Developing gridded air temperature data over cities using machine learning

Developing gridded air temperature data over cities using machine learning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Developing gridded air temperature data over cities using machine learning Marzie Naserikia, Melissa A. Hart, Elyas Asadi Shamsabadi, Katrin Meissner, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7987756/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Urban heat is a growing concern for public health, energy demand, and urban liveability. High-resolution air temperature (T a ) data is needed for developing effective, localised adaptation strategies, but obtaining such data at the city scale remains a challenge as weather stations are often sparse and unevenly distributed within cities. To address this data gap, we developed a machine learning (ML) framework that creates high-resolution, gridded T a maps using crowdsourced observations and diverse geospatial datasets describing the city. This framework uses satellite-derived Land Surface Temperature (LST), urban form and fabric datasets, and meteorological variables to train a Convolutional Neural Network (CNN) algorithm. The approach was implemented in Sydney, Australia, using multi-day observations from 2019 to 2024, producing 30 m gridded T a estimates with high accuracy (R² = 0.97, RMSE = 0.91°C, surpassing previously reported ML performances). We further assessed the generalisation and spatial transferability of this method, which revealed that the CNN model maintained strong predictive accuracy for unseen locations across the city (R² = 0.91–0.93; RMSE = 1.27–1.44°C). Model performance also remained stable when the number of T a stations was significantly reduced by ~ 80%. Performance slightly declined for unseen days (R² = 0.66–0.93; RMSE = 1.52–2.60°C), suggesting the need for incorporating a broader range of weather conditions. We find that high-resolution city-descriptive datasets are beneficial but not essential, as comparable accuracy was achieved using only globally available predictors. These results indicate that the proposed framework is transferable to other cities, including those in data-sparse regions. The study provides an effective and scalable approach for developing city-scale air temperature maps, which are urgently needed for urban heat assessment, climate adaptation, and public health planning. Air temperature mapping Land surface temperature Crowdsourcing Remote sensing Machine learning Urban heat Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 1. Introduction Urban heat has become a major concern globally, driven by the combined impacts of urbanization and climate change. It presents unevenly distributed impacts across space and socioeconomic groups, affecting public health [ 1 ], energy demand [ 2 ], economic productivity [ 3 ], and environmental quality [ 4 ], thereby constraining efforts toward resilient and climate-adaptive urban development [ 5 ]. Understanding and mitigating these impacts require high-resolution air temperature (T a ) data that captures the spatial and temporal variability of urban heat. The sparse distribution of ground-based sensors, however, limits the ability to represent fine-scale temperature variations across complex urban landscapes [ 6 , 7 ]. To provide spatially continuous estimates of near-surface air temperature, many studies have relied on mesoscale model simulations. Such endeavours typically require intricate physics-based approaches and significant computational resources [ 8 , 9 ]. Alternatively, statistical interpolation techniques based on weather station data have been used [ 10 , 11 ]. However, the limited density and uneven distribution of stations reduce the resolution and accuracy of these methods in capturing urban-scale temperature variability. Satellite-derived Land Surface Temperature (LST) data offers a partial solution by enabling spatially consistent observations of the land surface globally [ 12 , 13 ]. Much of the recent work on mapping urban heat and assessing mitigation strategies has therefore focused on satellite-based LST. However, canopy urban heat, measured by air temperature, is more directly relevant for public health and citizen thermal comfort [ 14 ], and there is no universal, direct, or linear relationship between LST and T a in urban areas [ 15 – 17 ]. Despite this, LST still captures key characteristics of the thermal environment, including the spatial variability of surface heating, and leveraging emerging methods, it can be used to estimate near-surface air temperature. Recent advances in machine learning (ML) have enabled more accurate and scalable approaches to model near-surface air temperature in urban areas. Using these techniques, several studies have integrated satellite-derived LST with ground-based T a measurements to produce gridded temperature estimates across cities [ 18 – 20 ]. Most of these studies rely on Moderate Resolution Imaging Spectroradiometer (MODIS) data to provide temperature estimates at a spatial resolution of approximately 1 km. Relatively few have explored the use of higher-resolution sensors such as Landsat or Sentinel, which offer finer spatial detail more suitable for urban-scale analysis [ 21 ]. Moreover, some studies have focused on mapping monthly average air temperatures rather than capturing finer temporal dynamics. For example, Xu et al. [ 22 ] developed a 1-km resolution dataset of monthly air temperature over the Tibetan Plateau using MODIS and auxiliary data from 2001 to 2015, but did not attempt to resolve daily or sub-daily variability. Conversely, other studies have aimed to map air temperature at higher temporal resolutions but relied on sparse temporal sampling. For instance, Ho et al. [ 23 ] used ML to estimate daily air temperatures over Vancouver, Canada, based on only six individual days scattered across five years, making it difficult to assess the determinants of model performance. Many of these earlier ML applications were constrained by the availability and distribution of ground observations, which were typically drawn from synoptic weather stations. While ML models offer a cost-effective alternative to physics-based simulations, particularly for capturing fine-scale spatial patterns of urban heat, their accuracy and reliability depend heavily on the quality and representativeness of the training data. Air temperature observations are often collected at peri-urban locations such as airports or parks, and therefore lack the spatial resolution needed to capture the heterogeneity of urban microclimates [ 24 , 25 ]. As a result, ML models trained on such data may struggle to accurately represent intra-urban temperature variability in urban areas with complex surface morphology. Moreover, the scalability and generalisability of these models have received limited attention in the previous studies. Addressing these gaps is needed to develop transferable ML frameworks that can reliably estimate air temperature under diverse conditions. The recent increase in crowdsourced weather stations in cities has enabled the analysis of T a distribution within heterogeneous urban environments [ 26 – 29 ]. Located within residential areas, these sensors provide locally relevant data for assessing thermal exposure in diverse neighbourhoods. Integrating these dense, crowd-sourced observations with satellite data holds great potential for improving the spatial mapping of urban air temperatures. Venter et al. [ 21 ] integrated data from over 1,300 crowdsourced weather stations across Oslo in 2018 with satellite observations to model urban air temperatures using a Random Forest approach. While their study is informative and demonstrates the value of leveraging dense crowdsourced measurements alongside satellite data, it is limited to a single year and relies on a conventional, tree-based machine learning model that does not explicitly capture spatial dependencies in the input data. Implementing urban air temperature mapping for real-world applications requires scalable models that not only produce accurate, high-resolution estimates, but also capture spatial relationships and remain robust across multiple years, which has not been the focus of previous efforts. In this study, we integrate crowdsourced T a data with satellite remote sensing imagery to create high-resolution spatially contiguous maps of T a over Sydney, Australia. We developed a machine learning framework using a convolutional neural network (CNN) to infill gaps in the crowdsourced sensor data, using LST derived from Landsat imagery (2019–2024) and urban datasets characterising urban form, fabric, and geography. We aim to assess how T a mapping accuracies are influenced by spatial and temporal variability in the data, as well as the availability of predictor variables and T a observations. This allows us to evaluate the generalisability of our model across different urban areas, time periods, and data constraints. The proposed framework offers a transferable approach for high-resolution urban air temperature mapping, with potential applications in climate adaptation and urban planning. 2. Data and Methods The study followed a three-stage workflow integrating data acquisition, model development, and generalisation assessment. First, satellite, meteorological, and urban form datasets were compiled and pre-processed to generate predictor variables. Second, a CNN model was developed to estimate gridded air temperature using crowdsourced sensor data for training. Finally, model performance, generalisability, and transferability were evaluated through multiple scenarios, including spatial and temporal hold-out tests and reduced-data experiments. Figure 1 summarises the overall workflow. 2.1. Study area The analysis is developed and tested for Sydney, located on the southeast coast of Australia (Fig. 2 ), which has a mean annual temperature of 17.8°C and experiences mild winters and warm summers with 1200 mm of precipitation throughout the year. Sydney houses over 5.5 million residents [ 30 ]. The city’s rapid urban expansion has significantly increased its vulnerability to urban warming effects [ 31 ]. Existing studies in Sydney have often relied on sparse, fixed-point air temperature measurements from meteorological stations and local monitoring networks [ 32 – 34 ]. To effectively characterise urban heat, more detailed spatial temperature data is needed. This approach would be particularly valuable for monitoring canopy urban heat and understanding the key drivers of intra-urban temperature variability across Sydney. 2.2. City-descriptive datasets Previous studies have shown that the physical characteristics in cities, such as building morphology and surface cover, can significantly affect spatial variations in urban temperatures [ 35 – 38 ]. To represent these characteristics, we used the Local Climate Zone (LCZ) data, with a resolution of 100 m, derived from the global dataset developed by Demuzere et al. [ 39 ]. While the LCZ dataset is valuable for representing urban form and surface cover, it is a class-based approach that lacks building height and fine-scale spatial detail. To gain a more detailed representation of the urban landscape, we incorporated a Lidar-based “Geoscape” dataset, which is based on building-resolved, 3D land cover information at 2 m resolution. This dataset contains over twenty variables describing the urban form and fabric, including building height, tree height, and fractional cover of buildings, vegetation, roads, and water, all aggregated to a 300 m resolution [ 40 ]. We also collected distance to the ocean and elevation to account for topographic and coastal influences on temperature patterns. Distance to the ocean was derived from a NASA-provided layer at 0.01° resolution ( https://oceancolor.gsfc.nasa.gov/docs/distfromcoast/ ) , and elevation was obtained from the NASADEM dataset at 30 m spatial resolution. 2.3. Satellite remote sensing data We selected Landsat 8 and 9 scenes covering Sydney for 61 cloud-less days between 2019 and 2024. The imagery corresponds to a mean local overpass time of 10:00 ± 15 min, and LST data were generated at 30 m spatial resolution using the Google Earth Engine (GEE) platform [ 41 ]. LST was retrieved using the Statistical Mono-Window (SMW) algorithm [ 42 ] implemented in GEE. This method estimates LST from top-of-atmosphere brightness temperatures in the thermal infrared band, incorporating surface emissivity from the ASTER Global Emissivity Database and applying a vegetation correction based on NDVI. The LST data derived from this method have been shown to yield satisfactory accuracy, with an RMSE of 1K (for further details, see Ermida et al. [ 43 ]). 