Transfer learning reveals large discrepancies between air and land surface temperatures in cities Authors: Yiwen Zhang, Lei Zhao, TC Chakraborty, Priyam Mazumdar, Keer Zhang, Pierre Gentine Published: 2026 ABSTRACT Understanding of urban weather and extreme events remains severely limited by data poverty resulting from a dearth of true urban weather stations. As a result, land surface temperature (Ts), obtained from remote sensing platforms, has been widely used as a stand-in for near-surface air temperature (Ta) despite their fundamental differences, especially in urban areas. Although Ts provides important scientific insights and practical utility for studying the urban thermal environment, this substitution risks introducing large uncertainties and biased characterization of impact-relevant urban heat stress. Here we develop an urban transfer-learning framework (U-TL) to address this critical gap and to provide urban high-resolution air temperature (U-HAT) data at large scales across the contiguous United States (CONUS). U-TL demonstrates high accuracy and strong robustness in predicting urban Ta, even with limited training data. The resulting U-HAT is a high-resolution urban Ta dataset capable of accurately reproducing observed and well-established urban climatology. U-HAT reveals substantial Ts–Ta discrepancies and therefore cautions the use of Ts to characterize urban heat. We show that satellite-measured Ts substantially overestimates both urban heat stress magnitude and intra-city spatial variability, which have consequential implications for urban heat exposure, vulnerability, and adaptation policy making. Introduction Cities, despite being widely recognized as global centers of climate hazards and exposure1–3, remain severely underrepresented in terms of in situ meteorological observations. This gap stems primarily from the scarcity of research-grade weather stations in urban areas4,5 (Fig. 1). Standard weather stations are disproportionately placed in non-urban landscapes (Fig. 1a, b). Even those stations that are classified as “urban” (see Fig. 1c–g) are most often located at airports or adjacent open fields, thereby failing to capture the true meteorological and climatological characteristics of the built environments6–8. A key reason for this misrepresentation lies in the World Meteorological Organization (WMO) siting guidelines, which stipulate weather stations to be installed on open, grass-covered ground and far away from built obstructions9. While the WMO guidelines aim to ensure consistent, uncontaminated measurements for capturing broader weather and climate signals, adherence to these standards effectively precludes the placement of conventional weather stations in true built environments, leaving only peripheral locations such as airports or city outskirts as apparent viable options10. Consequently, very few stations are situated in core urban areas worldwide. Even in the rare occasions where stations do exist in true built environments, their sparse presence (e.g., just one or two stations per city) fails to capture the inherent large intra-city heterogeneity of urban landscapes11. This pervasive urban data poverty has been a major obstacle to urban climate research for decades, hindering observation-based studies, compromising validation efforts for climate and weather models, and limiting data assimilation in historical reanalysis products and weather forecasting systems, particularly for urban areas12.Fig. 1Limited true urban representation in global weather stations.a Meteorological stations in the Global Historical Climatology Network Hourly97. Stations located within urban boundaries as defined by the Global Urban Boundaries dataset98 are shown as brown dots, while those outside are in green. Blue stars indicate selected stations that are considered “urban” based on this broad classification. b Proportion of meteorological stations categorized by impervious surface fraction within a 200-meter buffer. Impervious surface fraction is based on the “built-up” category in the ESA WorldCover dataset99. c–g Satellite images from Google Earth (http://earth.google.com) showing the surrounding environments of selected stations. Basemap from Natural Earth (https://www.naturalearthdata.com/). Limited true urban representation in global weather stations. a Meteorological stations in the Global Historical Climatology Network Hourly97. Stations located within urban boundaries as defined by the Global Urban Boundaries dataset98 are shown as brown dots, while those outside are in green. Blue stars indicate selected stations that are considered “urban” based on this broad classification. b Proportion of meteorological stations categorized by impervious surface fraction within a 200-meter buffer. Impervious surface fraction is based on the “built-up” category in the ESA WorldCover dataset99. c–g Satellite images from Google Earth (http://earth.google.com) showing the surrounding environments of selected stations. Basemap from Natural Earth (https://www.naturalearthdata.com/). Land surface temperature (Ts), a bulk radiative skin temperature of objects on the ground as estimated from a satellite view, has therefore been commonly used as a stand-in for near-surface air temperature at the screen height (Ta) to study the urban thermal environment in recent years13–17. In many cases, findings based on Ts have been used to discuss implications for an amorphous “urban heat”, which can often be confused with those from Ta18. Satellite-measured Ts offers several advantages over sparse weather stations, including global coverage, consistency through measurements from the same satellite sensor, and spatial continuity19,20, making it particularly suitable for large-scale studies and detailed intra-urban analyses. Recent advances in drone sensing can also provide on-demand, meter-to-submeter-level Ts observations that capture very fine-scale urban heterogeneity at both high spatial and temporal resolutions21. These Ts-based applications provide important scientific insights and practical utility in urban climate studies22,23. For example, Ts is directly related to the urban radiative thermal environment, provides the lower boundary condition for the atmosphere above, and directly modulates the surface energy balance and land-atmosphere interactions. However, unlike Ta, which is strongly correlated with human-perceived heat stress, Ts does not directly reflect physiologically relevant heat hazard and is thus subject to critical limitations when drawing impact- or risk-related implications such as those related to public health or urban adaptation policies24. Indeed, Ts itself does not inherently represent human-perceived thermal conditions or other impact-relevant variables (e.g., heat-related health risks or energy use). Therefore, studies relying solely on Ts to infer urban climate impacts and adaptation strategies have to assume – explicitly or implicitly – that either (i) Ts has a strong spatial correlation with Ta within cities, or (ii) urban-rural difference in Ts (surface urban heat island or UHIs) is strongly correlated with urban-rural difference in Ta (canopy urban heat island or UHIa). However, multiple recent observational studies in diverse cities or regions have demonstrated that neither of the two assumptions is justified6,24,25. Ts and Ta, while related, are conceptually and physically distinct26. Their relationship is influenced by the complex interplay of many factors, including land surface properties (e.g., vegetation cover, imperviousness, and aerodynamic roughness), sky conditions (e.g., water vapor and cloud cover), and boundary layer processes26–29, especially in highly heterogeneous urban landscapes. This raises significant concerns about the validity of using Ts as a proxy for Ta, as this substitution can lead to substantial uncertainties and systematic errors in the assessments of urban heat stress and climate impacts. For example, Ts generally shows stronger modifications due to urbanization than Ta or more comprehensive heat stress indices, thereby leading to quantitative and sometimes qualitative errors in such assessments. Recent observational-based attempts to address this critical urban data gap have primarily focused on two approaches: crowdsourced citizen weather station (CWS) measurements30,31 and empirical data-driven methods to infer urban Ta32,33. The former has emerged as a promising avenue for local and regional urban observational studies due to the recent proliferation of CWS networks24,34–36. Their measurements, however, are subject to two critical limitations. First, CWS measurements are prone to substantial errors, biases, and uncertainties due to improper station placement (e.g., on rooftops or near HVAC [Heating, ventilation, and air conditioning] exhausts), lack of calibration and maintenance, instrument or internet malfunctions, and inconsistent metadata31,37,38. Second, CWSs are subject to uneven distribution both within cities and across