1	18.16

).

SD-LST was calculated using the multiplicative scaling factor and additive offset from the Landsat Collection 2 Level 2 product guide [53] with the following equation:

$$\mathrm{S}\mathrm{D}-\mathrm{L}\mathrm{S}\mathrm{T}=\left(\right(\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d} 10 \mathrm{S}\mathrm{T}•0.00341802)+149)-273.15$$

where Band 10 ST is the thermal measurement taken with Landsat’s Thermal Infrared Sensor (TIRS), 0.00341802 is the multiplicative scaling factor, 149 is the additive offset, and 273.15 converts the measurement from Kelvin to Celsius. Calculations were conducted using the Raster Calculator tool (Esri, Inc., Redlands, CA, USA) in ArcGIS Pro (Version 3.5). Thirty-meter buffers were drawn around each bus stop in the study. Using the Zonal Statistics tool, SD-LST was then extracted for these buffers. Thus, 30 m resolution SD-LST measurements were extracted for each bus stop on the four dates that coincided with FLIR image collection.

3. Methods

3.1. Sampling Design and Methodology

Data were collected during an extensive five-week field campaign during the hottest summer months, July and August 2023. Measurements were made twice each weekday—once in the morning and once in the afternoon—during peak commute hours (7:30–10:30, 14:00–18:00). Bus stops were randomly selected for each day of the week, with the goal of obtaining a total of six replications per study site: three in the morning and three in the afternoon.

Hyper-local thermal dynamics of each bus stop were captured with thermographic images and biometeorological measurements. FLIR images were captured from three positions at the bus stop, forming a surface area of approximately 48 m². FLIR image capture coincided with biometeorological measurements using a series of three Kestrel 5400 sensors (Figure 2).

3.2. Biometeorological Measurements

Biometeorological measurements taken with the three Kestrel sensors included air temperature (T_Air), wind speed (V_a), relative humidity (RelHum), Dry Bulb Globe Temperature (T_Globe), and Wet Bulb Globe Temperature (WBGT). The three sensors were calibrated to metric units and positioned 4.8 m apart on tripods at 1.1 m above ground level at each site, following methods designed by Dzyuban et al. [49], to capture the micrometeorology of the bus stop. Sensors were set up to acclimate for five minutes prior to recording. Sensors recorded for two minutes, with their measurements averaged.

Thermal comfort indices, including T_MRT, were then calculated from these Kestrel measurements. T_MRT was calculated using a modified method of the ISO black globe thermometer equation found in Ouyang et al. [54]. This modified method was specifically calibrated for the Kestrel sensor with a different convection coefficient and has the following equation:

$${\mathrm{T}}_{\mathrm{MRT}}={\left({\left({\mathrm{T}}_{\mathrm{Globe}}+273.15\right)}^4+{\displaystyle \frac{\left(0.678\times {10}^8\cdot {\mathrm{V}}_{\mathrm{a}}^{0.019}\right)}{\left(0.95\cdot {150}^{0.4}\right)}}\cdot \left({\mathrm{T}}_{\mathrm{Globe}}-{\mathrm{T}}_{\mathrm{Air}}\right)\right)}^{0.25}-273.15$$

where T_Globe and T_Air are the globe temperature and air temperature in Celsius, respectively, and V_a is the wind velocity in meters per second. The thermal comfort index Physiological Equivalent Temperature (PET) was calculated using the software RayMan Pro (Version 0.1) [55,56], and used inputs from the Kestrel sensors, T_MRT, and self-reported personal and biometric factors from transit users who were willing to report these metrics while waiting for their bus. These inputs included weight, height, and sex, and clothing insulation as calculated by the clothing metric (clo), a metric that assigns values for different articles of clothing. The collection of these data was reviewed and approved by The University of British Columbia’s Behavioural Research Ethics Board under identification code H23-01399. Another thermal comfort index, the Universal Thermal Climate Index (UTCI), was calculated using the R package ‘comf’ (Version 0.1.12) [57] using T_Air, T_MRT, relative humidity, and wind velocity as inputs. A final thermal comfort index, WBGT, was obtained directly via the Kestrel sensors.

