A Spatially Explicit Satellite-Derived Surface Urban Heat Island Database for Urbanized 1 Areas in the United States : Characterization , Uncertainties , and Possible Applications 2 3

The urban heat island (UHI) effect is strongly modulated by urban-scale changes to the aerodynamic, thermal, and radiative properties of the Earth’s land surfaces. Interest in this phenomenon, both from the climatological and public health perspectives, has led to hundreds of UHI studies, mostly conducted on a city-by-city basis. These studies, however, do not provide a complete picture of the UHI for administrative units using a consistent methodology. To address this gap, we characterize clear-sky surface UHI (SUHI) intensities for all urbanized areas in the United States using a modified Simplified Urban-Extent (SUE) approach by combining a fusion of remotely-sensed data products with multiple US census-defined administrative urban delineations. We find the highest daytime SUHI intensities during summer (1.91 ± 0.97 °C) for 418 of the 497 urbanized areas, while the winter daytime SUHI intensity (0.87 ± 0.45 °C) is the lowest in 439 cases. Since urban vegetation has been frequently cited as an effective way to mitigate UHI, we use NDVI, a satellite-derived proxy for live green vegetation, and US census tract delineations to characterize how vegetation density modulates inter-urban, intra-urban, and inter-seasonal variability in SUHI intensity. In addition, we also explore how elevation and distance from the coast confound SUHI estimates. To further quantify the uncertainties in our estimates, we analyze and discuss some limitations in using these satellite-derived products across climate zones, particularly issues with using remotely sensed radiometric temperature and vegetation indices as proxies for urban heat and vegetation cover. We demonstrate an application of this spatially explicit dataset, showing that for the majority of the urbanized areas, SUHI intensity is lower in census tracts with higher median income and higher proportion of white people. Our analysis also suggests that poor and non-white urban residents may suffer the possible adverse effects of summer SUHI without reaping the potential benefits (e.g., warmer temperatures) during winter, though establishing this result would require future research using more comprehensive heat stress metrics. This study develops new methodological advancements to characterize SUHI and its intra-urban variability at levels of aggregation consistent with sources of other socioeconomic information, which can be relevant in future inter-disciplinary research and as a possible screening tool for policy-making. The dataset developed in this study can be visualized at: https://datadrivenlab.users.earthengine.app/view/usuhiapp.

algorithm have been independently validated against both observational and theoretical estimates 185 of SUHI intensity (Manoli et al., 2020a;Niu et al., 2020), there is debate regarding whether it 186 constitutes a 'true' rural reference (for an extended discussion, see Chakraborty and Lee (2019)). 187 For the purposes of this study, however, SUHI intensity is the average LST difference between 188 the average built-up pixel and the average non built-up pixel within each urbanized area. While a 189 similar method of delineating urban and rural references would not work for CUHI, this is primarily 190 due to the stronger effect of advection on T a compared to LST. Similar to the algorithm used for 191 SUHI, we also calculate urban-rural differentials in NDVI (ΔNDVI) and DEM (ΔDEM) for each 192 agglomeration. To examine the suitability of using NDVI as a proxy for vegetation, we calculate 193 the urban-rural differential in tree cover percentage (ΔTree Cover) for each urbanized area from   (Table S1). Note that the slightly higher SUHI values in the present study are due to primarily two 221 Urbanized areas in temperate and boreal climates show larger seasonal variations. Arid zones 232 also show the lowest intra-urban spatial variation in daytime SUHI intensity for all cases (Fig. 3). 233 During nighttime, urbanized areas in arid climate show the highest SUHI intensity (Fig. 3b) (Table 1). During nighttime, when one 247 would expect coastal areas to have relatively higher temperatures, summer SUHI intensity is 248 actually higher for non-coastal urbanized areas (0.53 ± 0.31 °C for coastal; 0.62 ± 0.25 °C for non-249 coastal). This difference is not due to a sampling issue since we essentially analyze all urbanized 250 areas, as defined by the US census bureau. While it is possible to extend this analysis to the 251 'urban areas', which the US census bureau defines as regions with a population of less than 252 50,000 people, some of these tend to be very small, with few census tracts. The limited size and 253 intra-area variation limits the both the ability to obtain sufficient representative pixels to reliably 254 calculate SUHI intensity and to conduct analysis regarding its relationship with socioeconomic 255 variables. 256 We find daytime SUHI intensity and the urban-rural differential in NDVI (ΔNDVI), a proxy for live 271 green vegetation, to be negatively correlated both within and between urbanized areas (Fig. 4), 272 except for the boreal climate. These correlations are especially strong during summer, which is 273 expected due to higher potential evaporative cooling from vegetated surfaces during this season 274 (Manoli et al., 2020a