regions39,40. They are, by design, primarily installed in more conveniently accessible locations (e.g., backyards of residential buildings) and often in more affluent neighborhoods or regions, leading to sampling biases that undermine the representativeness of their measurements. These factors significantly limit the reliability of CWS data and complicate their use for global or large-scale urban climate studies. On the other hand, data-driven approaches aim to empirically infer urban Ta using standard observations and/or auxiliary parameters. Broadly, these methods fall into two categories: spatial interpolation from standard weather stations41–43 and statistical/machine learning techniques to predict urban Ta44–51. Both, however, are fundamentally constrained by the overall dearth of stations in true urban environments. The former often produces interpolated Ta datasets substantially biased toward non-urban landscapes due to their much higher station density, resulting in substantial inaccuracy in urban areas32. The latter, particularly data-intensive deep learning techniques, struggle to establish robust relationships between predictors and urban Ta due to insufficient representative training samples from urban cores. Many of these methods have to rely heavily on Ts as a predominant input feature. Consequently, their Ta products often fail to reproduce observed urban climate signals, for example, the well-established reversal of diurnal patterns for UHIa and UHIs, preventing their applications in urban climate research52. To address this critical data poverty issue, we present a framework based on the idea of transfer learning, termed U-TL (Supplementary Fig. S1a), which provides urban high-resolution air temperature (U-HAT) data at large scales. U-HAT is developed at 1-km resolution, covering 384 of the largest cities in the contiguous United States (CONUS) from 04-01-2013 to 12-31-2023 twice a day, at 1:30 AM and 1:30 PM local time. Comprehensive validation of the U-HAT data demonstrates strong performance in capturing urban-specific climate signals and gradients with high fidelity. Unlike existing methods, our framework effectively reproduces observed urban climatology both within and across cities, addressing a longstanding barrier in urban climate research. Future releases of U-HAT will extend coverage to all urban areas globally. Combining U-HAT with Ts observations, we provide a comprehensive assessment of urban Ts–Ta differences on a pixel-to-pixel level at the continental scale. We find large discrepancies between air and land surface temperatures in cities in both the magnitude and intra-city spatial variability in urban heat exposure. Results U-TL: an urban transfer-learning framework for air temperature reconstruction We overcome the fundamental barrier of sparse true urban weather stations by leveraging a transfer-learning approach, a machine learning technique that addresses data scarcity by adapting knowledge from data-rich, pretrained models to data-limited tasks through “fine-tuning”53. The U-TL framework consists of two main steps (Supplementary Fig. S1a; “Methods” section). In the first step, we pretrain a deep neural network (DNN) model to predict urban Ts, a data-abundant target due to extensive satellite observations, as a function of meteorological forcings and surface properties. This pretrained model is then fine-tuned with representative, ground-based true urban Ta observations to estimate urban Ta in the second step. Both Ta and Ts are essentially governed by urban land-atmosphere interactions. U-TL leverages this shared physics by initially learning a generalized representation of the land-atmosphere interactions driving urban Ts in the first step, and then transferring that knowledge to a different, yet related, task of estimating Ta via fine-tuning in the second step. This physical basis distinguishes U-TL from previous statistical models that directly map Ts to Ta based solely on empirical correlations, which often break down in data-scarce and spatially heterogeneous urban environments. Because U-TL acquires most of its “understanding” of urban land-atmosphere interactions in a data-rich environment (i.e., using urban Ts labels), only a small amount of true urban Ta data is required during the fine-tuning step for accurate prediction, making the approach both efficient and scalable. In other words, the purpose of predicting Ts in the first step is to pretrain the model with a data-rich proxy so that the knowledge gained can be transferred to the target task of estimating Ta. Urban land-atmosphere interactions are primarily characterized by meteorological forcings and surface properties (Supplementary Fig. S1b). To emulate these interactions, the DNN in the pretraining step ingests the atmospheric forcing fields and urban surface properties as predictors (Supplementary Table S1) and outputs the corresponding urban Ts (“Methods” section). For the atmospheric forcings, we include all the fields commonly required to drive an urban land model in climate or weather models54,55. The urban surface properties are represented using satellite imagery, surface elevation, and building height. This design allows the DNN to predict urban Ts in a manner reminiscent of climate or weather modeling, but “solving” for Ts “statistically” rather than “numerically”, as the model takes the same input variables as a numerical model would require for solving a system of equations54. In the second step, the pretrained DNN is further fine-tuned using urban Ta observations from 52 weather stations that were carefully selected based on our site-selection criteria (“Methods” section). These stations are sourced from 18 cities across diverse climate regimes (Supplementary Fig. S2). We conduct a thorough validation of U-TL against unseen observations and established urban temperature climatology (“Methods” section). U-TL exhibits high accuracy in predicting urban Ta across scales (Supplementary Fig. S3), and this strong overall predictive power is achieved with very limited urban training labels. It is able to reproduce the observed diurnal, daily, and spatial variabilities of urban Ta (Supplementary Fig. S4). Importantly, such performance remains highly robust under increased data scarcity (Supplementary Fig. S5) – a challenge mirroring the real-world extreme scarcity of true urban observations. U-HAT reproduces observed urban climatology As discussed above, inconsistencies with observed urban climatology have long hindered the widespread adoption of existing Ta datasets in urban climate and weather studies. Many such products, including most reanalysis datasets56,57, fail to capture urban signals such as the UHI effect. Even those specifically designed to incorporate urban signals often fail to reproduce observed spatiotemporal patterns of UHIa. For instance, a majority of observational studies have concluded that UHIa is typically stronger at night than during the day, whereas UHIs often show the opposite6,20,58,59. However, many existing urban Ta datasets reveal a contradictory diurnal relationship, featuring a stronger UHIa during daytime rather than at night52. This discrepancy largely arises from the fact that those datasets derive urban Ta directly from urban Ts, effectively making them “magnitude-adjusted Ts” products. Here we compared the diurnal and seasonal UHIa patterns in U-HAT against two recent Ta datasets45,46 as illustrative examples. We employed a Simplified Urban Extent (SUE) method60 (“Methods” section), which defines UHI as the difference in temperatures between the urban and non-urban pixels within each urban extent. We acknowledge that using non-urban pixels inside urban boundaries might introduce exaggerated “rural” reference temperatures. To ensure the robustness of our findings, we repeated these analyses using the PRISM data61 and ERA5 reanalysis62 as rural reference temperatures (“Methods” section). Our results show that U-HAT consistently captures the observed UHIa trends remarkably well across calculation methods. With U-HAT, the nighttime UHIa intensity is consistently more pronounced than during the day (Fig. 2a, Supplementary Fig. S6). For example, the average summer (winter) daytime UHIa across CONUS cities shown by SUE is 0.15 K (0.14 K), which is 0.71 K (0.31 K) lower than nighttime values. In contrast, the average summer (winter) UHIs, measured by Ts is 1.55 K (0.72 K) higher during the day than at night. Observational studies have also noted that UHIs tend to be higher in summer and lower in winter, whereas UHIa does not necessarily63. These diurnal and seasonal differences are, however, misrepresented in the two existing Ta datasets—Yao et al.45 and Zhang et al.46 Using SUE, both datasets show no significant diurnal differences, and consistently report higher summer UHIa than winter. With PRISM and ERA5, they both demonstrate stronger UHIa during daytime than at night and a greater summer peak than winter, a diurnal and seasonal pattern largely reflecting