3.3. FLIR Image Capture, Segmentation, and T_S Measurements

All three FLIR images were captured facing the bus stop to capture the T_S of the ground and horizontal surfaces, including buildings and bus stop infrastructure. The FLIR C5 thermal camera was set to the standard emissivity of 0.95. Camera positions were standardized as follows: FLIR 1 (F1), oriented to the left of the bus stop; FLIR 2 (F2) positioned in the street facing the bus stop; and FLIR 3 (F3), oriented to the right of the bus stop. F1 and F3 were taken 9.6 m from the stop’s center point (defined as the pole displaying the unique bus stop identification number), while F2 was captured 3 m from this pole (Figure 2). These distances were chosen to approximate areas where most transit users wait. Images were then segmented into polygons using the proprietary software FLIR Thermal Studio Suite (Version 2.0). Segmentation was based on both surface type and camera placement. For F1 and F3, segments encompassed all surface types between the camera position and the central Kestrel (K2). For F2, segmentation included all surface types. Seven surface categories were defined: asphalt, concrete, fine vegetation (herbaceous surfaces such as grass), coarse vegetation (woody tissue of street trees), bare soil, building (walls or fences), and bus stop infrastructure (shelters, poles, benches, and other street furniture). An example of the images and their segmentation can be seen below (Figure 3).

Due to the heterogeneous composition of each bus stop and its associated surfaces, the segmented polygon size was not standardized. From each segmented polygon, an average T_S was determined from the software and was recorded within each image position and for each replicate. If multiple polygons of the same surface types were present within a single FLIR image (e.g., several polygons classified as grass), their T_S values were averaged. This value was defined as FLIR image segment T_S.

From these FLIR image segment (polygon) T_S measurements, a grand mean was calculated for each of the three FLIR images to produce a FLIR image average for a given bus stop. If a surface category was absent from an image (e.g., no coarse vegetation or bare soil at that stop), it was assigned a null value and excluded from the grand mean. Thus, three averages were generated per replicate, one for each FLIR image. Importantly, this grand mean was not the overall pixel-based T_S mean of each FLIR image, provided by the FLIR Thermal Studio, but the average of the defined surface-type segments within a given FLIR image. This approach allowed for comparison across camera positions and assessment of how consistently they measured T_S from their position across bus stops. This value was defined as the FLIR image T_S.

Finally, the average for T_S for each surface type was calculated across all three images. As with the FLIR image T_S, a grand mean of bus stop T_S was then derived by T_S values across all surface types at a given bus stop, with the surface averages representing the mean across the three camera positions. Importantly, this grand mean was not the overall mean of the three FLIR images, but rather an aggregate of surface type averages. This approach allowed assessment of the consistency of T_S measurements across surface types. This value was defined as FLIR bus stop T_S.

3.4. Analysis of FLIR Camera Position and Surface Type for T_S Measurement Consistency, Variable Selection

With multiple FLIR images taken from different camera positions, and numerous thermographic images segmented, we wanted to ensure our method for calculating the T_S of all surface types was consistent over the range of camera positions and segmented polygons at each bus stop. We first examined summary statistics (mean, median, standard deviation, coefficients of variation, and interquartile range) of our T_S sample for each surface type, as well as the average T_S for all FLIR images taken at a stop, to examine the relative variation between mean surface T_S and mean image T_S.

To determine whether we were measuring the T_S of shared surfaces consistently, we calculated intraclass correlations using the two-way random effects model for the mean of k raters (ICC2k). ICC2k treats raters as a random sample from a larger population and estimates the reliability of their average rating. In our case, the “raters” were delineated FLIR image segments (polygons), which varied in size and spatial shading patterns across bus stops. Because these polygons can be considered randomly sampled subdivisions of a heterogeneous surface, ICC2k was appropriate for evaluating the consistency of FLIR camera position across surface types (FLIR image segment T_S). We then calculated ICC3k, a two-way mixed-effects model for the mean of k raters, which assumes that the set of raters is fixed. Here, the raters were the three FLIR camera positions, which were held constant across all bus stops (same distances, same orientations). ICC3k was therefore used to examine the consistency of surface temperature measures across camera positions and to test how reliably the average of these three camera-derived values (one per image) represented the overall bus stop T_S, relative to the grand average of all segmented polygons. Across all cases, we used ICC estimates of consistency rather than absolute agreement, as our focus was on whether FLIR-derived measures covaried reliably across positions and image segments, rather than whether they produced identical values in a heterogeneous thermal environment.

ICCs were run using the package ‘psych’ in R [58]. Assessment of the ICC correlation coefficients, Cohen’s kappa, followed criteria with values less than 0.5 indicating poor consistency, between 0.5 and 0.75 indicating moderate consistency, between 0.75 and 0.9 indicating good consistency, and greater than 0.9 indicating excellent consistency [59]. The metric with the greatest consistency, or the highest Cohen’s kappa correlation coefficient, was selected as our representative variable of T_S values captured by FLIR. The ICC coefficients of this analysis are found in Appendix A. Ultimately, FLIR bus stop T_S was determined to be consistently measured across camera positions and was approved for use in this study.