2.4. Citizen Weather Station data We used open-access air temperature data from 793 citizen weather stations distributed across Sydney (Fig. 2 b), obtained via the public Netatmo API ( https://dev.netatmo.com/apidocumentation/weather ). The dataset captured measurements for 61 selected cloudless days between 2019 and 2024. Crowdsourced temperature data can be affected by factors such as sensor misplacement, direct solar exposure, or device malfunctions [ 27 , 44 ]. To minimise these issues and ensure data reliability, a quality control procedure was applied using the CrowdQC + package [ 45 ], which identifies and removes statistically implausible observations without requiring reference meteorological data. The first step flagged stations with duplicate geographic coordinates. The second step identified statistical outliers using a z-score approach, comparing each observation against the distribution of all stations for the same time step. The third step removed entire months of data from stations where more than 20% of values were excluded in the previous steps. The fourth step filtered out stations likely located indoors, identified by a weak correlation between their hourly T a measurements and the spatial median of all stations for the same period. The final step excluded unrealistically high values detected as outliers when compared with neighbouring stations. For each station and date, we then extracted the LST value from the pixel containing the station’s location and obtained the T a measurement corresponding to the satellite overpass time (10:00 local time). We also obtained daily minimum and maximum air temperatures and solar radiation from the Australian Bureau of Meteorology’s Observatory Hill station via Climate Data Online ( bom.gov.au ) to capture seasonal variations in the data. 2.5. Air temperature predictions To map gridded T a data, we used a custom Convolutional Neural Network (CNN), which is a deep learning method suitable for capturing spatial patterns and complex, non-linear relationships in spatial data [ 46 ]. Previous studies on gridded T a estimation have often relied on ensemble-based machine learning algorithms such as Random Forest (RF) and Gradient Boosting (GB) [ 47 , 48 ]. While these methods can model complex relationships between predictors and T a , they treat each location independently of their neighbouring areas and therefore might not be able to explicitly capture spatial dependencies and patterns. One possible workaround is to incorporate spatial context by augmenting the predictor set with neighbourhood statistics or by extracting values from surrounding pixels (often referred to as context windows or spatial chips) [ 49 , 50 ]. However, this approach increases data redundancy, may dilute local-scale features, and does not fully exploit the hierarchical feature learning capabilities of deep learning methods such as CNNs [ 51 ]. The CNN was trained to predict T a at grid cells with available observations, using a 7 × 7 neighbourhood of surrounding grid cells, where the observed T a at the centre of each patch served as the target variable and was obtained from crowdsourced measurements. The model used 30 input variables as predictors, including remotely sensed data, urban landscape characteristics, and meteorological variables, as detailed in Fig. 6 . All ancillary data variables were resampled to match the spatial resolution and grid of the LST data prior to model training. The trained model was then used to predict T a for each grid cell across the study area, using the corresponding 7 × 7 patch of input predictors centred on the target grid cell. To determine the optimal patch size, we tested multiple window sizes (3 × 3, 5 × 5, 7 × 7, and larger). Performance improved with increasing patch size, likely due to the additional spatial information captured. However, beyond 7 × 7 no substantial improvement was observed. Therefore, a 7 × 7 size was selected as the best balance between accuracy and computational efficiency. We applied z-score normalization to the continuous input layers using the mean and standard deviation computed from the training set, while excluding the categorical LCZ layer, which was one-hot encoded. Each 7 × 7 × 30 patch was provided as input to the CNN. The first convolutional layer applied a 7 × 7 kernel, reducing the spatial dimensions to 1 × 1 while increasing feature depth. Subsequent 1 × 1 convolutions progressively expanded the feature space to 1 × 1 × 2400. The final layer produced a single scalar output, representing the predicted T a for the grid cell at the sensor location, which was compared against observed values using the loss function. We used a custom loss function that combines Mean Squared Error (MSE) and the coefficient of determination (R²) to balance prediction accuracy and overall fit. The loss is computed as a weighted combination of MSE and R², where a moderate weighting factor (α = 2/3) was adopted to place slightly more emphasis on minimizing prediction errors while still encouraging a strong explained variance between predictions and targets. This balance provided stable convergence and consistent performance across experiments. The overall workflow and CNN architecture used in this study are summarised in Fig. 3 . 2.6. Generalisation and spatial transferability assessment To evaluate model generalisability, we designed two scenarios. Scenario 1 tested spatial generalisation by predicting T a for grid cells or sensors excluded from the training data. Scenario 2 assessed temporal generalisation by predicting T a for days not used during training. For both scenarios, the data were partitioned into six mutually exclusive subsets. In each iteration, one subset (either a group of sensors in scenario 1 or a set of days in scenario 2) was held out exclusively for testing, while the remaining subsets were used for training and validation. This procedure was repeated six times so that all subsets were used once as the test set. To further assess the spatial and temporal transferability of our predictions, we designed two complementary scenarios. The first scenario tested model performance under conditions where high-resolution local datasets are not accessible. In this case, we restricted the predictor set from 30 variables to the seven globally available features to examine the model’s robustness to reduced feature availability. The second scenario focused on the influence of the number and density of T a observations. Here, 15% of the data were reserved for independent testing, while the remaining 85% were used for training and validation (10%). To assess sensitivity to the number of available stations, we progressively increased the proportion of training data from the 85% subset in 2.5% increments (i.e., 2.5%, 5%, 7.5%, and so on) and evaluated performance on the same held-out test set after each step. This approach simulated the gradual expansion of station networks within a city while maintaining a fixed test dataset. The entire process was repeated three times with different random samplings of the training data to ensure the stability of the results. Model accuracy metrics (MAE and R²) were computed for each subset size to evaluate how model performance responds to varying numbers of T a sensors and to assess performance stability under reduced training data. We further examined how the density and spatial distribution of T a sensors influence model performance. Using a leave-one-station-out approach, we first obtained prediction errors for each station. These errors were then analysed against the spatial density of nearby sensors, expressed as the number of stations per km². We further examined the relationship between MAE and the distance to the nearest station to quantify how spatial isolation affects prediction accuracy. 3. Results and Discussion 3.1. Air temperature mapping and model performance Using all input variables and the proposed machine learning framework, we developed high-resolution maps of air temperature across Sydney. Figure 4 presents gridded daytime T a for four representative days. The modelled T a maps reveal spatial gradients across Sydney on both warm and cold days. On the warm days, higher temperatures were mostly concentrated in the western and south-western suburbs, likely due to dense built-up areas, limited vegetation cover, and distance to the ocean. On the cold days, coastal moderation was also evident, with relatively higher T a near the shoreline. Cooler areas were observed over vegetated and elevated regions, particularly in the northern parts of the city. These patterns demonstrate that Sydney’s air temperature variability is greatly influenced by coastal proximity, urban density, and land-surface characteristics, which the model effectively captured across warm and cold conditions. Similar spatial contrasts have been reported by Potgieter et al. [ 52 ], who found that distance from the ocean strongly influences air temperature variations across Sydney. The high-resolution T a maps developed in this study enable detailed identification of hot and cool spots across the city. These data can guide targeted mitigation strategies, such as vegetation planting or increasing surface albedo and allow their effectiveness to be monitored over time. They are also valuable for validating urban climate models that incorporate urban canopy schemes [ 53 , 54 ]. More broadly, fine-scale T a mapping enhances understanding of the factors driving urban cooling and supports improved heat-exposure and risk assessments. Figure 5 shows the overall performance of the CNN model. The dataset was randomly divided into training (75%), validation (10%), and test (15%) subsets. The model performed well, achieving an R² of 0.97 and an RMSE of 0.91°C on both validation and test sets. This accuracy compares favourably with other recent studies applying machine learning to air temperature estimation, Reported RMSE values in previous works typically range between 1.0 and 3.0°C and R² values between 0.3 and 0.94, depending on spatial resolution, data sources, model architecture, and urban complexity [ 8 , 48 , 55 , 56 ]. For instance, a recent study in Morocco, Africa, which used satellite-derived predictors and tested several algorithms, achieved its best performance with a Multilayer Perceptron (MLP), obtaining an RMSE below 1.6°C at 1 km resolution [ 57 ]. At a finer spatial resolution, Venter et al. [ 21 ] applied a Random Forest model in Oslo, Norway, using remote sensing and crowdsourced data, reporting an RMSE of 1.85°C and an R² of 0.05 for daily maximum air temperature. Compared with these approaches, the present CNN framework achieved both finer spatial resolution (30 m) and higher predictive accuracy, highlighting its ability to capture spatial patterns in urban air temperature while maintaining strong predictive performance. We also assessed how model accuracy varied within the city using the LCZ classification (not shown). The results indicated lower mapping errors in compact and open low-rise areas, consistent with the findings of Venter et al. [ 21 ] for Oslo. However, because the number of stations differed substantially among LCZ classes, the comparison was inherently biased. 3.2. Predictor variable importance We evaluated the contribution of each predictor to model performance using a perturbation approach. For each variable, its values in the test dataset were replaced with a constant outside the range of data to remove variability. The resulting decrease in predictive accuracy (RMSE) was compared across all predictors, and predictor variables causing the largest performance drops were ranked as most important. In our model, the most important predictors were satellite-derived LST, solar radiation, and daily maximum and minimum air temperature (Tmax and Tmin), followed by LCZ (Fig. 6 ). The high importance of LST and solar radiation aligns with previous findings highlighting their strong influence on maximum air temperature, particularly during hot summer days [ 58 ]. The model appears to rely more strongly on variables that exhibit temporal variability than on static predictors, which aligns with observations reported by Hobeichi et al. [ 49 ]. The strong contribution of solar radiation, Tmax, and Tmin in our results reflects their role in providing the model with information on day-to-day and seasonal variations in air temperature. Without such temporal indicators, the model would primarily rely on spatial features, which are all static and would therefore struggle to capture time-related variability in T a . After LCZ, tree height and water fraction were the most influential predictors. Tree height likely reflects the cooling effect of vegetation through shading and evapotranspiration, while water fraction captures the moderating influence of coastal and inland water bodies on near-surface air temperature. 