that of UHIs. For instance, Yao et al. 45 and Zhang et al. 46 show a 0.95 K (0.66 K) and 0.85 K (0.60 K) higher average summer (winter) daytime UHIa than nighttime using PRISM, respectively. In other words, UHIa patterns demonstrated by U-HAT, which align with long-term documented observations, are not captured by these existing Ta datasets.Fig. 2U-HAT accurately reproduces observed urban climatology compared to existing Ta (near-surface air temperature) datasets.a Mean UHI (urban heat island) intensities across cities in the CONUS (contiguous United States) during daytime and nighttime in JJA (June–August) and DJF (December–February) from 2013–2020 calculated using U-HAT, Yao et al.45, Zhang et al.46, and Ts. Error bars indicate the standard error of the mean. b Distributions of correlation coefficients (r) from linear regressions between Ta, Ts (land surface temperature), and the Normalized Difference Vegetation Index (NDVI) across CONUS cities. Ta estimates are from U-HAT, Yao et al.45, and Zhang et al.46. Each data point is computed using JJA mean values of variables from 2013–2020, based on all available pixels within one cluster. Center bars represent the median, box edges the 25th and 75th percentiles, and error bars extend to 3 × IQR (interquartile range) from the quartiles. Shaded violin plots indicate the underlying distribution. U-HAT accurately reproduces observed urban climatology compared to existing Ta (near-surface air temperature) datasets. a Mean UHI (urban heat island) intensities across cities in the CONUS (contiguous United States) during daytime and nighttime in JJA (June–August) and DJF (December–February) from 2013–2020 calculated using U-HAT, Yao et al.45, Zhang et al.46, and Ts. Error bars indicate the standard error of the mean. b Distributions of correlation coefficients (r) from linear regressions between Ta, Ts (land surface temperature), and the Normalized Difference Vegetation Index (NDVI) across CONUS cities. Ta estimates are from U-HAT, Yao et al.45, and Zhang et al.46. Each data point is computed using JJA mean values of variables from 2013–2020, based on all available pixels within one cluster. Center bars represent the median, box edges the 25th and 75th percentiles, and error bars extend to 3 × IQR (interquartile range) from the quartiles. Shaded violin plots indicate the underlying distribution. Recent studies using CWSs also observed relatively weak and diverse Ts–Ta correlations across cities, cautioning the use of Ts for urban heat assessment and mitigation recommendations24. These studies have further indicated that the Normalized Difference Vegetation Index (NDVI) had significantly weaker associations with Ta than with Ts. U-HAT reproduces these relationships well (Fig. 2b). Across CONUS cities, U-HAT shows relatively small yet variable Ts–Ta correlation coefficients (r) of 0.51 ± 0.28 (mean ± standard deviation), in broad agreement with recent studies over European cities using CWS observations24. In contrast, Yao et al.45 and Zhang et al.46 exhibit substantially high and more spatially consistent correlations (r = 0.81 ± 0.15 and 0.86 ± 0.16, respectively). Similarly, the Ta–NDVI correlation from U-HAT (r = −0.33 ± 0.31) is notably weaker than in the Yao et al.45 and Zhang et al. 46 datasets (r = −0.43 ± 0.31 and −0.52 ± 0.24, respectively). Collectively, these results suggest that existing Ta datasets fail to capture true urban Ta signals and instead largely reflect Ts-driven patterns. This shortcoming likely stems from the heavy reliance on urban Ts as a primary predictor in those Ta data products. Large skin-to-air temperature discrepancies U-HAT provides an unprecedented opportunity for pixel-to-pixel level comparisons with satellite-derived Ts, enabling characterization of urban human-perceived thermal variations both within and across cities. Previous Ts–Ta comparisons were primarily based on station observations on non-urban surfaces, revealing substantial differences between these two temperatures27,64. However, similar analyses in urban environments remain limited. Even among the few urban case studies, inconsistencies arise due to measurement biases and representativeness issues25. Here, by combining U-HAT and Ts estimated from the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor onboard NASA’s Aqua satellite (MYD11A1), we examined the 9-year average urban vertical temperature gradient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left(\Delta T={T}_{s}-{T}_{a}\right)$$\end{document}ΔT=Ts−Ta at 1 km resolution, as well as the intra-city variability of urban Ta vs. Ts \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left(\Delta {\sigma }_{T}={\sigma }_{{T}_{s}}-{\sigma }_{{T}_{a}}\right)$$\end{document}ΔσT=σTs−σTa across CONUS cities. Our results reveal substantial spatiotemporal discrepancies between Ts and Ta in urban areas, reinforcing that Ts alone can be a highly biased proxy for Ta when assessing urban heat hazard. We observe strong diurnal and climatic signals in the urban vertical temperature gradient (Fig. 3a, b). Across all available pixels in 384 CNOUS urban clusters, ΔT is, on average, positive during the day (5.01 ± 3.10 K) and negative at night (−1.58 ± 0.96 K). This means that using Ts in place of Ta can substantially overestimate the urban diurnal temperature range (DTR), an important metric of climate change. Hence, caution is warranted when drawing implications of urban effects on DTR using satellite Ts. These city-average ΔT signals exhibit a distinct climatic and geographic pattern: substantially larger (positive or negative) ΔT magnitudes occur in dry climates (desert and semiarid regions of CONUS) than in wetter temperate climate zones (e.g., eastern CONUS). This spatial pattern complies with previous observations over non-urban landscapes27, with cities showing more intra-city variation in ΔT (Fig. 3a, b). For example, in Seattle (Phoenix, Dallas, New York), the range of daytime ΔT spans approximately 20 K (10 K, 13 K, 17 K, respectively), and nighttime ΔT is about 11 K (4 K, 4 K, 6 K).Fig. 3U-HAT demonstrates substantial differences in magnitude and spatial variability between Ts (land surface temperature) and Ta (near-surface air temperature) over 384 cities in the CONUS (contiguous United States).Map of the city-wise mean vertical temperature gradient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left(\Delta T={T}_{s}-{T}_{a}\right)$$\end{document}ΔT=Ts−Ta averaged across 9 years at daytime (a) and nighttime (b). For each map, pixel-wise ΔT in Seattle (c, g), Phoenix (d, h), Dallas (e, i), and New York (f, j) are shown below. The frequency distribution of pixel-wise ΔT is also shown. Difference between the city-wise standard deviation of 9-year mean of Ts and Ta \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left(\Delta {\sigma }_{T}\right)$$\end{document}ΔσT at daytime (k) and nighttime (l). Basemap from Natural Earth (https://www.naturalearthdata.com/). U-HAT demonstrates substantial differences in magnitude and spatial variability between Ts (land surface temperature) and Ta (near-surface air temperature) over 384 cities in the CONUS (contiguous United States). Map of the city-wise mean vertical temperature gradient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left(\Delta T={T}_{s}-{T}_{a}\right)$$\end{document}ΔT=Ts−Ta averaged across 9 years at daytime (a) and nighttime (b). For each map, pixel-wise ΔT in Seattle (c, g), Phoenix (d, h), Dallas (e, i), and New York (f, j) are shown below. The frequency distribution of pixel-wise ΔT is also shown. Difference between the city-wise standard deviation of 9-year mean of Ts and Ta \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left(\Delta {\sigma }_{T}\right)$$\end{document}ΔσT at daytime (k) and nighttime (l). Basemap from Natural Earth (https://www.naturalearthdata.com/). These intra-city and intercity ΔT variabilities can be primarily attributed to two factors: background climate and urbanization. In dry regions, with less water available for evaporation, more net radiation is partitioned into sensible heat flux rather than latent heat, leading to reduced evaporative cooling and weaker land-atmosphere coupling, thereby increasing ΔT. In addition, lower moisture availability of vegetated surfaces further reduces thermal inertia and heat capacity, causing the land surface to heat up more quickly during the day and cool more quickly at night compared to the air above27. However, these climatic signals are diminished at more “urbanized” landscapes (Supplementary Fig. S7a, b). When considering only pixels with impervious surface fraction (ISF) greater than 0.5 instead of all available pixels, ΔT increases more significantly in wet climates but remains largely unchanged in dry climates, narrowing the difference between them. To further examine the effects of urbanization on ΔT, we categorized pixels into 10 equally spaced bins based on their ISF and compared the average ΔT within each bin. We find that Ts and Ta are