3.5. Statistical Analyses

A linear mixed effect model was fit using the R package ‘lme4’ [60] to examine differences in Landsat’s T_S (SD-LST) and FLIR T_S. As Landsat captures photos in the morning only, a smaller subset of the data was used: FLIR images captured in the morning that coincided with Landsat’s orbital cycle. Both methods of measurement, FLIR T_S and Landsat SD-LST, were placed as categorical fixed effects predicting bus stop T_S. To account for repeated measurements and clustering, random effects included the date of image capture for both FLIR and Landsat, as well as the unique bus stop ID (BSID) for each study site location [61]. The equation for this model was thus as follows:

$$\mathrm{T}\mathrm{S}\mathrm{i}\mathrm{j}=\mathsf{\beta }0+\mathsf{\beta }1\mathrm{M}\mathrm{e}\mathrm{t}\mathrm{h}\mathrm{o}\mathrm{d}\mathrm{i}\mathrm{j}+\mathrm{b}\mathrm{D}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{j}+\mathrm{b}\mathrm{B}\mathrm{S}\mathrm{I}\mathrm{D}\mathrm{i}+\mathsf{\epsilon }\mathrm{i}\mathrm{j}$$

where TS_ij is the bus stop surface temperature for BSID i on j date, Method_ij is the measurement method (0 = FLIR, 1 = Landsat), β0 is the mean T_S for FLIR, the reference category, β1 is the fixed effect of method, bDate_j and bBSID_i are the random effects for date and bus stop, respectively, and ε_ij is the residual error.

To better understand the sources of variability in surface temperature measurements, we additionally examined the contribution of each component of the mixed-effects model separately. Specifically, we fit FLIR-only and Landsat-only models, including the same random effects for Date and BSID. This allowed us to quantify how much of the total variance was attributable to site-specific differences (BSID), day-to-day variation (Date), and residual measurement error, independently for each measurement method.

FLIR measurements and Landsat measurements were then compared to hyper-local Kestrel measurements and thermal comfort indices with a Pearson product-moment correlation matrix to examine correlations between these two methods and biometeorological measurements. Pearson correlations were assessed first for a significant linear correlation at an alpha of 0.05. Correlations above alpha were said not to be linearly correlated. Pearson correlation coefficients were then assessed for strength. Strongly correlated measurements were determined to be a Pearson correlation coefficient of 0.8 or higher.

We then examined how well the average T_S, as captured by the FLIR, predicted hyper-local measurements of the bus stop, including T_Air and indices of thermal comfort: WBGT, T_MRT, UTCI, and PET. Another series of five linear mixed-effects models was generated, this time predicting these metrics.

$$\mathrm{Y}\mathrm{i}\mathrm{j}=\mathsf{\beta }0\mathrm{Y}+\mathsf{\beta }1\mathrm{F}\mathrm{L}\mathrm{I}\mathrm{R}\mathrm{T}\mathrm{S}\mathrm{i}\mathrm{j}+\mathrm{b}\mathrm{D}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{j}+\mathrm{b}\mathrm{B}\mathrm{S}\mathrm{I}\mathrm{D}\mathrm{i}+\mathsf{\epsilon }\mathrm{i}\mathrm{j}$$

where Y_ij represents one of the five biometeorological metrics (T_Air, WBGT, T_MRT, UTCI, or PET), β0 is the model intercept, β1 is the fixed effect of FLIR T_S, and bDate_j and bBSID_i are the random effects of date and bus stop, and ε_ij is the residual error.

As we were not limited by Landsat’s orbital rotation, the larger data set that coincided with the field campaign was used, with the exception being the model predicting PET, which required biometric information from willing transit users as inputs for the calculation of PET. As not all transit users were willing to report this, the sample size for PET remained lower. Random effects also included study site (BSID) and date; however, after likelihood ratio testing of nesting models, BSID was ultimately dropped as a random effect to avoid issues of singular fit that would occur if left as a random effect in the model: the random effect was too complex for these data, resulting in an overfitted model. Lastly, Root Mean Squared Error (RMSE) values were calculated for each model, both for fixed effects within the model only and for the full mixed-effects model, to see if these models were reasonable for use under ISO 7726 standards for thermal comfort methods, for which an error of less than five degrees Celsius indicates adequate model fit [62]. Significance within all models was evaluated at an alpha (a) of 0.05, and model diagnostics were visualized to check their performance and assumptions.