3.3. Generalisation and spatial transferability assessment We evaluated model generalisability under two scenarios. Scenario 1 addressed spatial transferability, in which the model was tested on T a sensors not included in training. In this setup, ~ 660 sensors were used for training and validation, while the remaining ~ 130 were held out exclusively for testing. Six different train–test combinations were created based on the total number of unique sensors (Fig. 7 ). Model performance on the validation set remained consistent with the all-data baseline (R 2 = 0.97, RMSE = 0.86–0.94°C). For the test sets, R 2 decreased slightly to 0.91–0.93, with RMSE ranging from 1.27 to 1.44°C in all splits, which still shows robust performance. These results indicate that the model can effectively predict air temperature in unseen locations based on spatial patterns learned from the training data. Similar spatial hold-out testing has been applied in a previous study in Guangzhou, China [ 61 ], based on observations recorded by around 300 stations from March 2020 to March 2021. They found a marginal increase in RMSE compared to the full-data baseline, indicating that the model effectively captured spatial patterns and could predict air temperature at unobserved locations. Scenario 2 examined temporal transferability by testing the model on days not included in training. The dataset comprised 61 cloudless days across multiple seasons from 2019 to 2024. In each experiment, data from 50 days were used for training and validation, and the remaining 10–11 days for testing, producing six different train–test combinations (Fig. 8 ). While the all-data baseline yielded R² around 0.97, for the test set, Scenario 2 results ranged from R² = 0.66 to 0.93, with RMSE values between 1.52 to 2.6°C. These values still remain within the performance ranges reported in previous ML-based air-temperature mapping studies discussed earlier. The reduced accuracy suggests that the model has greater difficulty capturing atmospheric or weather conditions not represented in the training set, reflecting the inherently more complex and less predictable nature of temporal patterns compared to spatial patterns. Unlike an earlier MODIS-based study in Los Angeles and Seoul [ 55 ], which reported that incorporating LST observations from additional days did not necessarily improve T a estimation, our findings suggest that expanding the temporal diversity of the training data is an important factor in enhancing model robustness, particularly when it comes to creating maps for future dates. Incorporating a broader range of days, including extreme or atypical conditions, can further strengthen temporal generalisation. We assessed the model’s spatial transferability under two additional scenarios. The first scenario examined model performance under limited input feature availability, representing cases where high-resolution local datasets are not accessible. The full model was trained using 30 predictor variables, comprising both globally available and locally specific features. However, only seven of these variables are accessible globally at comparable resolution: LST, LCZ, solar radiation, Tmax, Tmin, elevation, and distance to coast, where five of these predictors were identified among the most important contributing features (shown in Fig. 6 ). To assess transferability, we trained an alternative model using only these seven global predictors (Fig. 8 ). Compared to the full-input model (R² = 0.97, RMSE = 0.91°C), the reduced-input version showed only a marginal decline in performance (R² = 0.96, RMSE = 0.98°C). Thus, the model still maintained strong predictive capacity, indicating robustness to reduced feature availability. We further repeated the spatial and temporal generalisation tests using only the seven global predictors to evaluate how well the reduced-input model performs under unseen conditions. The results remained close to those of the full-input model, demonstrating that the proposed framework can generalise well across space and time, even with a limited set of globally available variables. This suggests strong potential for transferring the approach to other cities worldwide where local high-resolution datasets are not accessible. The availability of sufficient ground-truth weather station data in cities for training and validation is another important factor influencing model performance. Therefore, the second scenario assessed the model’s sensitivity to the density of available T a observations. Because the spatial coverage of station data varies greatly among cities, we tested how progressively increasing the number of T a sensors affected model accuracy. The results showed that performance stabilised when approximately 20% of the stations (~ 160) were used for training, with only marginal improvements from adding additional sensors (Fig. 9 ). This implies that the proposed approach can be effectively extended to cities with limited observational networks. Lastly, we evaluated the influence of station density, rather than just the total number of sensors, since cities differ in size and spatial coverage. We observed that station density had a limited influence on model accuracy (Fig. 10 ). Increasing station density from one to six per 1 km grid cell did not substantially reduce the mean absolute error (MAE) (Fig. 10 b). Model performance remained largely stable across density levels, except for very isolated stations, where errors tended to be higher (Fig. 10 a). These findings suggest that the approach is technically transferable: the model structure and methodology can be applied to cities worldwide, even where high-resolution local datasets or dense ground-truth T a observations are lacking. Using Random Forest regression modelling, Venter et al. [ 21 ] estimated that around 250 stations are required to train and validate a model effectively and noted that most European cities meet this threshold. However, such station densities are uncommon in many cities, particularly in the Global South. In cases where crowdsourced networks are sparse, official weather stations can provide a viable alternative source of training data. Incorporating these stations could therefore extend the applicability of the proposed ML framework to data-limited contexts, although local validation would remain essential to ensure reliable performance. 3.4. Limitations and future work Despite the benefits of this developed dataset and modelling framework, some limitations need to be acknowledged. As satellite LST cannot be collected on cloudy days, the resulting air temperature maps are biased toward clear-sky conditions. Moreover, this study was restricted to daytime conditions, since the Landsat satellites only provide surface temperature observations during the day. The fixed overpass time of ~ 10:00 a.m. limits the model to represent morning conditions, which may differ from temperature patterns observed later in the day. Thus, further investigation is needed to incorporate multi-temporal or higher-frequency satellite observations to better capture diurnal variations in air temperature and improve the robustness of gridded estimates. In addition, expanding the set of predictor variables to include key meteorological factors such as wind speed and humidity could further enhance the accuracy of air temperature estimation. 4. Conclusion High-resolution spatial air temperature data is critical for understanding urban heat, yet it is often difficult to obtain using sparse conventional weather station networks. While satellite-derived LST provides useful spatial coverage, it is not directly relevant to human thermal exposure and health in the same way that T a is. Crowdsourced T a data offer a valuable complement, as these stations are densely located within residential areas where people live and where heat anomalies most strongly affect population health. Their spatial coverage, however, is limited. Our study demonstrated the benefits of integrating satellite observations with crowdsourced T a data. By combining high-resolution satellite imagery, urban landscape metrics, and terrain information within a CNN framework, we produced 30 m resolution maps of urban T a over Sydney, Australia. Compared with conventional machine learning approaches, CNNs provide clear advantages by capturing spatial patterns and contextual information more effectively. The model achieved strong performance, with the most influential predictors being LST, solar radiation, daily maximum and minimum T a , and LCZ. Vegetation and water-related variables also played an important role in modulating local microclimates. In terms of generalisability, the framework performed consistently well across space, while temporal transfer proved more challenging, highlighting the importance of incorporating a wider range of days to improve robustness. Our transferability analysis showed that high-resolution urban datasets, while useful, are not essential. Models trained using only globally available predictors performed nearly as well as those using detailed local datasets, supporting the broader applicability of this approach in data-limited contexts. Moreover, the model maintained stable performance even when the number and density of training stations were substantially reduced, demonstrating robustness to sparse observational networks. These findings suggest that the proposed framework can be effectively applied to other cities worldwide, including those with limited data availability. Overall, this study demonstrates that integrating crowdsourced air temperature measurements with satellite and urban data through deep learning provides an effective, scalable, and transferable method for mapping fine-scale urban air temperatures. Such maps are directly relevant for stakeholders in urban planning and public health, offering critical insights for climate change adaptation and the design of heat-resilient cities in a rapidly urbanising world. Declarations Declaration of competing interest The authors declare no competing interests. Acknowledgements This work was supported by the Australian Research Council as part of the Centre of Excellence for Climate Extremes (CE170100023). We also acknowledge support from the Minderoo Foundation (PS74451). References Tan J et al (2010) The urban heat island and its impact on heat waves and human health in Shanghai. Int J Biometeorol 54:75–84. https://doi.org/10.1007/s00484-009-0256-x Santamouris M, Cartalis C, Synnefa A, Kolokotsa D (2015) On the impact of urban heat island and global warming on the power demand and electricity consumption of buildings—A review. Energy Build 98:119–124. https://doi.org/10.1016/j.enbuild.2014.09.052 Herbel I, Croitoru A-E, Rus AV, Roşca CF, Harpa GV, Ciupertea A-F, Rus I (2018) The impact of heat waves on surface urban heat island and local economy in Cluj-Napoca city, Romania. Theoret Appl Climatol 133(3):681–695. https://doi.org/10.1007/s00704-017-2196-4 Aw J, Kleeman MJ (2003) Evaluating the first-order effect of intraannual temperature variability on urban air pollution. J Geophys Research: Atmos 108(D12). p. https://doi.org/10.1029/2002JD002688 Nazarian N et al (2022) Integrated assessment of urban overheating impacts on human life. Earths Future 10(8). https://doi.org/10.1029/2022EF002682 . p. e2022EF002682 Baranka G, Bozó L, Ciglič R, Komac B (2016) Urban Heat Island Gold Standard and Urban Heat Island Atlas: Gold Standard for UHI Measurements and Introduction of The Central-European Urban Heat Island Atlas. Counteracting Urban Heat Island Effects in a Global Climate Change Scenario. Springer, pp 41–70. https://doi.org/10.1007/978-3-319-10425-6_2 Wang K, Jiang S, Wang J, Zhou C, Wang X, Lee X (2017) Comparing the diurnal and seasonal variabilities of atmospheric and surface urban heat islands based on the Beijing urban meteorological network. J Geophys Research: Atmos 122(4):2131–2154. https://doi.org/10.1002/2016JD025304 Hassani A, Santos GS, Schneider P, Castell N (2024) Interpolation, satellite-based machine learning, or meteorological simulation? A comparison analysis for spatio-temporal mapping of mesoscale urban air temperature. Environ Model Assess 29(2):291–306. https://doi.org/10.1007/s10666-023-09943-9 Martilli A et al (2024) WRF-Comfort: simulating microscale variability in outdoor heat stress at the city scale with a mesoscale model. Geosci Model Dev 17(12):5023–5039. https://doi.org/10.5194/gmd-17-5023-2024 Wang M, He G, Zhang Z, Wang G, Zhang Z, Cao X, Wu Z, Liu X (2017) Comparison of spatial interpolation and regression analysis models for an estimation of monthly near surface air temperature in China. Remote Sens 9(12):1278. https://doi.org/10.3390/rs9121278 Florio E, Lele S, Chi Chang Y, Sterner R, Glass G (2004) Integrating AVHRR satellite data and NOAA ground observations to predict surface air temperature: a statistical approach. Int J Remote Sens 25(15):2979–2994. https://doi.org/10.1080/01431160310001624593 Zhou D et al (2018) Satellite remote sensing of surface urban heat islands: Progress, challenges, and perspectives. Remote Sens 11(1):48. https://doi.org/10.3390/rs11010048 Naserikia M, Hart MA, Nazarian N, Bechtel B (2022) Background climate modulates the impact of land cover on urban surface temperature. Sci Rep 12(1):15433. https://doi.org/10.1038/s41598-022-19431-x Martilli A, Krayenhoff ES, Nazarian N (2020) Is the Urban Heat Island intensity relevant for heat mitigation studies? Urban Clim 31:100541. https://doi.org/10.1016/j.uclim.2019.100541 Naserikia M, Hart MA, Nazarian N, Bechtel B, Lipson M, Nice KA (2023) Land surface and air temperature dynamics: The role of urban form and seasonality. Sci Total Environ 905:167306. https://doi.org/10.1016/j.scitotenv.2023.167306 Zhan W et al (2025) Satellite-derived Land Surface Temperatures Strongly Mischaracterise Urban Heat Hazard. arXiv preprint arXiv:2509.16568 , p. https://doi.org/10.48550/arXiv.2509.16568 Venter ZS, Chakraborty T, Lee X (2021) Crowdsourced air temperatures contrast satellite measures of the urban heat island and its mechanisms. Sci Adv 7(22):eabb9569. https://doi.org/10.1126/sciadv.abb9569 Chen Y, Sun H, Li J (2016) Estimating daily maximum air temperature with MODIS data and a daytime temperature variation model in Beijing urban area. Remote Sens Lett 7(9):865–874. https://doi.org/10.1080/2150704X.2016.1193792 Noi PT, Kappas M, Degener J (2016) Estimating daily maximum and minimum land air surface temperature using MODIS land surface temperature data and ground truth data in Northern Vietnam. Remote Sens 8(12):1002. https://doi.org/10.3390/rs8121002 Shen H, Jiang Y, Li T, Cheng Q, Zeng C, Zhang L (2020) Deep learning-based air temperature mapping by fusing remote sensing, station, simulation and socioeconomic data. Remote Sens Environ 240:111692. https://doi.org/10.1016/j.rse.2020.111692 Venter ZS, Brousse O, Esau I, Meier F (2020) Hyperlocal mapping of urban air temperature using remote sensing and crowdsourced weather data. Remote Sens Environ 242:111791. https://doi.org/10.1016/j.rse.2020.111791 Xu Y, Knudby A, Shen Y, Liu Y (2018) Mapping monthly air temperature in the Tibetan Plateau from MODIS data based on machine learning methods. IEEE J Sel Top Appl earth observations remote Sens 11(2):345–354. https://doi.org/10.1109/JSTARS.2017.2787191 Ho HC, Knudby A, Sirovyak P, Xu Y, Hodul M, Henderson SB (2014) Mapping maximum urban air temperature on hot summer days. Remote Sens Environ 154:38–45. https://doi.org/10.1016/j.rse.2014.08.012 Schlünzen KH, Grimmond S, Baklanov A (2023) Guidance on measuring, modelling and monitoring the canopy layer urban heat island (CL-UHI). World