strongly decoupled over more urbanized surfaces. Specifically, highly impervious surfaces (ISF > 0.9) show 9.10 K higher daytime ΔT and 1.25 K lower nighttime ΔT than highly pervious surfaces (ISF < 0.1) (Supplementary Fig. S7c, d). This is primarily due to the thermal properties and energy partitioning toward sensible heat flux of impervious surfaces, which resemble those of dry natural surfaces as described above. These similarities between dry and impervious surfaces also explain why ΔT changes less with increasing ISF in dry climates. In these regions, transitioning from dry to impervious surfaces has little impact, whereas in wet climates, the conversion from moist soils or vegetated surfaces to impervious materials results in a much greater reduction in evaporative cooling and surface heat capacity and thus leads to a large increase in ΔT. This further cautions the use of Ts as a proxy of Ta over highly urbanized surfaces where Ts and Ta are largely decoupled, regardless of the background climate zones, especially during daytime. We also find a large discrepancy in intra-city variability between Ta and Ts. Across the CONUS, a majority of cities (99.2% at daytime and 69.8% at nighttime) show larger \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma }_{{T}_{s}}$$\end{document}σTs than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma }_{{T}_{a}}$$\end{document}σTa (Fig. 3c, d). The smaller variability of Ta within cities reflects the role of air advection and horizontal mixing in redistributing heat citywide, an effect that minimally influences Ts. Such an effect is especially evident during the daytime. At night, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma }_{{T}_{s}}$$\end{document}σTs and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma }_{{T}_{a}}$$\end{document}σTa become closer, likely due to weaker air mixing. These differences suggest that using Ts might misrepresent human-perceived heat hotspots in urban areas. Large skin-to-air temperature discrepancies in heat exposure This large Ts–Ta discrepancy in intra-city variability is further amplified when examining population-scale exposure to urban heat extremes. To illustrate this, we compared daytime extreme heat exposure measured by Ta (HEa) and by Ts (HEs) across each city. We define extreme heat exposure as the total number of days exceeding the city’s 99th percentile of temperature over the current entire U-HAT timespan (2013–2023), multiplied by the exposed population65 at each 1 km pixel. We find that HEs show much wider intra-city distributions than HEa across CONUS cities (Fig. 4, Supplementary Fig. S8a). Specifically, 96.9% of the cities examined show a larger coefficient of variation for HEs than for HEa. This discrepancy is often more pronounced in major, populated U.S. metropolitan areas, as illustrated in Fig. 4a, b. Numerous recent studies have utilized high-resolution satellite data to examine urban environmental disparities or inequities16,17,66,67. However, our results indicate that these Ts-based estimates of intra-city heat spatial variability are substantially overestimated, consistent with patterns previously identified using process-based modeling68. Insights derived from these measures, therefore, may not translate directly to effective, equitable policy. We also find that HEs distributions are commonly more skewed toward higher ends than those of HEa, as evidenced by a greater difference between their mean and median. 87.5% of the cities show higher skewness for HEs than HEa (Supplementary Fig. S8b), suggesting that not only intra-city spatial variability but also the overall magnitude of extreme heat exposure are likely overestimated when relying on satellite-observed Ts.Fig. 4U-HAT demonstrates substantial differences between daytime extreme heat exposure measured by Ts (HEs) and Ta (HEa) in their magnitude and spatial variability over 384 cities in the CONUS (contiguous United States).The map shows pixel-wise differences between daytime HEs and HEa. a–h Distributions of HEa and HEs in selected cities on a log scale. Center bars represent the median, box edges the 25th and 75th percentiles, and error bars the 1st and 99th percentiles. Basemap from Natural Earth (https://www.naturalearthdata.com/). U-HAT demonstrates substantial differences between daytime extreme heat exposure measured by Ts (HEs) and Ta (HEa) in their magnitude and spatial variability over 384 cities in the CONUS (contiguous United States). The map shows pixel-wise differences between daytime HEs and HEa. a–h Distributions of HEa and HEs in selected cities on a log scale. Center bars represent the median, box edges the 25th and 75th percentiles, and error bars the 1st and 99th percentiles. Basemap from Natural Earth (https://www.naturalearthdata.com/). This overestimation also displays an important intra-city pattern: Ts tends to overestimate heat exposure in urban cores while underestimating it in peripheral neighborhoods. Both HEa and HEs indicate that heat exposure is often higher in city centers (Supplementary Fig. S9), likely due to the combined effects of greater population density and elevated temperatures. However, their spatial distributions differ significantly. HEa typically shows a less steep spatial gradient than HEs within a city, encompassing both central and suburban areas – a pattern evident in Chicago, Atlanta, Dallas, and other major cities. The more uneven distribution of HEs – from the city center to its periphery – is shaped by a multiplicative effect of Ts and population. Ts has a more localized spatial pattern than Ta, as its redistribution is much less affected by air advection and horizontal mixing. This effect is further amplified by the uneven distribution of population, resulting in a much more intensified pattern of HEs than HEa in densely populated, high-temperature areas. These results suggest that relying on Ts alone risks biased policy interventions that focus disproportionately on city centers while overlooking broader, dispersed populations in outlying residential areas who are also dangerously exposed to extreme heat stress. In the U.S., these often include retired and elderly individuals who are highly vulnerable to heat17. Ta-based analyses are therefore essential in ensuring effective and equitable resource allocation, strategy development, and decision-making for urban heat mitigation efforts and climate actions. Nevertheless, we note that U-HAT is not uncertainty-free. While more impact- and exposure-relevant than Ts, U-HAT remains subject to uncertainties tied to limited training domains and spatial transferability. These uncertainties should be considered when interpreting equity and policy implications. Discussion In this work, we developed a transfer-learning framework, U-TL, to reconstruct urban high-resolution air temperature. Accurate urban Ta estimation has long been challenging due to scarce true urban observations and strong urban heterogeneity. U-TL addresses these fundamental barriers and provides the urban Ta dataset capable of accurately reproducing observed urban climatology. Notably, U-TL achieves high predictive power as a “global” model, meaning it forgoes any local empirical tuning or spatially varying parameters. Importantly, U-TL demonstrates the possibility of using data-rich proxies to support accurate prediction of data-poor variables in highly heterogeneous environments. This framework is not restricted to air temperature prediction or to urban systems. With appropriate pretraining targets and urban observations, the framework can be trained to predict other sparsely measured urban variables such as humidity, precipitation, and radiative or turbulent fluxes. Nor is data poverty an issue unique to urban landscapes; many other regions around the world, such as the Global South, also significantly lack quality observations. Similar transfer-learning models can be adapted for those contexts or systems, such as rivers and lakes, wildlands, croplands, or other vegetated ecosystems, where in situ data are scarce. The 1-km U-HAT dataset produced by U-TL fills a critical urban data gap that has hindered urban climate and weather research for decades, and provides an impact-relevant thermal characterization both within and across cities at a large scale, thus advancing our understanding of the urban thermal environment. Due to the absence of high-fidelity, high-resolution, and spatially continuous urban Ta data, satellite-derived Ts has been widely used, often informing impact-, decision-, or policy-oriented applications under the assumption that Ts (or UHIs) strongly correlates with Ta (UHIa). U-HAT enables direct evaluations of these assumptions at the pixel-to-pixel level. Our findings reveal substantial discrepancies between Ta and Ts both within and across cities. These discrepancies are further exacerbated in urban heat exposure. These results underscore the criticality of Ta-based measures in urban studies and have consequential implications for