4. Results

4.1. Description of Collected Data and Bus Stop Structure

The study originally aimed to collect 360 replicates of the 60 study sites (bus stops), 180 each for both the morning and afternoon. However, during summer of 2023, the Denver metropolitan area received increased afternoon precipitation. Compared with the previous decade, it was a historically wet summer and year (Appendix B, Figure A3). Data were not collected during this period of active rainfall, as wet surfaces would confound our FLIR T_S measurements. While all sites saw at least one replicate for both the morning and the afternoon, only 43% of all 60 study sites had a full three replicates in the morning and the afternoon. A total of 93% of all 60 sites had at least two replicates in the morning and the afternoon. After accounting for this, the representative sample of bus stop FLIR thermographic image capture was n = 270 across 19 unique dates. Additionally, only a subset of FLIR data coincided with Landsat’s image capture, as Landsat is limited by its orbital cycles. The representative subsample is n = 66, with 33 measurements of each of Landsat and FLIR across 4 days at 19 unique bus stops, after removing sites obstructed by clouds.

Descriptive statistics for the percentages of land cover composition, captured from 30 m buffers of each bus stop, are presented below (

Table 2. Descriptive statistics for percent land cover composition at each bus stop.

Land Cover	Mean	Stand Dev.	Median	IQR	Min.	Max.
%Impervious	57.26	20.26	56.80	23.23	13.95	98.23
%Tree Canopy	18.55	17.66	13.34	21.82	0	62.72
%Structures	10.60	12.26	8.25	12.44	0	72.87
%Vegetation	29.84	22.57	28.19	30.56	0	85.14

and Figure 4). On average, bus stops were predominantly surrounded by impervious surfaces (Mean = 57.26%, SD = 20.26%), followed by softscapes/vegetation (Mean = 29.84%, SD = 22.57%). In terms of obstruction, the tree canopy was greater (Mean = 18.55%, SD = 17.66%) than built structures (Mean = 10.60%, SD = 12.26%).

4.2. Differences Between Landsat SD-LST and FLIR T_S

Descriptive statistics for differences between T_S as measured by FLIR and Landsat were detailed (

Table 3. Descriptive statistics for bus stop surface temperatures (T_S) measured by FLIR thermal imagery and Landsat, as well as biometeorological metrics. Summary statistics include mean, median, standard deviation (SD), standard error (SE), 95% confidence interval (CI), interquartile range (IQR), minimum, maximum, range, and coefficient of variation (CV) for each variable. FLIR and Landsat represent remote sensing surface temperature measurements, while air temperature (T_Air), Wet Bulb Globe Temperature (WBGT), Mean Radiant Temperature (T_MRT), and Universal Thermal Climate Index (UTCI) capture local biometeorological conditions at bus stops.

Method	Mean	Median	SD	SE	CI	IQR	Min.	Max.	Range	CV
FLIR	25.0	23.8	5.36	0.91	1.77	6.91	15.0	36.7	21.8	0.21
Landsat	35.8	38.7	12.4	2.09	4.10	18.2	4.71	48.8	44.1	0.35
TAir	26.1	26.4	3.27	0.55	1.08	5.32	20.6	32.4	11.7	0.13
WBGT	20.8	21.1	2.59	0.44	0.86	4.27	16.6	25.8	9.12	0.13
TMRT	30.4	31.1	7.44	1.26	2.46	10.00	10.6	42.6	31.9	0.24
UTCI	26.6	27.7	3.82	0.65	1.27	5.77	19.6	33.0	13.4	0.14

). On average, Landsat recorded higher surface temperatures (Mean = 35.8 °C) compared to FLIR (Mean = 25.0 °C), with a notably wider spread in values. The interquartile range (IQR) for Landsat 8 was 18.2 °C—nearly three times that of FLIR (6.91 °C)—indicating greater variability. The standard deviation and coefficient of variation were also higher for Landsat (SD = 12.4 °C, CV = 0.35) than for FLIR (SD = 5.36 °C, CV = 0.21). Furthermore, the minimum and maximum temperatures recorded by Landsat 8 spanned a broader range (4.71 °C to 48.8 °C) than those recorded by FLIR (15.0 °C to 36.7 °C), suggesting Landsat captured more extreme values.

For the biometeorological variables, T_Air had a mean of 26.1 °C with moderate variability (SD = 3.27 °C, CV = 0.13). WBGT averaged 20.8 °C (SD = 2.59 °C, CV = 0.13), while T_MRT exhibited higher variation (Mean = 30.4 °C, SD = 7.44 °C, CV = 0.24), reflecting microclimatic differences at bus stops. UTCI averaged 26.6 °C (SD = 3.82 °C, CV = 0.14), indicating moderate thermal stress.

These results are further visualized (Figure 5). The boxplots show a clear difference between SD-LST as measured by the Landsat satellites and bus stop average T_S as measured by the FLIR. The FLIR displays less variation and tends to be cooler, while Landsat tends to be hotter with greater variance (Figure 5). In addition, median FLIR T_S is closer to other biometeorological metrics, including T_Air, WBGT, T_MRT, and UTCI (Figure 5).