Meteorological Origanization World Meteorological Organization (WMO) (2021) Guide to Instruments and Methods of Observation Vol. V—Quality Assurance and Management of Observing Systems . WMO Geneva, Switzerland Fenner D, Meier F, Bechtel B, Otto M, Scherer D (2017) Intra and inter local climate zone variability of air temperature as observed by crowdsourced citizen weather stations in Berlin, Germany. Meteorol Z 26:525–547. https://doi.org/10.1127/metz/2017/0861 Meier F, Fenner D, Grassmann T, Otto M, Scherer D (2017) Crowdsourcing air temperature from citizen weather stations for urban climate research. Urban Clim 19:170–191. https://doi.org/10.1016/j.uclim.2017.01.006 Muller C, Chapman L, Johnston S, Kidd C, Illingworth S, Foody G, Overeem A, Leigh R (2015) Crowdsourcing for climate and atmospheric sciences: current status and future potential. Int J Climatol 35(11):3185–3203. https://doi.org/10.1002/joc.4210 Chapman L, Bell C, Bell S (2016) Can the crowdsourcing data paradigm take atmospheric science to a new level? A case study of the Urban Heat Island of London quantified using Netatmo weather stations. Int J Climatol. https://doi.org/10.1002/joc.4940 (ABS) (2024) A.B.o.S., Australian Statistical Geography Standard (ASGS) Argüeso D, Evans JP, Pitman AJ, Di A, Luca (2015) Effects of city expansion on heat stress under climate change conditions. PLoS ONE 10(2):e0117066. https://doi.org/10.1371/journal.pone.0117066 Pfautsch S, Wujeska-Klause A, Walters JR (2025) Spatiotemporal variation of intra-urban heat and heatwaves across Greater Sydney, Australia. Weather Clim Extremes 47:100741. https://doi.org/10.1016/j.wace.2024.100741 Sidiqui P, Huete A, Devadas R (2016) Spatio-temporal mapping and monitoring of Urban Heat Island patterns over Sydney, Australia using MODIS and Landsat-8. in 4th International Workshop on Earth observation and Remote Sensing applications (EORSA) . IEEE Santamouris M, Haddad S, Fiorito F, Osmond P, Ding L, Prasad D, Zhai X, Wang R (2017) Urban heat island and overheating characteristics in Sydney, Australia. An analysis of multiyear measurements. Sustainability 9(5):712. https://doi.org/10.3390/su9050712 Nice KA et al (2022) Isolating the impacts of urban form and fabric from geography on urban heat and human thermal comfort. Build Environ 224:109502. https://doi.org/10.1016/j.buildenv.2022.109502 Konarska J, Holmer B, Lindberg F, Thorsson S (2016) Influence of vegetation and building geometry on the spatial variations of air temperature and cooling rates in a high-latitude city. Int J Climatol 36(5). p. https://doi.org/10.1002/joc.4502 Yin C, Yuan M, Lu Y, Huang Y, Liu Y (2018) Effects of urban form on the urban heat island effect based on spatial regression model. Sci Total Environ 634:696–704. https://doi.org/10.1016/j.scitotenv.2018.03.350 Naserikia M, Asadi Shamsabadi E, Rafieian M, Leal Filho W (2019) The urban heat island in an urban context: A case study of Mashhad, Iran. Int J Environ Res Public Health 16(3):313. https://doi.org/10.3390/ijerph16030313 Demuzere M, Kittner J, Martilli A, Mills G, Moede C, Stewart ID, Van Vliet J, Bechtel B (2022) A global map of Local Climate Zones to support earth system modelling and urban scale environmental science. Earth System Science Data Discussions , 2022: pp. 1–57. https://doi.org/10.5194/essd-14-3835-2022 Lipson MJ, Nazarian N, Hart MA, Nice KA, Conroy B (2022) A transformation in city-descriptive input data for urban climate models. Front Environ Sci 10:866398. https://doi.org/10.3389/fenvs.2022.866398 Gorelick N, Hancher M, Dixon M, Ilyushchenko S, Thau D, Moore R (2017) Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sens Environ 202:18–27. https://doi.org/10.1016/j.rse.2017.06.031 Duguay-Tetzlaff A et al (2015) Meteosat land surface temperature climate data record: Achievable accuracy and potential uncertainties. Remote Sens 7(10):13139–13156. https://doi.org/10.3390/rs71013139 Ermida SL, Soares P, Mantas V, Göttsche F-M, Trigo IF (2020) Google earth engine open-source code for land surface temperature estimation from the landsat series. Remote Sens 12(9):1471. https://doi.org/10.3390/rs12091471 Varentsov M, Konstantinov P, Shartova N, Samsonov T, Kargashin P, Varentsov A, Fenner D, Meier F (2020) Urban heat island of the Moscow megacity: the long-term trends and new approaches for monitoring and research based on crowdsourcing data. in IOP Conference Series: Earth and Environmental Science . IOP Publishing. http://doi.org/10.1088/1755-1315/606/1/012063 Fenner D, Bechtel B, Demuzere M, Kittner J, Meier F (2021) CrowdQC+—a quality-control for crowdsourced air-temperature observations enabling world-wide urban climate applications. Front Environ Sci 9:720747. https://doi.org/10.3389/fenvs.2021.720747 Zhao X, Wang L, Zhang Y, Han X, Deveci M, Parmar M (2024) A review of convolutional neural networks in computer vision. Artif Intell Rev 57(4):99. https://doi.org/10.1007/s10462-024-10721-6 Chen G, Shi Y, Wang R, Ren C, Ng E, Fang X, Ren Z (2022) Integrating weather observations and local-climate-zone-based landscape patterns for regional hourly air temperature mapping using machine learning. Sci Total Environ 841:156737. https://doi.org/10.1016/j.scitotenv.2022.156737 dos Santos RS (2020) Estimating spatio-temporal air temperature in London (UK) using machine learning and earth observation satellite data. Int J Appl Earth Obs Geoinf 88:102066. https://doi.org/10.1016/j.jag.2020.102066 Hobeichi S, Nishant N, Shao Y, Abramowitz G, Pitman A, Sherwood S, Bishop C, Green S (2023) Using machine learning to cut the cost of dynamical downscaling. Earths Future 11(3). https://doi.org/10.1029/2022EF003291 . p. e2022EF003291 Nishant N, Hobeichi S, Sherwood S, Abramowitz G, Shao Y, Bishop C, Pitman A (2023) Comparison of a novel machine learning approach with dynamical downscaling for Australian precipitation. Environ Res Lett 18(9):094006. https://doi.org/10.1088/1748-9326/ace463 Rampal N et al (2024) Enhancing regional climate downscaling through advances in machine learning. Artif Intell Earth Syst 3(2):230066. https://doi.org/10.1175/AIES-D-23-0066.1 Potgieter J, Nazarian N, Lipson MJ, Hart MA, Ulpiani G, Morrison W, Benjamin K (2021) Combining high-resolution land use data with crowdsourced air temperature to investigate intra-urban microclimate. Front Environ Sci 9:720323. https://doi.org/10.3389/fenvs.2021.720323 Chen F et al (2011) The integrated WRF/urban modelling system: development, evaluation, and applications to urban environmental problems. Int J Climatol 31(2):273–288. https://doi.org/10.1002/joc.2158 Salamanca F, Krpo A, Martilli A, Clappier A (2010) A new building energy model coupled with an urban canopy parameterization for urban climate simulations—part I. formulation, verification, and sensitivity analysis of the model. Theoret Appl Climatol 99(3):331–344. https://doi.org/10.1007/s00704-009-0142-9 Yoo C, Im J, Park S, Quackenbush LJ (2018) Estimation of daily maximum and minimum air temperatures in urban landscapes using MODIS time series satellite data. ISPRS J photogrammetry remote Sens 137:149–162. https://doi.org/10.1016/j.isprsjprs.2018.01.018 Wang W, Brönnimann S, Zhou J, Li S, Wang Z (2025) Near-surface air temperature estimation for areas with sparse observations based on transfer learning. ISPRS J photogrammetry remote Sens 220:712–727. https://doi.org/10.1016/j.isprsjprs.2025.01.021 Salih W et al (2025) A High-Resolution Air Temperature Estimation Using Machine Learning and Land Surface Data Across Morocco. Earth Syst Environ 9(2):1311–1331. https://doi.org/10.1007/s41748-025-00629-8 Vancutsem C, Ceccato P, Dinku T, Connor SJ (2010) Evaluation of MODIS land surface temperature data to estimate air temperature in different ecosystems over Africa. Remote Sens Environ 114(2):449–465. https://doi.org/10.1016/j.rse.2009.10.002 Macdonald R, Griffiths R, Hall D (1998) An improved method for the estimation of surface roughness of obstacle arrays. Atmos Environ 32(11):1857–1864. https://doi.org/10.1016/S1352-2310(97)00403-2 Kanda M, Inagaki A, Miyamoto T, Gryschka M, Raasch S (2013) A new aerodynamic parametrization for real urban surfaces. Boundary Layer Meteorol 148(2):357–377. https://doi.org/10.1007/s10546-013-9818-x Ding X, Zhao Y, Fan Y, Li Y, Ge J (2023) Machine learning-assisted mapping of city-scale air temperature: Using sparse meteorological data for urban climate modeling and adaptation. Build Environ 234:110211. https://doi.org/10.1016/j.buildenv.2023.110211 Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7987756","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":537392194,"identity":"5294bbc9-8d06-4774-b5d5-0aaf24bab7d2","order_by":0,"name":"Marzie Naserikia","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA8klEQVRIiWNgGAWjYBADHglmBsYHUI4B0VqYYUqJ08IgwcDAJkGUFt0G3oMff1Tck5FsZ35WzbujLrGBvXmbBEPNYZxazA7wJUvznCnmkWZmM7vNe+ZwYgPPsTIJhmP4tPAYSDO2JfDIMTMAtbQdyG2QyDGTYGDDq8X4589/IC3s34p52+pyG+TfALX8w6vFTIK3IQHoMB4zZt42ZqAtQBHGNjxaDvOYWfMcS+CRbOYplpzbdri+jSet2CKxLx23luM9xjd/1CTYS5w/vvHD27Y6Y372wxtvfPhmjVMLAzO6ABuISMCtYRSMglEwCkYBEQAAs1ZH8XmUYo4AAAAASUVORK5CYII=","orcid":"https://orcid.org/0009-0009-8358-3603","institution":"University of New South Wales","correspondingAuthor":true,"prefix":"","firstName":"Marzie","middleName":"","lastName":"Naserikia","suffix":""},{"id":537394263,"identity":"1973ee57-6af4-4959-b3e2-611454b879c2","order_by":1,"name":"Melissa A. Hart","email":"","orcid":"https://orcid.org/0000-0002-3512-8843","institution":"University of Tasmania","correspondingAuthor":false,"prefix":"","firstName":"Melissa","middleName":"A.","lastName":"Hart","suffix":""},{"id":537394264,"identity":"1afceee4-bb53-4651-9e04-ad0ab0ea0799","order_by":2,"name":"Elyas Asadi Shamsabadi","email":"","orcid":"https://orcid.org/0000-0001-7314-5217","institution":"The University of Sydney","correspondingAuthor":false,"prefix":"","firstName":"Elyas","middleName":"Asadi","lastName":"Shamsabadi","suffix":""},{"id":537394265,"identity":"8b89b2a3-df22-4086-9ced-5ad9fabd572f","order_by":3,"name":"Katrin Meissner","email":"","orcid":"https://orcid.org/0000-0002-0716-7415","institution":"University of New South Wales","correspondingAuthor":false,"prefix":"","firstName":"Katrin","middleName":"","lastName":"Meissner","suffix":""},{"id":537394266,"identity":"cd0b1af7-e747-4c58-ae53-2994a59ee74d","order_by":4,"name":"Sanaa Hobeichi","email":"","orcid":"https://orcid.org/0000-0001-6825-3854","institution":"University of New South Wales","correspondingAuthor":false,"prefix":"","firstName":"Sanaa","middleName":"","lastName":"Hobeichi","suffix":""},{"id":537394267,"identity":"d677f1f5-7dc6-43d5-b9b2-ebc49a4e88e3","order_by":5,"name":"Benjamin Bechtel","email":"","orcid":"https://orcid.org/0000-0001-8802-7934","institution":"Ruhr-University Bochum","correspondingAuthor":false,"prefix":"","firstName":"Benjamin","middleName":"","lastName":"Bechtel","suffix":""},{"id":537394268,"identity":"ecf64b25-94d9-4ee3-838f-0cd31ea9296e","order_by":6,"name":"Negin Nazarian","email":"","orcid":"https://orcid.org/0000-0001-9839-7234","institution":"University of New South Wales","correspondingAuthor":false,"prefix":"","firstName":"Negin","middleName":"","lastName":"Nazarian","suffix":""}],"badges":[],"createdAt":"2025-10-30 09:44:02","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7987756/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7987756/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":94814948,"identity":"15d1afe6-54c4-4ce1-b3bb-4089f3b630b1","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":22098224,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/a4fdf3ae6782383ce1e7ae31.docx"},{"id":94814947,"identity":"d3b92d5a-c16c-4488-b232-cf04f882dda6","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs7987756.json","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/808f204897e83d74dc2f8c64.json"},{"id":94814943,"identity":"877c1d5e-be35-4328-9fc4-c2bf25bd7921","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":150209,"visible":true,"origin":"","legend":"","description":"","filename":"rs79877560enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/2634a15f1485f7208bafa00b.xml"},{"id":94814951,"identity":"f9bb0145-6006-49d1-81da-2d7238a5c821","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":130747,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/93bc427b5a8782a49326d137.png"},{"id":94814944,"identity":"0634c6e5-4571-444f-ac0b-9b4057635f45","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":615394,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/3df016ca63456ed6d4d71a5f.png"},{"id":94814952,"identity":"c2722ce2-57dc-4126-89ef-be876d7b24d1","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4924003,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/9856e1c8c9051a2afc3638a3.png"},{"id":94814961,"identity":"7ee07236-ecfd-43a5-bbb4-54791972d87f","added_by":"auto","created_at":"2025-10-31 04:19:53","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1060162,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/da6ec73d8ac9a4b8e83cd86c.png"},{"id":94814954,"identity":"93d06133-a96a-45d3-ba98-8492fdc9d818","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4548999,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/45e4d83ba8d0c4bb2bf7ea2d.png"},{"id":94814968,"identity":"45d41c8b-931b-4225-b0a8-2d26724cd6da","added_by":"auto","created_at":"2025-10-31 04:19:53","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":225042,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/a379a0e58a11fc913303e4a6.png"},{"id":94825511,"identity":"848103ba-1494-4003-a88a-cb7b8621a0b1","added_by":"auto","created_at":"2025-10-31 