urban heat mitigation. For instance, although numerous remote sensing studies confirm the effectiveness of vegetation-based strategies in reducing urban Ts, such cooling efficacy does not necessarily translate to equivalent air temperature reductions. Compared with Ts, the much weaker correlation (r = −0.33 vs. −0.56) (Fig. 2b) and lower sensitivity (−1.28 K vs. −13.26 K) of urban Ta to NDVI in U-HAT (Supplementary Fig. S10) suggest substantially lower cooling potential of urban greenery for air temperature. This cooling benefit for human-perceived heat stress is likely further dampened when air humidity is factored in refs. 24,69–71. That said, Ts estimates in urban climate studies are still important as they provide globally consistent metrics for mechanistic understanding of urban thermal environments and how it relates to the surface energy balance23,72,73. Moreover, as also seen from our U-TL workflow, it remains a critical input for state-of-the-art algorithms that can more accurately capture physiologically relevant urban climate signals. U-HAT opens promising new opportunities for longstanding challenges stemming from the shortage of spatiotemporal urban-resolving Ta data. For example, weather forecasting models and data assimilation systems are known to integrate little to no urban ground observations, and this dataset could potentially improve weather prediction and reanalysis products over urban areas significantly. In large-scale climate and Earth system modeling, U-HAT can provide spatiotemporally consistent urban Ta data to evaluate, benchmark, and constrain urban climate simulations. Furthermore, many urban epidemiological and socioeconomic studies have long been constrained by the lack of high-quality, high-resolution urban air temperature data16,74. U-HAT can provide a crucial input feature for applications in these fields. We note three main limitations of this study. First, the resolution and accuracy of our final data are limited by the uncertainty and resolution of satellite data. The spatial resolution of the current U-HAT is 1 km as it is constrained by the resolution of MODIS Ts. However, one could easily train U-TL to produce higher-resolution Ta when a higher-resolution satellite Ts is used. Part of the U-HAT uncertainty is inherited from the uncertainty of Ts, which is less than 2 K (“Methods” section). However, because the transfer-learning approach extracts generalizable features from the pretrained task rather than relying on exact pixel values, the impact of the Ts uncertainty on U-TL’s performance is likely mitigated, especially when our pretrained task is sufficiently general, i.e., trained on monthly mean Ts instead of daily values (“Methods” section). Second, the model currently provides estimations only for non-cloudy days, as the availability and accuracy of satellite-derived data under cloudy conditions remain limited. However, the large variability of the training data encompasses a wide range of atmospheric and surface conditions, presumably enabling the model to partly generalize beyond clear-sky scenarios. This suggests that the framework has the potential to predict Tₐ under cloudy conditions, although larger biases may be introduced. As more advanced methods are being developed to enhance spatial and temporal resolutions, improve retrieval accuracy, and reconstruct missing observations of satellite products, these challenges can likely be addressed by leveraging higher-resolution, higher-accuracy, and gap-filled datasets as they become available. Third, the physical relationship between Ts and Ta is variable and context-dependent across diverse urban morphologies and climates. Our validation results demonstrate strong generalizability and spatial transferability of U-TL within CONUS. U-TL nevertheless provides an approximation of the underlying complex, variable Ts–Ta mapping, and thus its accuracy may be sensitive to input uncertainties. Methods The U-TL framework The U-TL framework (Supplementary Fig. S1a) leverages a transfer-learning approach that adapts knowledge learned from data-rich tasks to improve performance on data-limited tasks53,75,76. We achieve this by initially training a deep neural network (DNN) to predict urban land surface temperature (Ts), and then fine-tuning it to estimate urban near-surface air temperature at the screen height (Ta) in the second step. Step 1: Pretraining to estimate urban Ts In the first step, we trained the model to estimate Ts at each urban pixel using its corresponding surface properties (i.e., satellite imagery, building height, and elevation), atmospheric forcing fields, and month indicators as predictors. The backbone of the model architecture is a DNN modified from a 2-layer ResNet77 concatenated with a separate set of linear layers. The DNN ingests satellite imagery and elevation as inputs, and its output is average-pooled to generate a feature vector. The input satellite imagery includes blue (SR_B2), green (SR_B3), red (SR_B4), near infrared (SR_B5), shortwave infrared 1 (SR_B6), normalized difference vegetation index (NDVI), and normalized difference built-up index (NDBI) bands. In addition, we applied three linear layers to extract another feature vector from forcing fields, building height, and month indicators. The atmospheric forcing fields include surface solar radiation downward (ssrd), surface thermal radiation downward (strd), surface pressure (sp), total precipitation (tp), u-component of wind (u), v-component of wind (v), temperature (t), and relative humidity (r) at a blending height. The two feature vectors are concatenated and passed through three linear layers to mimic the land-atmosphere interactions for the final Ts prediction. The Ts data is obtained from the Aqua MODIS daily land surface temperature product (MYD11A1) at 1 km resolution, retrieved twice a day at approximately 1:30 PM and 1:30 AM local time for daytime and nighttime, respectively78. Note that the U-TL framework itself is not limited to Aqua samples and can be applied to Terra data seamlessly. Here, Aqua is chosen over Terra to generate urban Ta data because its overpass times better correspond to daily air temperature extremes. We only retained pixels that are labeled with an estimated Ts error of less than 2 K according to the quality control bands and included only days with at least 90% available pixels to minimize cloud contamination effects. For each sample, satellite imagery and elevation at 30 m resolution are cropped into 33 × 33-pixel patches, centered on the corresponding Ts grid cell. All other input variables are represented as scalars. In transfer learning, coarse-graining is often applied to the pretraining task to improve its generalizability for the subsequent target task75. Therefore, we used monthly-averaged Ts and forcing fields instead of daily values for model training. Two separate models were trained to predict daytime and nighttime Ts. The nighttime model takes in one fewer input from the forcing fields than the daytime model, as downward surface solar radiation is absent at night. Step 2: Transfer learning to estimate urban Ta In the second step of U-TL, we fine-tune the pretrained Ts model using urban ground Ta observations. To optimize the knowledge transfer, we conduct several experiments to determine the best configuration. While model performance remains stable across structural adjustments, we selected the overall best-performing model. First, while the initial layers of the Ts model can resolve generalized representations of the land-atmosphere interactions, the final layers are more task-specific and may be less useful for Ta prediction due to Ts–Ta differences. We therefore experimented with progressively unfreezing or discarding layers of the pretrained Ts model from the final layer upward. Here, “Unfreezing” refers to allowing the weights of pretrained layers to be updated during fine-tuning, while “discarding” removes them entirely. Optimal configurations were determined separately for daytime and nighttime models. For both, unfreezing all layers proved most effective, likely due to the need to adjust for temperature distribution differences between Ts and Ta. In the daytime model, the last three linear layers of the Ts model were discarded and replaced with three new layers containing fewer parameters, initialized with random weights. For the nighttime model, only the last layer was discarded and replaced with a randomly initialized one. Second, we found that larger Ts–Ta differences were associated with higher residual errors during fine-tuning, particularly in the nighttime model, suggesting that greater temperature differences may hinder effective knowledge transfer. To mitigate this issue, we introduced a set of linear layers as skip connections79 using Ts as input to learn the residuals from the transfer-learning step (Supplementary Fig. S1a). We found that this implementation improved performance, particularly for the nighttime model. Third, we incorporated a curriculum