We then further quantified these differences by accounting for date and bus stop as random effects in a linear mixed-effects model (

Table 4. Linear mixed-effects regression model parameters comparing T_S as measured by FLIR and Landsat. On average, Landsat measurements are 10.77 degrees Celsius hotter than FLIR measurements, while the date of capture (τ₀₀ Date) influences variance more than study sites (τ₀₀ BSID). Asterisks (***) indicate significant fixed effects.

Predictors	β	CI	p-Value
FLIR (Intercept)	25.51	19.36–31.67	<0.001 ***
Landsat	10.77	6.81–14.73	<0.001 ***
Random Effects
σ2	64.74
τ00 BSID	7.59
τ00 Date	28.04
ICC	0.35
Marginal R2/Conditional R2	0.227/0.501

). Landsat 8 is on average 10.77 degrees Celsius hotter than the average bus stop T_S measured by the FLIR (Table 4, β). Measurements display a poor intraclass correlation (ICC = 0.35). Additionally, surface temperature varies more by date (τ₀₀ Date = 28.04) than by study site (τ₀₀ BSID = 7.59). Overall, this model explains a moderate amount of variation (Conditional R² = 0.5) in surface temperature between the two methods. Differences in variance in these measurements did not affect the model’s residual variance, with model diagnostics suggesting homoscedasticity.

To examine the variability of our random effects, we then fit linear mixed-effects models to FLIR-only and Landsat-only T_S measurements, including random intercepts for BSID and Date to account for repeated observations (Appendix B,

Table A6. Comparison of FLIR-only and Landsat-only surface temperature (T_S) measurements using linear mixed-effects models. The table reports the mean T_S (β) with 95% confidence intervals (CI), random-effect standard deviations for bus stop (BSID) and date, residual standard deviation (σ), and intraclass correlation coefficient (ICC) for each method. FLIR measurements capture more site-specific variation with lower day-to-day variability, while Landsat measurements exhibit higher day-to-day variability and a higher ICC, indicating that most variance is attributable to clustering by date.

Component	FLIR	Landsat
Mean TS (β, °C)	26.13	36.15
95% CI	23.39–28.87	24.24–48.06
Residual SD (σ)	3.14	6.81
BSID SD (τ00)	4.32	3.43
Date SD (τ00)	1.47	11.28
ICC	0.37	0.73

). FLIR measurements had a mean T_S of 26.1 °C (95% CI: 23.4–28.9 °C), with most variation occurring at the site level (SD_BSID = 4.32 °C) and relatively low day-to-day variation (SD_Date = 1.47 °C). In contrast, Landsat measurements were higher on average (36.2 °C, 95% CI: 24.2–48.1 °C) and dominated by day-to-day variability (SD_Date = 11.28 °C), with smaller site-to-site differences (SD_BSID = 3.43 °C). Residual variation was greater for Landsat (SD = 6.81 °C) than FLIR (SD = 3.14 °C), and the ICC was higher for Landsat (0.73) than FLIR (0.37), indicating that a larger proportion of Landsat variance was explained by clustering, particularly by Date. These results indicate that FLIR captures fine-scale, site-specific variation more effectively, whereas Landsat reflects broader temporal variability.

We then sought to see how closely these measurements were correlated to hyper-local biometeorological measurements, namely T_Air, WBGT, T_MRT, and UTCI. A Pearson correlation matrix was generated using a Pearson product-moment correlation. Significant linear correlations are denoted with asterisks (

Table 5. Pearson product-moment correlation coefficients (r) between surface temperature (Ts) and hyper-local biometeorological measurements Air Temperature (T_Air), Wet Bulb Globe Temperature (WBGT), Mean Radiant Temperature (T_MRT), and the Universal Thermal Climate Index (UTCI). The top row displays correlations derived from the Forward-Looking Infrared (FLIR) method, while the bottom panel shows correlations from the Landsat method. Landsat Ts exhibits weak and non-significant correlations with biometeorological measurements, whereas FLIR Ts demonstrates strong and significant correlations with these variables. Asterisks (***) indicate levels of statistical significance: p < 0.001.

Surface Temperature	TAir (r, Significance)	WBGT (r, Significance)	TMRT (r, Significance)	UTCI (R, Significance)
TS FLIR	0.84 ***	0.84 ***	0.91 ***	0.92 ***
Landsat SD-LST	0.16	0.20	0.26	0.24

).