06:50:22","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2331561,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/517c71961f3424b7434c8aab.png"},{"id":94814950,"identity":"7ba35792-4544-4698-94cc-58ad27038028","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1543382,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/a0caec0bacf44f7fce679039.png"},{"id":94814949,"identity":"7e97a28e-c38f-40f2-bb08-1f2987a0b485","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1392746,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/d9f5083b86836038b942b24f.png"},{"id":94814956,"identity":"44278041-6710-4e50-8ade-c80d72b6c67e","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2932055,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/a576ddf14e147f2462e14d6d.png"},{"id":94814958,"identity":"6d066212-adf5-4592-b465-fdeffe0d7465","added_by":"auto","created_at":"2025-10-31 04:19:53","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":76939,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/9ae60062a9dac63280969655.png"},{"id":94814957,"identity":"42036860-8a67-4b40-941a-581c63f2f02f","added_by":"auto","created_at":"2025-10-31 04:19:53","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":122365,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/f38531b2162f4572abd81ceb.png"},{"id":94814962,"identity":"7c8adc38-c63d-491a-a00e-49336a677542","added_by":"auto","created_at":"2025-10-31 04:19:53","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":752630,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/64d1225d1188a23ccf637b22.png"},{"id":94814965,"identity":"d93681e8-a655-4940-8f20-b54ffaa98119","added_by":"auto","created_at":"2025-10-31 04:19:53","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":199929,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/2b71d970a71cb056a96ac133.png"},{"id":94826318,"identity":"bf07cd58-4e36-4352-8b25-820084a5c6ea","added_by":"auto","created_at":"2025-10-31 06:51:24","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":780251,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/7c2d9678deb55a179f83d302.png"},{"id":94814963,"identity":"987f9ce6-90e4-4b45-ac19-0c7ae92fc9bd","added_by":"auto","created_at":"2025-10-31 04:19:53","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":55689,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/c530b5be33b96b6cc16bc983.png"},{"id":94814966,"identity":"b1483f3c-2b5e-4b00-9ec1-824d269e70d3","added_by":"auto","created_at":"2025-10-31 04:19:53","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":429154,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/11109b21205a83c260dce4a4.png"},{"id":94826273,"identity":"e3123e2e-b96d-4399-a5cb-1d54d0f059e1","added_by":"auto","created_at":"2025-10-31 06:51:20","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":290957,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/6f2c215ca4b910f7c6f1d0b6.png"},{"id":94825599,"identity":"f695388a-b95e-48a0-99cd-06f780d29ba7","added_by":"auto","created_at":"2025-10-31 06:50:28","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":282124,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/b27d75a4f0e4c52f89e97494.png"},{"id":94814972,"identity":"a97d649d-072f-49f7-b736-c9158014b0f8","added_by":"auto","created_at":"2025-10-31 04:19:53","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":884710,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/ef5364ad3cf2e86c83103de1.png"},{"id":94814964,"identity":"9e7f8153-9262-4955-8c33-2f0d4de33786","added_by":"auto","created_at":"2025-10-31 04:19:53","extension":"xml","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":149401,"visible":true,"origin":"","legend":"","description":"","filename":"rs79877560structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/e3ec408936f2c3f7a722ad28.xml"},{"id":94814973,"identity":"4beafd1f-7a58-4b18-b9e9-eeb6c1e405f9","added_by":"auto","created_at":"2025-10-31 04:19:54","extension":"html","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":159877,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/dde2b0df863fa61662d943c6.html"},{"id":94814941,"identity":"dc74b6e3-c6ea-4cb0-8ba3-ba7716834c68","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":132916,"visible":true,"origin":"","legend":"\u003cp\u003eThe research framework of the study.\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/27af1c1741666fb0b1b82891.png"},{"id":94814940,"identity":"c3a7c71b-d7ba-45e6-97cb-63b12599d76b","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":4931462,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Location of the study area, (b) distribution of 793 crowdsourced weather stations across Sydney, Australia, (c) Local Climate Zone (LCZ) map of Sydney, and (d) building fraction map from Geoscape dataset.\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/e7d0cb056d8b0aa7db19dca5.png"},{"id":94814974,"identity":"7700a9ef-e53e-4473-97af-01a86eca0f45","added_by":"auto","created_at":"2025-10-31 04:20:09","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":959475,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverall workflow and CNN architecture designed for predicting gridded Ta\u003c/strong\u003e\u003cem\u003e\u003cstrong\u003e \u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003efrom 30 input variables\u003c/strong\u003e. The network comprised six layers in total: one input layer, four intermediate convolutional layers (the first with a 7 × 7 kernel followed by three 1 × 1 convolutions), and one output layer producing a single scalar value representing the predicted Ta for the grid cell at the weather station location.\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/eb79533197654cbaa5a5a688.png"},{"id":94814967,"identity":"fabe854f-1aa5-4a9b-8d76-8c49e47fac14","added_by":"auto","created_at":"2025-10-31 04:19:53","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":4439851,"visible":true,"origin":"","legend":"\u003cp\u003eModelled air temperature maps at 30 m resolution for Sydney, developed using Landsat LST, LCZ, urban form, and terrain data. The top row shows two warm days, while the bottom row illustrates two cold days.\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/d47f1437a92e566152638d39.png"},{"id":94814969,"identity":"63e93bd6-7173-4e95-92a6-0ed240c59d33","added_by":"auto","created_at":"2025-10-31 04:19:53","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":208501,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eModel accuracy. \u003c/strong\u003eObserved T\u003csub\u003ea\u003c/sub\u003e from crowdsourced weather stations in the validation (10 %, n = 1436) and withheld test dataset (15 %, n = 2155) is regressed against model-predicted T\u003csub\u003ea\u003c/sub\u003e to obtain the coefficient of determination (R²) and root mean square error (RMSE). The black dashed line represents the fitted linear regression.\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/2665501e3ab6766b9f6da678.png"},{"id":94814946,"identity":"ef07037c-fb23-4f15-95e8-2aa2fcfe20c2","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":2310217,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eRelative importance of input variables in predicting T\u003c/strong\u003e\u003csub\u003e\u003cstrong\u003ea\u003c/strong\u003e\u003c/sub\u003e\u003cstrong\u003e over Sydney.\u003c/strong\u003e Variable importance was derived from the change in model performance (RMSE) after perturbing the values of each predictor. Urban landscape variables (LCZ and high-resolution gridded (Geoscape) datasets) and terrain features (elevation and distance to coast) are static, while LST and meteorological variables (Tmin, Tmax, solar radiation) vary in time. Variables labelled with ‘mac’ and ‘kan’ correspond to aerodynamic parameters derived following Macdonald et al. [59] and Kanda et al. [60], as implemented in the Geoscape dataset [40].\u003c/p\u003e","description":"","filename":"image7.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/ee6d6a037dfc97923fe0cce4.png"},{"id":94826526,"identity":"dc88c07b-d179-4627-b4fa-3f22b41d3d76","added_by":"auto","created_at":"2025-10-31 06:52:00","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":1564582,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSpatial generalisation performance of the model. \u003c/strong\u003eEach panel shows the relationship between predicted and observed T\u003csub\u003ea\u003c/sub\u003e for the validation and spatially independent test datasets across six train–test sensor combinations. The black dashed line represents the fitted linear regression.\u003c/p\u003e","description":"","filename":"image8.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/1fb5aac161d35d5cbc7ecd55.png"},{"id":94814945,"identity":"670c79ba-86f2-48f7-b908-5384dded803a","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":1421975,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTemporal generalisation performance of the model. \u003c/strong\u003eEach panel shows the relationship between predicted and observed T\u003csub\u003ea\u003c/sub\u003e for the validation and temporally independent test datasets across six train–test day combinations. The black dashed line represents the fitted linear regression.\u003c/p\u003e","description":"","filename":"image9.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/d0cc3f6f31c81d51dbf79c28.png"},{"id":94814942,"identity":"1820f6c2-9f8b-4179-b443-f27fdbdb8e75","added_by":"auto","created_at":"2025-10-31 04:19:52","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":2948352,"visible":true,"origin":"","legend":"","description":"","filename":"image10.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/db92e8fe7e37c320c12e6b7c.png"},{"id":94826084,"identity":"c7431761-e68c-4450-a978-a76286b9c05a","added_by":"auto","created_at":"2025-10-31 06:51:03","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":624905,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eVariation in model performance with station density and spatial distribution.\u003c/strong\u003e (a) Relationship between MAE and distance to the nearest station (km). (b) Variation in MAE across different density classes, expressed as the number of stations within each 1 km grid cell. (c) Spatial distribution of T\u003csub\u003ea\u003c/sub\u003e stations across the study area, colour-coded by the number of stations per 1 km grid cell.\u003c/p\u003e","description":"","filename":"image11.png","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/3795fb63a2054112c22060d9.png"},{"id":94827404,"identity":"0950995c-354f-4464-89f4-e7bfa312a90d","added_by":"auto","created_at":"2025-10-31 06:58:23","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":19207568,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7987756/v1/591a9122-a445-4a34-8788-86007a8709e8.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eDeveloping gridded air temperature data over cities using machine learning\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eUrban heat has become a major concern globally, driven by the combined impacts of urbanization and climate change. It presents unevenly distributed impacts across space and socioeconomic groups, affecting public health [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], energy demand [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], economic productivity [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], and environmental quality [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], thereby constraining efforts toward resilient and climate-adaptive urban development [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eUnderstanding and mitigating these impacts require high-resolution air temperature (T\u003csub\u003ea\u003c/sub\u003e) data that captures the spatial and temporal variability of urban heat. The sparse distribution of ground-based sensors, however, limits the ability to represent fine-scale temperature variations across complex urban landscapes [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. To provide spatially continuous estimates of near-surface air temperature, many studies have relied on mesoscale model simulations. Such endeavours typically require intricate physics-based approaches and significant computational resources [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Alternatively, statistical interpolation techniques based on weather station data have been used [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. However, the limited density and uneven distribution of stations reduce the resolution and accuracy of these methods in capturing urban-scale temperature variability.\u003c/p\u003e\u003cp\u003eSatellite-derived Land Surface Temperature (LST) data offers a partial solution by enabling spatially consistent observations of the land surface globally [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Much of the recent work on mapping urban heat and assessing mitigation strategies has therefore focused on satellite-based LST. However, canopy urban heat, measured by air temperature, is more directly relevant for public health and citizen thermal comfort [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], and there is no universal, direct, or linear relationship between LST and T\u003csub\u003ea\u003c/sub\u003e in urban areas [\u003cspan additionalcitationids=\"CR16\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Despite this, LST still captures key characteristics of the thermal environment, including the spatial variability of surface heating, and leveraging emerging methods, it can be used to estimate near-surface air temperature.