learning-inspired loss function to improve training stability and the learning of hard samples for both the pretraining and fine-tuning steps. Curriculum learning prioritizes easier samples at the beginning of training and gradually increases the influence of harder ones as the model improves80. We implemented this by dynamically adjusting sample importance based on their loss81. By initially downweighing high-loss samples, the model first captures fundamental patterns, leading to better feature extraction for complex cases later. This prevents the model from being overwhelmed early and ensures that harder samples are learned more effectively over time. Data preprocessing Urban extent Urban boundaries are defined using the TIGER/Line Files and Shapefiles provided by the United States Census Bureau. Cities are ranked by their sizes, and the largest 400 are initially selected. After excluding 16 cities located outside the Contiguous United States (CONUS), the final sample includes 384 cities. Station data preparation For the fine-tuning step, we carefully screened 52 stations from 18 cities with hourly Ta records (Supplementary Fig. S2). We applied a set of selection criteria to ensure the representativeness of true urban environments, though some subjectivity remains. Each meteorological network was scrutinized to verify that the screen height is at 1.3–3 m above the surface. Generally, stations were selected if they were located on surfaces with an impervious surface fraction (ISF) greater than 0.5 within a 200 m buffer, and we also visually inspected station locations and surrounding environments using Google Earth (http://earth.google.com). Airport urban stations were only included when they were within or close enough to cities. The selected stations cover diverse landscapes and climate zones. To match the sampling time of Ts, we calculated mean Ta values between 1 PM and 2 PM local time as daytime Ta, and mean Ta values between 1 AM and 2 AM as nighttime Ta. Stations came from Synoptic data, Integrated Surface Database (ISD), and local sources. Auxiliary datasets All auxiliary datasets are summarized in Supplementary Table S1. Satellite imagery is from the Landsat 8 surface reflectance product82. Although the model can infer vegetation and built-up features from the available bands, we explicitly calculated and included NDVI and NDBI (Eqs. 1 and 2) as inputs, as prior testing showed that their inclusion enhances model performance.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${NDVI}=\frac{{SR}\_B5-{SR}\_B4}{{SR}\_B5+{SR}\_B4}$$\end{document}NDVI=SR_B5−SR_B4SR_B5+SR_B42\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${NDBI}=\frac{{SR}\_B6-{SR}\_B5}{{SR}\_B6+{SR}\_B5}$$\end{document}NDBI=SR_B6−SR_B5SR_B6+SR_B5Where SR_B4, SR_B5, and SR_B6 are red, near infrared, and shortwave infrared 1 bands, respectively. Clouds and cloud shadows are masked using the quality control bands. We converted raw values into surface reflectance and masked pixels exceeding the radiometric saturation range (0–1) to remove physically unrealistic values, which can result from Landsat’s retrieval issues over water bodies and snow surfaces. As urban surface characteristics (except vegetation phenology) are generally static over a short period, for input satellite imagery, we use the monthly median composites over the corresponding five-year period (2013–2018 or 2019–2023) for each training sample. This choice is considered reasonable because urban expansion, particularly in U.S. cities, typically occurs on decadal timescales83. However, this assumption may not hold universally. In regions experiencing rapid urban development over short timescales, the use of multi-year composites could obscure temporal evolution in urban characteristics and potentially introduce bias into Ta predictions. In such cases, yearly composites may be more appropriate, albeit at the cost of insufficient cloud-free surface imagery availability. To test the effect of the composite window length on model performance, we conducted a sensitivity analysis comparing model performance when using 3-year versus 5-year monthly median composite surface imagery in both the Ts pretraining and Ta fine-tuning steps. For this analysis, we trained both models using data from the 18 cities with available station observations to ensure a fair comparison. We find negligible differences in model performance: the daytime MAE is 0.93 K for the 3-year composite compared to 0.90 K for the 5-year composite experiment, while the nighttime MAE is 1.22 K for both cases (Supplementary Fig. S11). To further assess the sensitivity of the resulting Ta estimates to shorter composite windows, we examined the temporal characteristics of estimates derived from models using 1-year, 3-year, and 5-year monthly median composite surface imagery as inputs. The results show highly consistent temporal patterns across all three composite windows (Supplementary Fig. S12). The mean bias differences between the 1-year and 3-year composites and between the 3-year and 5-year composites are −0.3 K and −0.5 K, respectively. Together, these results indicate that the choice of composite window length (up to 5 years) has minimal influence on the resulting Ta estimates, demonstrating the robustness of the model to this design choice over the CONUS region. We used only images without masked pixels for model training and data production, which results in an average of 84.4% available samples across the entire dataset spanning 384 cities. For building height84, we applied gap-filling with the mean height of neighboring 1.5 pixels. We reprojected Landsat imagery, elevation85, and building height data to the Ts projection, and then preprocessed them using the Google Earth Engine platform86. The atmospheric forcing fields are from the ERA5 hourly reanalysis dataset. Specifically, the variables u, v, t, and r are at the 1000 hPa level87, and ssrd, strd, sp, and tp are at the surface level88. Precipitation values were aggregated daily, whereas all other variables were extracted at Aqua’s overpass times in the local time zone. At the forcing height (i.e., blending height in the atmospheric boundary layer), the air is well blended, and thus the influence of fine-scale urban structure on the forcing meteorology at this height is largely diminished. The rationale of using coarse large-scale atmospheric forcings as inputs in U-TL aligns with standard methodological structure in climate modeling, where coarse-resolution atmospheric forcings (such as ERA5 reanalysis or GCM outputs) are routinely used to drive high-resolution land surface models or WRF54,55. Therefore, the fine-scale variability captured in U-HAT arises from the interaction between large-scale atmospheric forcing and fine-scale surface characteristics. To ensure that these forcing variables are compatible with the 1-km output temperatures and can be consistently applied to the model, we re-gridded all forcing fields from 0.25° to 1 km using a “patch” method provided by xESMF89. Validation of U-TL Validation of the Ts model We randomly selected approximately 35 million samples from 70% of each city’s data for training the Ts model. We drew an equal number of samples from each city to ensure sample balance. To compensate for scarcer data in smaller cities, we applied a bootstrap stratified sampling method in which existing samples from data-scarce cities were stochastically resampled during training as needed. An additional 3.5 million samples were randomly selected from the remaining 30% for testing. Model validation on the test data demonstrates high accuracy in predicting urban Ts (Supplementary Fig. S13), with mean absolute errors (MAEs) of 0.61 K and 0.37 K, and R2 scores of 0.99 and 1 for daytime and nighttime, respectively. The model also shows very low spatial biases: 99.5% (day) and 99.8% (night) of pixels have a mean error between −2 and 2 K (i.e., within the quality control threshold of the Ts data), with errors randomly distributed within cities (Supplementary Fig. S14). Our result underscores the pretrained model’s ability to capture the mapping from land-atmosphere interactions to urban surface temperatures, laying a strong foundation for the subsequent transfer-learning step. Validation of the Ta model To validate the generalizability of U-TL to unseen locations, we randomly divided stations into 10 folds, each containing 5–6 stations. In each iteration, 1 fold was held out for testing, while a subset of 3, 6, or 9 folds from the remaining ones was used for training. This process was repeated 10 times, ensuring that each fold was used for testing once. Note that the entire data from held-out stations, rather than individual samples, were excluded from the training. This design ensures that the testing results better reflect the model’s performance and generalizability to unseen locations. We repeated the whole process 5 times to ensure the robustness of the results. The U-TL framework exhibits high accuracy and strong