Landsat’s Ts measurements are more poorly correlated than the hyper-local measurements taken with the Kestrel (Table 5, All r < 0.8, bottom row). Additionally, with insignificant Pearson correlation coefficients, we fail to reject the null hypothesis that Landsat’s correlation to the micrometeorological measurements is equal to zero, thus failing to demonstrate a linear correlation between Landsat’s SD-LST measurements and biometeorological metrics. The FLIR measurements remain much more strongly correlated with these biometeorological measurements, especially thermal comfort indices including T_MRT, UTCI, and WBGT (Table 5, All r > 0.8, top row). With all metrics having significant Pearson correlation coefficients, we can reject the null hypothesis that these correlations are equal to zero. We conclude that the T_S measured by the FLIR is significantly correlated with the human biometeorology at bus stops in a semi-arid system, while the T_S measured by Landsat is not.

4.3. FLIR T_S and Thermal Comfort

We then wanted to determine how well the average bus stop T_S, as measured by the FLIR, could predict hyper-local measurements, particularly indices of thermal comfort for use in future urban heat studies. The results of five linear mixed-effects models predicting T_Air, WBGT, UTCI, PET, and T_MRT are displayed below (

Table 6. Linear mixed-effects regression model parameters for FLIR T_S predicting biometeorological metrics. Asterisks (***) indicate levels of statistical significance: p < 0.001.

Response Variable	Intercept (β [CI])	TS FLIR (β [CI])	p-Value	σ2	τ00 (Random Effect)	ICC	Marginal R2	Conditional R2
TAir	15.45 [14.12, 16.79]	0.43 [0.39, 0.47]	<0.001 ***	4.82	1.87 (Date)	0.28	0.61	0.72
WBGT	13.73 [12.63, 14.83]	[0.29 [0.25, 0.32]	<0.001 ***	4.31	0.29 (Date)	0.06	0.50	0.53
UTCI	14.79 [13.47, 16.12]	0.47 [0.43, 0.51]	<0.001 ***	5.73	0.82 (Date)	0.13	0.66	0.70
PET	12.69 [7.24, 18.14]	0.58 [0.41, 0.74]	<0.001 ***	10.91	7.56 (Date)	0.41	0.51	0.71
TMRT	11.54 [9.17, 13.92]	0.75 [0.67, 0.82]	<0.001 ***	21.73	1.32 (BSID)	0.06	0.58	0.60

). Representative sample for T_Air, WBGT, UTCI, and T_MRT was n = 270, while for PET it was n = 47.

All hyper-local measurements showed significant positive correlations with the bus stop average T_S as measured by the FLIR (Table 6, p-values). The strongest relationship was between T_S and T_MRT, for which a per-unit increase in T_S results in a 0.75 increase in T_MRT (Table 5, β). In addition, this model is notable for being the only one out of the five for which the study sites contributed the most variance as a random effect instead of date (τ₀₀ BSID). Other indices of thermal comfort, including UTCI and PET, also showed significant relationships with bus stop average T_S, 0.47 for UTCI and 0.58 for PET (Table 6, β). In addition, both UTCI and T_Air explained the largest amounts of variance as fixed effects, 0.66 for UTCI and 0.61 for T_Air (Table 6, Marginal R²).

RMSE values for assessment under ISO 7726 were then calculated for both fixed effects and for the entire mixed-effect models. Regression results are further visualized below (Figure 6). All models display RMSE values under 5 degrees Celsius, for both fixed effects (blue) and full effects (red), indicating adequate fit for thermal comfort studies under ISO 7726.

5. Discussion and Limitations

5.1. Landsat SD-LST as a Hotter and Weak Thermal Comfort Correlate

This study demonstrated that the T_S, as measured from the Landsat 8 and 9 orbitals (SD-LST), is on average 10.7 degrees Celsius hotter than the T_S measured from the FLIR C5, and that it is not significantly correlated with other hyper-local measurements, including indices of human thermal comfort. This corroborates recent calls for urban heat studies to move beyond SD-LST for accurate accounting of pedestrian thermal comfort [12,23,24,28]. It seems that ground-based, hyper-local investigations of urban heat remain crucial to our understanding of human biometeorology.

This study utilized the method outlined by the USGS Level 2 Science Products Guide [53] for SD-LST calculation; however, it is worth noting that other satellites, methods, product levels, and equations exist for its quantification [1,2]. We are not precluding the potential usefulness of these alternatives in assessing pedestrian thermal comfort. However, with a majority of UHI studies utilizing Landsat [9], it seems that the USGS Level 2 Science Product 30 m SD-LST metric is not useful for hyper-local pedestrian thermal comfort analysis. Future studies could compare these alternatives to FLIR photogrammetry and other biometeorological measurements to see how they differ. For now, this calculation method seems incongruous with the current needs of outdoor pedestrian thermal comfort inquiry.