\u003c/p\u003e\u003cp\u003eRecent advances in machine learning (ML) have enabled more accurate and scalable approaches to model near-surface air temperature in urban areas. Using these techniques, several studies have integrated satellite-derived LST with ground-based T\u003csub\u003ea\u003c/sub\u003e measurements to produce gridded temperature estimates across cities [\u003cspan additionalcitationids=\"CR19\" citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Most of these studies rely on Moderate Resolution Imaging Spectroradiometer (MODIS) data to provide temperature estimates at a spatial resolution of approximately 1 km. Relatively few have explored the use of higher-resolution sensors such as Landsat or Sentinel, which offer finer spatial detail more suitable for urban-scale analysis [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. Moreover, some studies have focused on mapping monthly average air temperatures rather than capturing finer temporal dynamics. For example, Xu et al. [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] developed a 1-km resolution dataset of monthly air temperature over the Tibetan Plateau using MODIS and auxiliary data from 2001 to 2015, but did not attempt to resolve daily or sub-daily variability. Conversely, other studies have aimed to map air temperature at higher temporal resolutions but relied on sparse temporal sampling. For instance, Ho et al. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] used ML to estimate daily air temperatures over Vancouver, Canada, based on only six individual days scattered across five years, making it difficult to assess the determinants of model performance.\u003c/p\u003e\u003cp\u003eMany of these earlier ML applications were constrained by the availability and distribution of ground observations, which were typically drawn from synoptic weather stations. While ML models offer a cost-effective alternative to physics-based simulations, particularly for capturing fine-scale spatial patterns of urban heat, their accuracy and reliability depend heavily on the quality and representativeness of the training data. Air temperature observations are often collected at peri-urban locations such as airports or parks, and therefore lack the spatial resolution needed to capture the heterogeneity of urban microclimates [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. As a result, ML models trained on such data may struggle to accurately represent intra-urban temperature variability in urban areas with complex surface morphology. Moreover, the scalability and generalisability of these models have received limited attention in the previous studies. Addressing these gaps is needed to develop transferable ML frameworks that can reliably estimate air temperature under diverse conditions.\u003c/p\u003e\u003cp\u003eThe recent increase in crowdsourced weather stations in cities has enabled the analysis of T\u003csub\u003ea\u003c/sub\u003e distribution within heterogeneous urban environments [\u003cspan additionalcitationids=\"CR27 CR28\" citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. Located within residential areas, these sensors provide locally relevant data for assessing thermal exposure in diverse neighbourhoods. Integrating these dense, crowd-sourced observations with satellite data holds great potential for improving the spatial mapping of urban air temperatures. Venter et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] integrated data from over 1,300 crowdsourced weather stations across Oslo in 2018 with satellite observations to model urban air temperatures using a Random Forest approach. While their study is informative and demonstrates the value of leveraging dense crowdsourced measurements alongside satellite data, it is limited to a single year and relies on a conventional, tree-based machine learning model that does not explicitly capture spatial dependencies in the input data. Implementing urban air temperature mapping for real-world applications requires scalable models that not only produce accurate, high-resolution estimates, but also capture spatial relationships and remain robust across multiple years, which has not been the focus of previous efforts.\u003c/p\u003e\u003cp\u003eIn this study, we integrate crowdsourced T\u003csub\u003ea\u003c/sub\u003e data with satellite remote sensing imagery to create high-resolution spatially contiguous maps of T\u003csub\u003ea\u003c/sub\u003e over Sydney, Australia. We developed a machine learning framework using a convolutional neural network (CNN) to infill gaps in the crowdsourced sensor data, using LST derived from Landsat imagery (2019\u0026ndash;2024) and urban datasets characterising urban form, fabric, and geography. We aim to assess how T\u003csub\u003ea\u003c/sub\u003e mapping accuracies are influenced by spatial and temporal variability in the data, as well as the availability of predictor variables and T\u003csub\u003ea\u003c/sub\u003e observations. This allows us to evaluate the generalisability of our model across different urban areas, time periods, and data constraints. The proposed framework offers a transferable approach for high-resolution urban air temperature mapping, with potential applications in climate adaptation and urban planning.\u003c/p\u003e"},{"header":"2. Data and Methods","content":"\u003cp\u003eThe study followed a three-stage workflow integrating data acquisition, model development, and generalisation assessment. First, satellite, meteorological, and urban form datasets were compiled and pre-processed to generate predictor variables. Second, a CNN model was developed to estimate gridded air temperature using crowdsourced sensor data for training. Finally, model performance, generalisability, and transferability were evaluated through multiple scenarios, including spatial and temporal hold-out tests and reduced-data experiments. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarises the overall workflow.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1. Study area\u003c/h2\u003e\u003cp\u003eThe analysis is developed and tested for Sydney, located on the southeast coast of Australia (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), which has a mean annual temperature of 17.8\u0026deg;C and experiences mild winters and warm summers with 1200 mm of precipitation throughout the year. Sydney houses over 5.5\u0026nbsp;million residents [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. The city\u0026rsquo;s rapid urban expansion has significantly increased its vulnerability to urban warming effects [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. Existing studies in Sydney have often relied on sparse, fixed-point air temperature measurements from meteorological stations and local monitoring networks [\u003cspan additionalcitationids=\"CR33\" citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. To effectively characterise urban heat, more detailed spatial temperature data is needed. This approach would be particularly valuable for monitoring canopy urban heat and understanding the key drivers of intra-urban temperature variability across Sydney.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2. City-descriptive datasets\u003c/h2\u003e\u003cp\u003ePrevious studies have shown that the physical characteristics in cities, such as building morphology and surface cover, can significantly affect spatial variations in urban temperatures [\u003cspan additionalcitationids=\"CR36 CR37\" citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. To represent these characteristics, we used the Local Climate Zone (LCZ) data, with a resolution of 100 m, derived from the global dataset developed by Demuzere et al. [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. While the LCZ dataset is valuable for representing urban form and surface cover, it is a class-based approach that lacks building height and fine-scale spatial detail. To gain a more detailed representation of the urban landscape, we incorporated a Lidar-based \u0026ldquo;Geoscape\u0026rdquo; dataset, which is based on building-resolved, 3D land cover information at 2 m resolution. This dataset contains over twenty variables describing the urban form and fabric, including building height, tree height, and fractional cover of buildings, vegetation, roads, and water, all aggregated to a 300 m resolution [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]. We also collected distance to the ocean and elevation to account for topographic and coastal influences on temperature patterns. Distance to the ocean was derived from a NASA-provided layer at 0.01\u0026deg; resolution (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://oceancolor.gsfc.nasa.gov/docs/distfromcoast/\u003c/span\u003e\u003cspan address=\"https://oceancolor.gsfc.nasa.gov/docs/distfromcoast/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e)\u003c/span\u003e, and elevation was obtained from the NASADEM dataset at 30 m spatial resolution.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3. Satellite remote sensing data\u003c/h2\u003e\u003cp\u003eWe selected Landsat 8 and 9 scenes covering Sydney for 61 cloud-less days between 2019 and 2024. The imagery corresponds to a mean local overpass time of 10:00\u0026thinsp;\u0026plusmn;\u0026thinsp;15 min, and LST data were generated at 30 m spatial resolution using the Google Earth Engine (GEE) platform [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e]. LST was retrieved using the Statistical Mono-Window (SMW) algorithm [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e] implemented in GEE. This method estimates LST from top-of-atmosphere brightness temperatures in the thermal infrared band, incorporating surface emissivity from the ASTER Global Emissivity Database and applying a vegetation correction based on NDVI. The LST data derived from this method have been shown to yield satisfactory accuracy, with an RMSE of 1K (for further details, see Ermida et al. [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e]).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.4. Citizen Weather Station data\u003c/h2\u003e\u003cp\u003eWe used open-access air temperature data from 793 citizen weather stations distributed across Sydney (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb), obtained via the public Netatmo API (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://dev.netatmo.com/apidocumentation/weather\u003c/span\u003e\u003cspan address=\"https://dev.netatmo.com/apidocumentation/weather\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). The dataset captured measurements for 61 selected cloudless days between 2019 and 2024. Crowdsourced temperature data can be affected by factors such as sensor misplacement, direct solar exposure, or device malfunctions [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e, \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. To minimise these issues and ensure data reliability, a quality control procedure was applied using the CrowdQC\u0026thinsp;+\u0026thinsp;package [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e], which identifies and removes statistically implausible observations without requiring reference meteorological data. The first step flagged stations with duplicate geographic coordinates. The second step identified statistical outliers using a z-score approach, comparing each observation against the distribution of all stations for the same time step. The third step removed entire months of data from stations where more than 20% of values were excluded in the previous steps. The fourth step filtered out stations likely located indoors, identified by a weak correlation between their hourly T\u003csub\u003ea\u003c/sub\u003e measurements and the spatial median of all stations for the same period. The final step excluded unrealistically high values detected as outliers when compared with neighbouring stations. For each station and date, we then extracted the LST value from the pixel containing the station\u0026rsquo;s location and obtained the T\u003csub\u003ea\u003c/sub\u003e measurement corresponding to the satellite overpass time (10:00 local time). We also obtained daily minimum and maximum air temperatures and solar radiation from the Australian Bureau of Meteorology\u0026rsquo;s Observatory Hill station via Climate Data Online (\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003ebom.gov.au\u003c/span\u003e) to capture seasonal variations in the data.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e2.5. Air temperature predictions\u003c/h2\u003e\u003cp\u003eTo map gridded T\u003csub\u003ea\u003c/sub\u003e data, we used a custom Convolutional Neural Network (CNN), which is a deep learning method suitable for capturing spatial patterns and complex, non-linear relationships in spatial data [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]. Previous studies on gridded T\u003csub\u003ea\u003c/sub\u003e estimation have often relied on ensemble-based machine learning algorithms such as Random Forest (RF) and Gradient Boosting (GB) [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e]. While these methods can model complex relationships between predictors and T\u003csub\u003ea\u003c/sub\u003e, they treat each location independently of their neighbouring areas and therefore might not be able to explicitly capture spatial dependencies and patterns. One possible workaround is to incorporate spatial context by augmenting the predictor set with neighbourhood statistics or by extracting values from surrounding pixels (often referred to as context windows or spatial chips) [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e, \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e]. However, this approach increases data redundancy, may dilute local-scale features, and does not fully exploit the hierarchical feature learning capabilities of deep learning methods such as CNNs [\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe CNN was trained to predict T\u003csub\u003ea\u003c/sub\u003e at grid cells with available observations, using a 7 \u0026times; 7 neighbourhood of surrounding grid cells, where the observed T\u003csub\u003ea\u003c/sub\u003e at the centre of each patch served as the target variable and was obtained from crowdsourced measurements. The model used 30 input variables as predictors, including remotely sensed data, urban landscape characteristics, and meteorological variables, as detailed in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. All ancillary data variables were resampled to match the spatial resolution and grid of the LST data prior to model training. The trained model was then used to predict T\u003csub\u003ea\u003c/sub\u003e for each grid cell across the study area, using the corresponding 7 \u0026times; 7 patch of input predictors centred on the target grid cell.