SUHI Intensity and Distance from the Coast 293
We examine the coastal influence on SUHI intensity by calculating the mean and standard 294 deviation of the correlation coefficients (after Fisher's z transformation and back-transformation) 295 between the distance of the census tract centroids from the nearest coast and the annual, 296 summer, and winter SUHI intensities (Table 2). This analysis is only done for the 110 census tract 297 groups adjoining the coast. On average, the correlation coefficients are negative for both daytime 298 (-0.09 ± 0.42 for annual) and nighttime (-0.5 ± 0.43 for annual). The strong negative correlations 299 are expected during nighttime due to the thermal inertia of water. We examine the correlation 300 coefficients between distance from the coast and ΔNDVI to resolve the seemingly counter-intuitive 301 decreasing daytime SUHI with distance from the coast. For all cases considered, ΔNDVI is 302 positively correlated with distance from the coast (around 0.28 ± 0.33 for all cases). This means 303 that for the coastal urbanized areas in the US, vegetation density tends to increase farther from 304 the waterfront, thereby counteracting the coastal influence on SUHI. Partial correlations that 305 account for the ΔNDVI variability gives us slightly positive correlation coefficients between SUHI 306 intensity and distance from the coast, at least for the annual and summer cases. It should be noted 307 that isolating the influence of coastal advection on UHI intensity is much more complicated than 308 can be inferred from the bulk statistical analysis performed here (Steneveld et al., 2011); and 309 requires considerations of wind speed and direction, land-sea thermal gradients, and other factors 310 beyond the scope of the present study. 311 312

Census-tract Elevation: A Possible Confounding Factor 313
Since temperature varies with altitude, comparing UHI intensities at different elevations is not 314 ideal. The UHI literature typically accounts for this limitation by setting elevation differential 315 thresholds for entire cities (in multi-city analysis) or for individual pixels before calculating SUHI. 316 For illustration, we examine the relationship between SUHI intensity and the urban-rural elevation 317 differential (ΔDEM) for each urbanized area (Fig. S2). The elevation differential is indeed 318 important, showing a negative correlation with SUHI intensity for a slight majority of the urbanized 319 areas considered. While there is not as much inter-seasonal trend, roughly two-thirds of urbanized 320 areas (316 for year, 320 for summer, and 342 for winter) demonstrate this negative correlation, 321 confirming that census tracts with a higher average elevation have lower temperature. The 322 negative correlations are slightly lower at night. Note that while elevation is an unwelcome 323 confounder when dealing with SUHI intensity itself, it is less problematic from a human welfare 324 perspective. Since it is not necessarily true that higher elevation areas will not be inhabited, using 325 such elevation thresholds in the present study would mask out entire census tracts or large parts 326 of the population who live in the higher elevation regions of the urbanized areas. Therefore, with 327 the aim of consistent assessment of the local distribution of SUHI as a bulk parameter, we do not 328 use elevation thresholds, acknowledging that this omission leads to some uncertainties in 329 urbanized areas with large terrain gradients. 330 SUHI; Fig. 6 shows the patterns for summer). Overall, white is the only racial group for which the 351 mean correlation between SUHI intensity and proportion of population is negative, while the mean 352 positive correlation is highest for the black racial group. These patterns persist even after 353 accounting for income, as seen from the distribution of partial correlation coefficients between the 354 two variables (Fig. 6b). For winter nights, the association between SUHI intensity and race 355 practically disappears (Fig. 6c), especially after accounting for income (Fig. 6d). In general, the highest percentages of available data are over arid urbanized areas since they 399 are relatively cloud free, with the lowest percentages over boreal and tropical climates. While 400 this distribution is consistent for both 8-day composites and daily scenes, the percentage of 401 available LST data are much lower at the daily scale. Note that missing data are due to both 402 cloudy pixels and the 3 °C uncertainty limit specified during quality control. We generally expect 403 similar percentages of valid pixels across the different climate zones for NDVI. 404 405 While the use of 8-day composites instead of daily scenes could also lead to biases in our SUHI 406 estimates (Hu and Brunsell, 2013), we find surprisingly strong correlations between SUHI 407 intensities calculated from the two levels of temporal aggregation, with r 2 over 0.90 and the slope 408 of the linear fit close to 1 in most cases (Fig. 7). Exceptions include winter daytime and annual 409 nighttime, with the largest deviations seen for the boreal climate. Noting that the mean 410 percentage of valid urban pixels for winter daytime for the boreal climate is only 17.9% (39.6% 411 for annual nighttime) when using the daily scenes (66.6% when using 8-day composites; 94.3% 412 for annual nighttime), we are more confident in the representativeness of the 8-day composites 413 for calculating clear-sky SUHI estimates. Low missing data in the daily LST product in Table S3  414 (for instance, in the arid zone) is also a good proxy for regions and seasons for which our clear-415 sky estimates would approach the true LST climatology. This variability in representativeness 416 across seasons and climate zones should be kept in mind when using this dataset. 417