robustness in predicting urban Ta across scales (Supplementary Figs. S3–5). We validate U-TL’s predicted urban Ta against unseen observations via the leave-station-out 10-fold cross-validation across 18 cities in the CONUS (Supplementary Fig. S3). The results indicate excellent accuracy of our Ta predictions, with out-of-sample mean absolute errors (MAEs) of 0.91 K during daytime and 1.24 K at night (Supplementary Fig. S3). The model accurately captures daily Ta variabilities across space and time, with R2 of 0.98 and 0.97 for daytime and nighttime, respectively. This high predictive accuracy is observed both across and within cities. When evaluated at each individual city, R2 remains high with values ranging from 0.88 to 0.99 at daytime and from 0.86 to 0.98 at nighttime (Supplementary Figs. S15–S16). Notably, this high overall predictive performance is achieved with very limited urban Ta labels, illustrating the power of our transfer-learning approach. Daytime results slightly outperform nighttime ones, likely because the model does not explicitly account for the heat storage term, which plays a more significant role at night. We further demonstrate U-HAT’s high accuracy by comparison against the observationally based PRISM dataset61. Ten metropolitan areas across a range of background climate zones are selected for assessment. We acknowledge that PRISM data is spatially interpolated from standard weather stations and thus, in general, subject to rural biases. However, previous studies have shown that it could partially capture urban signals in certain large metropolitan areas due to a relatively higher density of urban stations within these areas54,90. Our results show that U-HAT accurately reproduces the observed daily distributions of diurnal average, daytime, and nighttime urban Ta within each city and the spatial variability across the ten cities for the whole timespan, as well as summer and winter months from 2013 to 2023 (Supplementary Fig. S4). The higher median of U-HAT’s diurnal average Ta compared to PRISM is likely due to PRISM’s limited incorporation of true urban stations, which leads to its underestimation of urban temperature. Another caveat is the difference in sampling times between these two datasets (PRISM only reports daily maximum and minimum Ta), which might cause differences in their median values across times of day. Nevertheless, these results show strong evidence for U-TL’s ability to reproduce the observed diurnal, daily, and spatial variabilities of urban Ta. To evaluate whether our approach improves upon direct prediction of urban Ta from scratch, we compared U-TL with a model that directly predicts urban Ta using the same predictors without skip connections – hereinafter referred to as the “scratch model”. We divided the selected 52 stations into 10 folds and tested both models’ performance under fewer training samples by progressively reducing the number of training folds from 9 to 6 to 3. The results reveal that U-TL significantly enhances the accuracy, generalizability, and robustness in urban Ta prediction compared to the scratch model, particularly under data-scarce conditions (Supplementary Fig. S5). When relatively abundant training samples are available (i.e., training with 9 folds), U-TL reduces mean absolute error (MAE) by 13.1% and 26.2% for daytime and nighttime, respectively. U-TL also shows narrower error distributions than the scratch model, with standard deviation reductions of 12.6% during the day and 26.0% at night, demonstrating higher robustness of our approach. This advantage becomes more pronounced when data are even scarcer. The performance of U-TL remains remarkably stable as the number of training samples is largely reduced, especially at nighttime, while the scratch model’s errors worsen substantially (Supplementary Fig. S5). From 9 to 3 training folds, U-TL’s MAE increases by only 8.5% during daytime and 7.1% at nighttime, compared to 13.4% and 18.0% for the scratch model. Similarly, the error standard deviation increases by only 8.2% during daytime and 6.9% at nighttime for U-TL, compared to 13.7% and 15.6% for the scratch model. Generalizability of U-TL across climates, surface types, and urban forms To validate U-TL’s generalizability to unseen climates, we conducted a leave-climate-zone-out cross-validation experiment. In addition, we assessed U-TL’s generalizability to unseen climates and urban forms using leave-climate-zone-out and leave-city-out cross-validation, respectively. First, for the leave-climate-zone-out experiment, we grouped all stations by their Köppen–Geiger climate zone and, in each training round, withheld one entire climate-zone group for testing while training on stations from all other zones. Overall, U-TL demonstrates strong spatial generalization ability: across all held-out climate zones, the mean absolute error (MAE) is 1.04 K and 1.41 K, with corresponding R2 values of 0.97 and 0.96 for daytime and nighttime, respectively (Supplementary Fig. S17). Similarly, the leave-city-out cross-validation experiment withholds all stations from each city for testing while training on stations from all other cities in each training round. U-TL remains highly robust under this setting, with MAE of 1.07 K during daytime and 1.32 K during nighttime, and R2 values of 0.97 for both times (Supplementary Figs. S18 and 19). Performance is generally consistent across climate zones and cities, although we do observe slightly reduced skill at coastal sites (e.g., Csa: hot-summer Mediterranean climate; Csb: warm-summer Mediterranean climate; and Am: tropical monsoon climate). This reduction is likely due to local processes, such as sea-breeze circulations, which might only be partially represented in the atmospheric forcing inputs but not fully resolved by our current model. However, when stations with similar coastal characteristics are included in the training set, as in our standard 9-fold cross-validation, the performance improves substantially (Supplementary Figs. S15 and 16). For example, MAE and R2 in Csa (hot-summer Mediterranean climate) zones improve to 1.2 K and 0.92 at daytime and 1.22 K and 0.89 at nighttime, respectively. This suggests that U-TL is able to transfer knowledge from stations with similar local dynamics when such information is represented in the training data, rather than being fundamentally limited in representing these processes. We further assess whether specific surface types or urban geometries challenge the framework by examining the relationship between MAE and indicators of local surface characteristics, for which we use both National Land Cover Dataset (NLCD)91 and Local Climate Zone (LCZ)92 classes to ensure the robustness of the results. For each station, we identified the dominant NLCD or LCZ class within the 1-km buffer and then computed the average MAE from the 9-fold cross-validation experiment across all stations in that class (Supplementary Tables S2 and S3). Results show that MAE varies modestly across classes. Among NLCD classes (Supplementary Table S2), values span 0.84 K (developed medium intensity) to 1.04 K (developed low intensity) at daytime and 1.15 K (developed medium intensity) to 1.56 K (shrub/scrub) at nighttime. Among LCZ classes (Supplementary Table S3), errors range from 0.77 K (sparsely built) to 1.23 K (bare rock or paved) at daytime and 1.05 K (sparsely built) to 1.78 K (bare rock or paved) at nighttime. These results suggest that model performance shows little bias with respect to surface type or urban form. Comparison with existing Ta models and datasets To evaluate U-TL against previous models, we compiled a summary of major gridded air temperature datasets covering the contiguous United States and their reported error metrics (Supplementary Table S4). Note that because most products report overall accuracy based on predominantly rural stations, those statistics are not urban-specific MAE and therefore likely provide a lower bound on their potential errors in urban areas. In addition, the magnitude of MAE for different products can be influenced by their temporal scales; for example, monthly Ta is likely to show lower errors than daily Ta, as evidenced by the comparison between the MAE of daily values and monthly means from U-HAT. Nevertheless, among datasets at similar spatial and temporal scales (i.e., ~1 km and daily scale), U-TL generally achieves lower errors, particularly at daytime. We further computed the urban-specific MAE for each product against the 52 urban stations used in this study over a common period (04-01-2013 to 12-31-2016). Unlike our leave-station-out cross-validation, these evaluations are not necessarily out-of-sample, since some of these datasets may have been trained on these stations. Even so, U-HAT shows lower MAE than most previous datasets, particularly those derived from statistical or traditional machine learning techniques (i.e., Zhang et al.). Some interpolation-based products (e.g., PRISM) exhibit lower errors against