\u003c/p\u003e\u003cp\u003eTo determine the optimal patch size, we tested multiple window sizes (3 \u0026times; 3, 5 \u0026times; 5, 7 \u0026times; 7, and larger). Performance improved with increasing patch size, likely due to the additional spatial information captured. However, beyond 7 \u0026times; 7 no substantial improvement was observed. Therefore, a 7 \u0026times; 7 size was selected as the best balance between accuracy and computational efficiency. We applied z-score normalization to the continuous input layers using the mean and standard deviation computed from the training set, while excluding the categorical LCZ layer, which was one-hot encoded. Each 7 \u0026times; 7 \u0026times; 30 patch was provided as input to the CNN. The first convolutional layer applied a 7 \u0026times; 7 kernel, reducing the spatial dimensions to 1 \u0026times; 1 while increasing feature depth. Subsequent 1 \u0026times; 1 convolutions progressively expanded the feature space to 1 \u0026times; 1 \u0026times; 2400. The final layer produced a single scalar output, representing the predicted T\u003csub\u003ea\u003c/sub\u003e for the grid cell at the sensor location, which was compared against observed values using the loss function. We used a custom loss function that combines Mean Squared Error (MSE) and the coefficient of determination (R\u0026sup2;) to balance prediction accuracy and overall fit. The loss is computed as a weighted combination of MSE and R\u0026sup2;, where a moderate weighting factor (α\u0026thinsp;=\u0026thinsp;2/3) was adopted to place slightly more emphasis on minimizing prediction errors while still encouraging a strong explained variance between predictions and targets. This balance provided stable convergence and consistent performance across experiments. The overall workflow and CNN architecture used in this study are summarised in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e2.6. Generalisation and spatial transferability assessment\u003c/h2\u003e\u003cp\u003eTo evaluate model generalisability, we designed two scenarios. Scenario 1 tested spatial generalisation by predicting T\u003csub\u003ea\u003c/sub\u003e for grid cells or sensors excluded from the training data. Scenario 2 assessed temporal generalisation by predicting T\u003csub\u003ea\u003c/sub\u003e for days not used during training. For both scenarios, the data were partitioned into six mutually exclusive subsets. In each iteration, one subset (either a group of sensors in scenario 1 or a set of days in scenario 2) was held out exclusively for testing, while the remaining subsets were used for training and validation. This procedure was repeated six times so that all subsets were used once as the test set.\u003c/p\u003e\u003cp\u003eTo further assess the spatial and temporal transferability of our predictions, we designed two complementary scenarios. The first scenario tested model performance under conditions where high-resolution local datasets are not accessible. In this case, we restricted the predictor set from 30 variables to the seven globally available features to examine the model\u0026rsquo;s robustness to reduced feature availability. The second scenario focused on the influence of the number and density of T\u003csub\u003ea\u003c/sub\u003e observations. Here, 15% of the data were reserved for independent testing, while the remaining 85% were used for training and validation (10%). To assess sensitivity to the number of available stations, we progressively increased the proportion of training data from the 85% subset in 2.5% increments (i.e., 2.5%, 5%, 7.5%, and so on) and evaluated performance on the same held-out test set after each step. This approach simulated the gradual expansion of station networks within a city while maintaining a fixed test dataset. The entire process was repeated three times with different random samplings of the training data to ensure the stability of the results. Model accuracy metrics (MAE and R\u0026sup2;) were computed for each subset size to evaluate how model performance responds to varying numbers of T\u003csub\u003ea\u003c/sub\u003e sensors and to assess performance stability under reduced training data. We further examined how the density and spatial distribution of T\u003csub\u003ea\u003c/sub\u003e sensors influence model performance. Using a leave-one-station-out approach, we first obtained prediction errors for each station. These errors were then analysed against the spatial density of nearby sensors, expressed as the number of stations per km\u0026sup2;. We further examined the relationship between MAE and the distance to the nearest station to quantify how spatial isolation affects prediction accuracy.\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Results and Discussion","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e3.1. Air temperature mapping and model performance\u003c/h2\u003e\u003cp\u003eUsing all input variables and the proposed machine learning framework, we developed high-resolution maps of air temperature across Sydney. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e presents gridded daytime T\u003csub\u003ea\u003c/sub\u003e for four representative days. The modelled T\u003csub\u003ea\u003c/sub\u003e maps reveal spatial gradients across Sydney on both warm and cold days. On the warm days, higher temperatures were mostly concentrated in the western and south-western suburbs, likely due to dense built-up areas, limited vegetation cover, and distance to the ocean. On the cold days, coastal moderation was also evident, with relatively higher T\u003csub\u003ea\u003c/sub\u003e near the shoreline. Cooler areas were observed over vegetated and elevated regions, particularly in the northern parts of the city. These patterns demonstrate that Sydney\u0026rsquo;s air temperature variability is greatly influenced by coastal proximity, urban density, and land-surface characteristics, which the model effectively captured across warm and cold conditions. Similar spatial contrasts have been reported by Potgieter et al. [\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e], who found that distance from the ocean strongly influences air temperature variations across Sydney.\u003c/p\u003e\u003cp\u003eThe high-resolution T\u003csub\u003ea\u003c/sub\u003e maps developed in this study enable detailed identification of hot and cool spots across the city. These data can guide targeted mitigation strategies, such as vegetation planting or increasing surface albedo and allow their effectiveness to be monitored over time. They are also valuable for validating urban climate models that incorporate urban canopy schemes [\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e, \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e]. More broadly, fine-scale T\u003csub\u003ea\u003c/sub\u003e mapping enhances understanding of the factors driving urban cooling and supports improved heat-exposure and risk assessments.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e shows the overall performance of the CNN model. The dataset was randomly divided into training (75%), validation (10%), and test (15%) subsets. The model performed well, achieving an R\u0026sup2; of 0.97 and an RMSE of 0.91\u0026deg;C on both validation and test sets. This accuracy compares favourably with other recent studies applying machine learning to air temperature estimation, Reported RMSE values in previous works typically range between 1.0 and 3.0\u0026deg;C and R\u0026sup2; values between 0.3 and 0.94, depending on spatial resolution, data sources, model architecture, and urban complexity [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e, \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e, \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e]. For instance, a recent study in Morocco, Africa, which used satellite-derived predictors and tested several algorithms, achieved its best performance with a Multilayer Perceptron (MLP), obtaining an RMSE below 1.6\u0026deg;C at 1 km resolution [\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e]. At a finer spatial resolution, Venter et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] applied a Random Forest model in Oslo, Norway, using remote sensing and crowdsourced data, reporting an RMSE of 1.85\u0026deg;C and an R\u0026sup2; of 0.05 for daily maximum air temperature. Compared with these approaches, the present CNN framework achieved both finer spatial resolution (30 m) and higher predictive accuracy, highlighting its ability to capture spatial patterns in urban air temperature while maintaining strong predictive performance. We also assessed how model accuracy varied within the city using the LCZ classification (not shown). The results indicated lower mapping errors in compact and open low-rise areas, consistent with the findings of Venter et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] for Oslo. However, because the number of stations differed substantially among LCZ classes, the comparison was inherently biased.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003e3.2. Predictor variable importance\u003c/h2\u003e\u003cp\u003eWe evaluated the contribution of each predictor to model performance using a perturbation approach. For each variable, its values in the test dataset were replaced with a constant outside the range of data to remove variability. The resulting decrease in predictive accuracy (RMSE) was compared across all predictors, and predictor variables causing the largest performance drops were ranked as most important. In our model, the most important predictors were satellite-derived LST, solar radiation, and daily maximum and minimum air temperature (Tmax and Tmin), followed by LCZ (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e). The high importance of LST and solar radiation aligns with previous findings highlighting their strong influence on maximum air temperature, particularly during hot summer days [\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e]. The model appears to rely more strongly on variables that exhibit temporal variability than on static predictors, which aligns with observations reported by Hobeichi et al. [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]. The strong contribution of solar radiation, Tmax, and Tmin in our results reflects their role in providing the model with information on day-to-day and seasonal variations in air temperature. Without such temporal indicators, the model would primarily rely on spatial features, which are all static and would therefore struggle to capture time-related variability in T\u003csub\u003ea\u003c/sub\u003e. After LCZ, tree height and water fraction were the most influential predictors. Tree height likely reflects the cooling effect of vegetation through shading and evapotranspiration, while water fraction captures the moderating influence of coastal and inland water bodies on near-surface air temperature.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e3.3. Generalisation and spatial transferability assessment\u003c/h2\u003e\u003cp\u003eWe evaluated model generalisability under two scenarios. Scenario 1 addressed spatial transferability, in which the model was tested on T\u003csub\u003ea\u003c/sub\u003e sensors not included in training. In this setup, ~ 660 sensors were used for training and validation, while the remaining\u0026thinsp;~\u0026thinsp;130 were held out exclusively for testing. Six different train\u0026ndash;test combinations were created based on the total number of unique sensors (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e). Model performance on the validation set remained consistent with the all-data baseline (R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.97, RMSE\u0026thinsp;=\u0026thinsp;0.86\u0026ndash;0.94\u0026deg;C). For the test sets, R\u003csup\u003e2\u003c/sup\u003e decreased slightly to 0.91\u0026ndash;0.93, with RMSE ranging from 1.27 to 1.44\u0026deg;C in all splits, which still shows robust performance. These results indicate that the model can effectively predict air temperature in unseen locations based on spatial patterns learned from the training data. Similar spatial hold-out testing has been applied in a previous study in Guangzhou, China [\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e], based on observations recorded by around 300 stations from March 2020 to March 2021. They found a marginal increase in RMSE compared to the full-data baseline, indicating that the model effectively captured spatial patterns and could predict air temperature at unobserved locations.