Critically, we did not examine why there are differences in T_S measured from Landsat and the FLIR C5. There are two notable inconsistencies between the methods of measurement. The first is temporal. Landsat’s image capture occurs between 10:00 and 10:30 a.m., with our FLIR images being captured between 7:30 and 10:30 a.m., resulting in slight incongruities at the time of image capture. The other notable difference between the two methods is resolution; Landsat is resampled to 30 m, and our FLIR area is roughly 10 m. While investigating reasons for the difference in T_S measurements was not a question posed in this study, we wish to highlight these limitations and speculate on some sources of these differences.

Most importantly, the size of a Landsat thermal pixel relative to a bus stop introduces substantial spectral mixing, where multiple surface types (e.g., roadway, sidewalk, vegetation, built structures, shade) are aggregated into a single 30 m pixel. This mixed-pixel effect likely produced regression attenuation simply due to the scale mismatch between Landsat and FLIR. Namely, given the propensity of bus stops to be located near major vehicle roadways, these impervious surfaces were likely partially located within Landsat pixels, which could result in higher SD-LST temperatures. Future studies examining factors that make up this difference, including land cover metrics [63], as well as different compositions of shade may yield helpful insight into microclimatic variability [43].

Additional contributors to the mismatch may include differences in emissivity assumptions between Landsat and the FLIR sensors, as well as the presence or absence of shade structures at the time of image capture, both of which can shift T_S estimates. More explicit examination of these factors would help explain if the divergence between FLIR and Landsat is largely a scale- and physics-driven issue rather than purely methodological. Although methods exist to refine the resolution of Landsat’s coarse computation of SD-LST [64], Landsat still remains limited by its orbital cycle and often obstructive cloud cover [4,5]. Furthermore, the SD-LST from Landsat is increasingly recognized as being a poor proxy for measures of thermal comfort [28], such as UTCI and PET [23,24], and may need to be validated with in-situ measurements when used for this application [15].

Therefore, despite the uncertainty in differences between these measures of surface temperature, we ultimately conclude that T_S captured by the FLIR offers greater insights into the micrometeorology, shade, and vertical representations of a hyper-local thermal environment. Coupled with its ability to measure the T_S of outdoor spaces as they are being actively used, we conclude that FLIR has more utility than Landsat for analyzing the T_S of hyper-local thermal environments, particularly semi-arid transit systems.

5.2. FLIR Photogrammetry and Pedestrian Thermal Comfort

This study yielded significant results in hyper-local FLIR thermal imagery’s ability to predict both modeled indices of thermal comfort, such as T_MRT and UTCI, and demonstrates the ability of consumer-grade FLIR technology to be effective in the needs for thermal comfort studies under ISO 7726. Conversely, SD-LST measurements taken with the Landsat were weakly correlated with these hyper-local FLIR measurements, and, therefore, solely relying on SD-LST continues to be ineffective at portraying human thermal experience [12,13,65].

Previous studies have demonstrated this relationship with more complex methods. Middel et al. [65] utilized a “PanoMRT” system, by which T_S recorded alongside the MaRTy mobile weather station, documenting both shortwave and longwave radiation from six directions. It was found that the six-directional FLIR imagery from the PanoMRT system predicted UTCI best, with the lowest RMSE, but that PET and T_MRT are also well-predicted, with their models also falling under the five-degree threshold put forth by ISO standards for thermal comfort studies. The system’s ability to predict these metrics outperformed RayMan, a conventional method, ultimately postulating that its ability to measure longwave radiation fluxes is a key to its success.

Our study yielded similar results with simpler methods and instruments. Three-directional FLIR photogrammetry predicted both T_MRT, PET, and UTCI with similarly adequate RMSE values (4.53 for T_MRT, 2.94 for PET, and 2.33 for UTCI). While [65]’s methods utilized the MaRTy mobile weather station, measuring both longwave and shortwave radiation for more accurate measurements of T_MRT, we still achieved significant results with the less costly use of a black globe thermometer. While we contend that the MaRTy mobile weather station and its ability to measure radiation directly are likely superior to black globe thermometers and their ever-expanding list of T_MRT equations, we still posit that simpler methods are adequate for use under ISO standards. Both instruments used in this study, the Kestrel 5400 and FLIR C5 thermal camera, are consumer-grade products and adequate for non-research applications. It seems that the FLIR C5 camera, which only captures longwave radiation, can also predict human thermal comfort indices with adequate accuracy.

Stewart et al. [12] also highlighted the need for hyper-local T_S measurements of multiple surfaces, namely those that cannot be captured by the 2-D photography on a satellite, such as walls and other vertical surfaces. Proposing a method called T_ped, they outlined a need for weighted averages of these surfaces to capture human thermal experience. While this study largely took this approach, the surface types were still averaged, resulting in a loss of variance among surface types. Future studies could consider controlling for this loss of variance and select study sites based on the composition of surface materials to see which surfaces provide lower T_S as well as lower indices of thermal comfort, especially since different surface and ground cover types are known to influence T_MRT [33,34,35,66], which in turn has implications for the calculations of thermal comfort indices.