418
The use of NDVI as a proxy for vegetation cover may be inaccurate, particularly during winter 419 and for coastal regions, due to the influence of water bodies, snow cover, and clouds. These between ΔNDVI and median income (Figs 7a, 7b, and 8c), implying richer urban residents live in 497 'greener' census tracts. However, for coastal urbanized areas, we see weaker correlations 498 between ΔNDVI and median income (r=0.28 ± 0.30 for coastal and -0.45 ± 0.29 for non-coastal 499 urbanized areas for the year; r=0.27 ± 0.30 for coastal and -0.46 ± 0.30 for non-coastal 500 urbanized areas for summer), which is not surprising since ocean-adjacent census tracts, which 501 tend to have less vegetation cover (Table 2), generally house richer populations. 502 503 We separated the difference in summer and winter NDVI for the low-income tracts (below 25 504 percentile of income) and high-income tracts (above 75 percentile of income) for each urbanized 505 area (Fig. 9d). We find that this mean difference (of summer NDVI-winter NDVI) is greater in 506 high income tracts for temperate and boreal climate zones (p<0.01), but not for arid and tropical 507 climate. This heterogeneity is due to the stronger vegetation phenology in temperate and boreal 508 climate due to the larger abundance of deciduous trees and shrubs. Similar values in the 509 difference in summer and winter NDVI in both low and high-income tracts for tropical and arid 510 cases explain the practically non-varying relationships between daytime SUHI intensity and 511 median income for urbanized areas in these climate zones. Similarly, the difference between 512 summer and winter NDVI is significantly (p<0.01) higher for white-dominant tracts (over 75% 513 white residents) than non-white dominant tracts (under 25% white residents) for temperate and 514 boreal climate. 515 516

Implications 517
The UHI is not an additional environmental stressor due to urbanization under all circumstances, 518 since in some cases, especially in boreal climate and winter nights, a higher temperature may be 519 preferable (Yang and Bou-Zeid, 2018). As we note from Fig. 5, the negative association 520 between SUHI and median income is much weaker at night, practically disappearing during 521 winter. For many US urban areas, since we can reasonably assume that the UHI has primarily 522 negative health effects during summer days and primarily positive health effects during winter 523 nights, our results imply that poor people may be suffering the adverse effects of UHI without 524 reaping the potential wintertime benefits. This result holds for race as well, with lower potential 525 SUHI intensity for white-dominant census tracts during summer days and a relatively even 526 distribution of SUHI intensity regardless of race for winter nights (Fig. 6). It is important to note 527 however, that verifying the possible health connotations of these trends requires using more 528 comprehensive metrics than LST. While Laaidi et al. (2012) found nighttime LST to be 529 associated with increased mortality during a heatwave period, it should be noted that T a , which 530 is more relevant to public health, is more strongly coupled with LST at nighttime, both within bearing the economic burden of UHI during both seasons, an aspect that could be further 538 explored in comparative analysis based on an initial screening using the tool presented in this 539 paper. With reference to these economic consequences, the SUHI, which is heavily influenced 540 by roof and wall temperatures, is also more directly relevant. 541 542 Evident from Fig. 9, seasonal trends in SUHI disparity are particularly strong for boreal and 543 temperate urbanized areas in the US. It remains to be seen whether these patterns would be 544 consistent for T a (and thus CUHI), and urban heat stress. For the overall spatial disparities 545 however, since CUHI also tends to be higher for the urban core references. While this is ideal since SUHIs are primarily due to changes in the physical 558 characteristics of the land surface, the mismatch between physical boundaries and 559 administrative boundaries makes comparisons between and within cities difficult. Here we use a 560 fusion of remotely-sensed products and multiple administrative boundary definitions to 561 characterize the intra and inter-city variation in the annual, summer, and winter SUHI intensities 562 during daytime and nighttime in the US. We find that SUHI intensity is negatively correlated with 563 income and percentage of white population for the vast majority of the urbanized areas. 564 Moreover, poorer and non-white urban residents tend to be exposed to higher summer daytime 565 SUHI, when heat stress would be at its maximum, and similar winter nighttime SUHI, when 566 poorer urban residents could potentially benefit from higher ambient temperatures. Since SUHI 567 intensity, its seasonality, and spatial variability are strongly associated with the degree of 568 vegetation cover in and within urbanized areas, strategically placing urban parks and green 569 spaces can be a useful way to reduce both the mean SUHI, as well as its spatial variability. The 570 dataset created in this study can be accessed through the web application 571 https://datadrivenlab.users.earthengine.app/view/usuhiapp , and companion data set 572  for urban pixels for the cases and climate zones considered in the present study.