selected stations because of a relatively higher density of urban stations incorporated in these areas. However, their accuracy is likely to degrade where stations are sparse, which is the very problem U-TL is intended to address. To provide a more independent validation of model performance, we conducted an additional evaluation using only stations from local observational networks that are not used in the training of other Ta products. When evaluated against this independent subset of local stations, we find an even larger performance advantage of U-HAT relative to existing datasets (Supplementary Table S4). Most models exhibit degraded performance, particularly those relying on spatial interpolation methods (e.g., PRISM, TopoWx, MacDonald et al., and Daymet). In contrast, U-HAT shows better performance, with lower MAE by 25.3% during daytime and 8.9% during nighttime. These comparisons further confirm the high accuracy of U-HAT and its value for urban areas with limited in situ observations. To provide a more robust evaluation, we further added a regression-based baseline model – specifically, a multiple linear regression model using the same predictors as U-TL (hereafter referred to as Baseline-MLR) – and conducted additional analyses, as the majority of the existing traditional methods reported in the literature are regression-based approaches. We then repeated the same 10-fold cross-validation applied to the main model using 3 folds of training data. Our results show that U-TL substantially outperforms this linear regression baseline, where mean absolute error (MAE) is reduced by 11.4 % during daytime and 14.6 % during nighttime (Supplementary Fig. S20). In addition, U-TL yields narrower error distributions, with standard deviation reductions of 8.1 % (day) and 11.6 % (night), indicating improved robustness relative to the linear model. More importantly, we evaluated whether these datasets reproduce observed urban climatology by conducting the same urban heat island (UHI) analysis using the same Simplified Urban Extent (SUE) method as Fig. 2a for consistency. As shown in Supplementary Fig. S21, none of the alternative datasets captures the observed UHI patterns due to inherent limitations in their approaches. For example, the Baseline-MLR model shows cooler urban temperatures than rural areas at daytime, contradicting the widely observed daytime UHI phenomenon. Daymet and PRISM show near-zero UHI across seasons and times of day because their estimates are produced through spatial interpolation from stations that are mostly located in rural or open areas, essentially missing the urban signal. HUMID, while generated with a physics-based model and a bias-corrected approach, produces much stronger daytime than nighttime UHI, which is again inconsistent with widely observed and well-established UHI climatology. This is possibly because HUMID is produced from land-only simulations, therefore missing the two-way land-atmosphere interactions and lateral advection68. These results further suggest that lower errors (e.g., for PRISM) do not equal faithful urban signal representation, particularly where station coverage is sparse. Collectively, these findings underscore the criticality of the TL principle and U-TL’s advantage over traditional methods. These findings highlight U-TL’s strong performance in estimating high-resolution urban Ta with high generalizability to unseen locations and robustness under limited training data – a challenge mirroring the real-world extreme scarcity of true urban observations. This advance largely stems from the transfer-learning approach, in which abundant Ts data from the pretraining step helps the model learn a generalized representation of land-atmosphere interactions, thereby enabling its adaptability across diverse urban environments and under data-scarce conditions. Notably, U-TL achieves this high predictive performance through a unified global model and not a locally refined one. Unlike previous statistical or machine learning methods that are often only locally applicable and may fail in regions with limited training data, U-TL’s reliance on underlying physical principles makes it universally applicable. Therefore, although the number of selected stations is limited because of the extreme scarcity of true urban weather stations, this data limitation is overcome by the transfer-learning approach. That said, U-TL will benefit when more true and representative urban weather stations become available in the future. While Ta variability is primarily influenced by atmospheric forcing fields, surface properties also play a role through land-atmosphere interactions. To assess the significance of surface properties in Ta prediction, we conducted an experiment where training samples were assigned randomly shuffled images and the correct forcings. We compared four scenarios: (1) correct images in both pretraining and fine-tuning steps, (2) correct images in the pretraining step but mismatched images in the fine-tuning step, (3) mismatched images in the pretraining step but correct images in the fine-tuning step, and (4) mismatched images in both pretraining and fine-tuning steps. We followed the same 10-fold cross-validation procedure with 5 repetitions. We found that scenario (1) outperformed all other scenarios in both accuracy and stability, with a reduction of 7.1–14.5% in MAE and 8.4–30.0% in error standard deviation for daytime and a reduction of 6.3–8.7% in MAE and 5.9–8.3% in error standard deviation for nighttime predictions (Supplementary Fig. S22). The result highlights the critical role of surface properties in predicting urban Ta. Urban heat island calculations UHI is defined as the temperature differences between an urban area and its surrounding rural area. Here we compare UHI intensities derived from U-HAT, Yao et al.45, Zhang et al.46, and Ts at daytime and nighttime in summer and winter months of 2013–2020. Both Yao et al.45 and Zhang et al.46 provide seamless 1 km air temperature datasets, with the former offering monthly-averaged values of daily maximum and minimum temperatures, and the latter providing daily values. Since U-HAT provides urban temperature only, we first adopted the simplified urban extent (SUE) method60 to ensure fair comparisons. SUE defines UHI as the temperature difference between urban and non-urban pixels within each urban cluster, and shows UHI characteristics in line with results from other algorithms60. To classify pixels into “urban” and “non-urban”, we used the MODIS Land Cover Type dataset (MCD12Q1), where urban pixels correspond to urban land use, and non-urban pixels include all other non-urban and non-water land uses. To minimize elevation-related effects, we further filtered non-urban pixels to retain only those with an elevation difference of less than 50 m from the median urban elevation of each cluster. Elevation data were obtained from the Global Multi-Resolution Terrain Elevation Data 2010 (GMTED2010)93, and all datasets were reprojected to match their projection. We then computed the mean temperatures over 2013–2020 for the urban and non-urban subsets within each cluster, and their difference yields the UHI intensity. Urban areas lacking available non-urban pixels were excluded, resulting in a final dataset of 335 urban clusters. To ensure the robustness of our findings, we repeated these analyses using an equal-area rural buffer method following a previous study94. We use both PRISM61 and ERA562 as the rural reference temperatures for the canopy UHI (UHIa) calculations. For surface UHI (UHIs), we use the MODIS land surface temperature (LST) data for both urban and rural Ts. The urban extent is defined using the Global Human Settlement Index for 201595. For the rural reference of each urban cluster, we generated an equal-area buffer around the urban region using an iterative 300 m step size. To prevent overlap with nearby urban clusters, pixels classified as urban land use within the rural buffer were excluded. This results in a final dataset of 183 urban clusters. Given the differences between 1:30 AM/PM overpass times and the timing of daily extremes, to test the robustness of the UHI results by U-HAT and ensure comparability with other datasets, we trained U-TL using Tmin and Tmax observations and recomputed the UHI results using the same SUE method. We found close agreement between these two results (Supplementary Fig. S23). This is consistent with previous studies showing similar UHIa intensity between 1:30 AM/PM and daily Ta extremes6 further supporting the validity of our results. We conducted all UHI calculations using the Google Earth Engine platform86. Definitions of all the abbreviations used in this study are listed in Supplementary Table S5. Supplementary information Supplementary Information Transparent Peer Review file Supplementary Information Transparent Peer Review file Source data Source Data 1 Source Data 2 Source Data 3 Source Data 4 Source Data 1 Source Data 2 Source Data 3 Source Data 4