In further terms of variance, Date continued to account for more variance than BSID in nearly all thermal comfort models, the notable exception being when FLIR T_S predicted T_MRT (Table 4 τ₀₀ BSID). While this speaks to the possibility of daily meteorological conditions and extreme heat events driving thermal comfort over bus stop structural heterogeneity, it was noteworthy that this pattern did not hold for all thermal comfort indices, namely one that is difficult to quantify [19]. While a different formulation than the modified equation used in this study might yield different results, the variance captured by this site-level random effect warrants further exploration, especially given that past urban heat mitigation strategies have overlooked this variable and prioritized the use of SD-LST to inform interventions [31].

Similar approaches are also found in the use of ENVI-met (Version 4.4.4), which creates 3-D simulations to determine surface temperatures based on various urban materials [67]. However, it too faces challenges when representing vertical mixing of radiative heat transfer [68]. With inaccurate measures of longwave radiation from the ground suspected as a contributor to these challenges [69], perhaps this simpler method could provide useful measurements for ground-based surface temperatures used in ENVI-met simulations.

Lastly, in the systematic literature review outlined by Wu et al. [41], there is scant literature linking outdoor T_S from thermal imagery to subjective thermal perception. When it does occur, participants’ facial T_S is measured rather than the T_S of the surrounding environment. Given the linkages among FLIR T_S and indices of thermal comfort, future research could consider measuring T_S with the FLIR alongside subjective thermal perception, i.e., thermal comfort surveys, to see if FLIR technology can also predict the subjective experience of the thermal environment.

5.3. Limitations

A limitation in this study’s use of FLIR thermal cameras is the potential uncertainty introduced by radiometric assumptions inherent to infrared thermography. Thermal cameras measure upwelling longwave radiation and convert this to a radiometric temperature using an assumed surface emissivity (a constant 0.95 in this study). Spatially heterogeneous emissivity assignments (for example, across different materials or coatings) can bias derived brightness temperatures and flux estimates if not explicitly measured or corrected [70]. Although we accounted for different bus stop surface heterogeneity by segmenting our FLIR images, specific corrections to emissivity values for certain materials could be made in future studies.

In addition to emissivity uncertainty, urban surfaces exhibit directional (anisotropic) thermal emission, meaning that apparent radiometric temperature varies with sensor viewing angle because of the three-dimensional geometry and orientation of surface facets. This anisotropy arises from differences in the proportion of sunlit and shaded surfaces within the sensor’s field of view, compounded by variable emissivities across materials and orientations [71,72]. While we hope we accounted for some anisotropic bias by positioning three FLIR cameras at different directions towards the center of the bus stop, this method cannot eliminate anisotropic uncertainty in its entirety. Continuing to document and test emissivity assumptions for each material, quantifying residual anisotropy after multi-view averaging, and incorporating geometric or radiative modeling to correct for directional effects are needed to refine similar temperature studies in complex urban environments.

6. Conclusions

This study examined the differences in T_S at bus stops from two measurement methods: macroscale SD-LST from the Landsat 8 satellite and hyper-local, FLIR thermography. In addition, it linked T_S derived from FLIR photogrammetry to some commonly used thermal comfort indices to examine its use in predicting human thermal experience, including T_MRT, PET, and UTCI.

We demonstrated that T_S, as measured by Landsat, is on average 10.7 degrees hotter than FLIR measurements. Additionally, FLIR measurements are strong and significantly correlated to the micrometeorological measurements of the bus stop (r > 0.8, p < 0.001), while Landsat measurements have no significant correlations. Lastly, the average T_S measured by the FLIR was able to explain over 50% of the variation in T_Air, WBGT, UTCI, PET, and T_MRT. With these models having RMSE values below five degrees Celsius, segmented FLIR image averages are adequate for use in thermal comfort studies under standards put forth by the ISO.

Ultimately, we find that this novel method of utilizing thermal image photogrammetry is sufficient as a simple method and low-cost alternative for analyzing the T_S of bus stops, overcoming some of the challenges of scale posed by the Landsat satellites, which were demonstrated to be insignificantly correlated to hyper-local biometeorological measurements. Continuing to examine why these differences exist, along with other measurement methodologies for capturing SD-LST, would be helpful for advancing analyses derived from satellite imagery. In sum, hyper-local thermographic images are effective at predicting indices of human thermal comfort at bus stops, offering potential solutions for prioritizing heat-resilient transit design in semi-arid transit systems.