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eScenario 2 examined temporal transferability by testing the model on days not included in training. The dataset comprised 61 cloudless days across multiple seasons from 2019 to 2024. In each experiment, data from 50 days were used for training and validation, and the remaining 10\u0026ndash;11 days for testing, producing six different train\u0026ndash;test combinations (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e). While the all-data baseline yielded R\u0026sup2; around 0.97, for the test set, Scenario 2 results ranged from R\u0026sup2; = 0.66 to 0.93, with RMSE values between 1.52 to 2.6\u0026deg;C. These values still remain within the performance ranges reported in previous ML-based air-temperature mapping studies discussed earlier. The reduced accuracy suggests that the model has greater difficulty capturing atmospheric or weather conditions not represented in the training set, reflecting the inherently more complex and less predictable nature of temporal patterns compared to spatial patterns. Unlike an earlier MODIS-based study in Los Angeles and Seoul [\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e], which reported that incorporating LST observations from additional days did not necessarily improve T\u003csub\u003ea\u003c/sub\u003e estimation, our findings suggest that expanding the temporal diversity of the training data is an important factor in enhancing model robustness, particularly when it comes to creating maps for future dates. Incorporating a broader range of days, including extreme or atypical conditions, can further strengthen temporal generalisation.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eWe assessed the model\u0026rsquo;s spatial transferability under two additional scenarios. The first scenario examined model performance under limited input feature availability, representing cases where high-resolution local datasets are not accessible. The full model was trained using 30 predictor variables, comprising both globally available and locally specific features. However, only seven of these variables are accessible globally at comparable resolution: LST, LCZ, solar radiation, Tmax, Tmin, elevation, and distance to coast, where five of these predictors were identified among the most important contributing features (shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e). To assess transferability, we trained an alternative model using only these seven global predictors (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e). Compared to the full-input model (R\u0026sup2; = 0.97, RMSE\u0026thinsp;=\u0026thinsp;0.91\u0026deg;C), the reduced-input version showed only a marginal decline in performance (R\u0026sup2; = 0.96, RMSE\u0026thinsp;=\u0026thinsp;0.98\u0026deg;C). Thus, the model still maintained strong predictive capacity, indicating robustness to reduced feature availability. We further repeated the spatial and temporal generalisation tests using only the seven global predictors to evaluate how well the reduced-input model performs under unseen conditions. The results remained close to those of the full-input model, demonstrating that the proposed framework can generalise well across space and time, even with a limited set of globally available variables. This suggests strong potential for transferring the approach to other cities worldwide where local high-resolution datasets are not accessible.\u003c/p\u003e\u003cp\u003eThe availability of sufficient ground-truth weather station data in cities for training and validation is another important factor influencing model performance. Therefore, the second scenario assessed the model\u0026rsquo;s sensitivity to the density of available T\u003csub\u003ea\u003c/sub\u003e observations. Because the spatial coverage of station data varies greatly among cities, we tested how progressively increasing the number of T\u003csub\u003ea\u003c/sub\u003e sensors affected model accuracy. The results showed that performance stabilised when approximately 20% of the stations (~\u0026thinsp;160) were used for training, with only marginal improvements from adding additional sensors (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e). This implies that the proposed approach can be effectively extended to cities with limited observational networks.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eLastly, we evaluated the influence of station density, rather than just the total number of sensors, since cities differ in size and spatial coverage. We observed that station density had a limited influence on model accuracy (Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e). Increasing station density from one to six per 1 km grid cell did not substantially reduce the mean absolute error (MAE) (Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003eb). Model performance remained largely stable across density levels, except for very isolated stations, where errors tended to be higher (Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003ea). These findings suggest that the approach is technically transferable: the model structure and methodology can be applied to cities worldwide, even where high-resolution local datasets or dense ground-truth T\u003csub\u003ea\u003c/sub\u003e observations are lacking. Using Random Forest regression modelling, Venter et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] estimated that around 250 stations are required to train and validate a model effectively and noted that most European cities meet this threshold. However, such station densities are uncommon in many cities, particularly in the Global South. In cases where crowdsourced networks are sparse, official weather stations can provide a viable alternative source of training data. Incorporating these stations could therefore extend the applicability of the proposed ML framework to data-limited contexts, although local validation would remain essential to ensure reliable performance.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e3.4. Limitations and future work\u003c/h2\u003e\u003cp\u003eDespite the benefits of this developed dataset and modelling framework, some limitations need to be acknowledged. As satellite LST cannot be collected on cloudy days, the resulting air temperature maps are biased toward clear-sky conditions. Moreover, this study was restricted to daytime conditions, since the Landsat satellites only provide surface temperature observations during the day. The fixed overpass time of ~\u0026thinsp;10:00 a.m. limits the model to represent morning conditions, which may differ from temperature patterns observed later in the day. Thus, further investigation is needed to incorporate multi-temporal or higher-frequency satellite observations to better capture diurnal variations in air temperature and improve the robustness of gridded estimates. In addition, expanding the set of predictor variables to include key meteorological factors such as wind speed and humidity could further enhance the accuracy of air temperature estimation.\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Conclusion","content":"\u003cp\u003eHigh-resolution spatial air temperature data is critical for understanding urban heat, yet it is often difficult to obtain using sparse conventional weather station networks. While satellite-derived LST provides useful spatial coverage, it is not directly relevant to human thermal exposure and health in the same way that T\u003csub\u003ea\u003c/sub\u003e is. Crowdsourced T\u003csub\u003ea\u003c/sub\u003e data offer a valuable complement, as these stations are densely located within residential areas where people live and where heat anomalies most strongly affect population health. Their spatial coverage, however, is limited. Our study demonstrated the benefits of integrating satellite observations with crowdsourced T\u003csub\u003ea\u003c/sub\u003e data. By combining high-resolution satellite imagery, urban landscape metrics, and terrain information within a CNN framework, we produced 30 m resolution maps of urban T\u003csub\u003ea\u003c/sub\u003e over Sydney, Australia. Compared with conventional machine learning approaches, CNNs provide clear advantages by capturing spatial patterns and contextual information more effectively.\u003c/p\u003e\u003cp\u003eThe model achieved strong performance, with the most influential predictors being LST, solar radiation, daily maximum and minimum T\u003csub\u003ea\u003c/sub\u003e, and LCZ. Vegetation and water-related variables also played an important role in modulating local microclimates. In terms of generalisability, the framework performed consistently well across space, while temporal transfer proved more challenging, highlighting the importance of incorporating a wider range of days to improve robustness. Our transferability analysis showed that high-resolution urban datasets, while useful, are not essential. Models trained using only globally available predictors performed nearly as well as those using detailed local datasets, supporting the broader applicability of this approach in data-limited contexts. Moreover, the model maintained stable performance even when the number and density of training stations were substantially reduced, demonstrating robustness to sparse observational networks. These findings suggest that the proposed framework can be effectively applied to other cities worldwide, including those with limited data availability.\u003c/p\u003e\u003cp\u003eOverall, this study demonstrates that integrating crowdsourced air temperature measurements with satellite and urban data through deep learning provides an effective, scalable, and transferable method for mapping fine-scale urban air temperatures. Such maps are directly relevant for stakeholders in urban planning and public health, offering critical insights for climate change adaptation and the design of heat-resilient cities in a rapidly urbanising world.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003ch2\u003eDeclaration of competing interest\u003c/h2\u003e\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e\u003cp\u003eThis work was supported by the Australian Research Council as part of the Centre of Excellence for Climate Extremes (CE170100023). We also acknowledge support from the Minderoo Foundation (PS74451).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eTan J et al (2010) The urban heat island and its impact on heat waves and human health in Shanghai. Int J Biometeorol 54:75\u0026ndash;84. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00484-009-0256-x\u003c/span\u003e\u003cspan address=\"10.1007/s00484-009-0256-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSantamouris M, Cartalis C, Synnefa A, Kolokotsa D (2015) On the impact of urban heat island and global warming on the power demand and electricity consumption of buildings\u0026mdash;A review. Energy Build 98:119\u0026ndash;124. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.enbuild.2014.09.052\u003c/span\u003e\u003cspan address=\"10.1016/j.enbuild.2014.09.052\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHerbel I, Croitoru A-E, Rus AV, Roşca CF, Harpa GV, Ciupertea A-F, Rus I (2018) The impact of heat waves on surface urban heat island and local economy in Cluj-Napoca city, Romania. Theoret Appl Climatol 133(3):681\u0026ndash;695. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00704-017-2196-4\u003c/span\u003e\u003cspan address=\"10.1007/s00704-017-2196-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAw J, Kleeman MJ (2003) Evaluating the first-order effect of intraannual temperature variability on urban air pollution. J Geophys Research: Atmos 108(D12). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ep. https://doi.org/10.1029/2002JD002688\u003c/span\u003e\u003cspan address=\"p. 10.1029/2002JD002688\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNazarian N et al (2022) Integrated assessment of urban overheating impacts on human life. Earths Future 10(8). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1029/2022EF002682\u003c/span\u003e\u003cspan address=\"10.1029/2022EF002682\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. p. e2022EF002682\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBaranka G, Boz\u0026oacute; L, Ciglič R, Komac B (2016) Urban Heat Island Gold Standard and Urban Heat Island Atlas: Gold Standard for UHI Measurements and Introduction of The Central-European Urban Heat Island Atlas. Counteracting Urban Heat Island Effects in a Global Climate Change Scenario. Springer, pp 41\u0026ndash;70. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/978-3-319-10425-6_2\u003c/span\u003e\u003cspan address=\"10.1007/978-3-319-10425-6_2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang K, Jiang S, Wang J, Zhou C, Wang X, Lee X (2017) Comparing the diurnal and seasonal variabilities of atmospheric and surface urban heat islands based on the Beijing urban meteorological network. J Geophys Research: Atmos 122(4):2131\u0026ndash;2154. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/2016JD025304\u003c/span\u003e\u003cspan address=\"10.1002/2016JD025304\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHassani A, Santos GS, Schneider P, Castell N (2024) Interpolation, satellite-based machine learning, or meteorological simulation? A comparison analysis for spatio-temporal mapping of mesoscale urban air temperature. Environ Model Assess 29(2):291\u0026ndash;306. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10666-023-09943-9\u003c/span\u003e\u003cspan address=\"10.1007/s10666-023-09943-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMartilli A et al (2024) WRF-Comfort: simulating microscale variability in outdoor heat stress at the city scale with a mesoscale model. Geosci Model Dev 17(12):5023\u0026ndash;5039. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5194/gmd-17-5023-2024\u003c/span\u003e\u003cspan address=\"10.5194/gmd-17-5023-2024\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang M, He G, Zhang Z, Wang G, Zhang Z, Cao X, Wu Z, Liu X (2017) Comparison of spatial interpolation and regression analysis models for an estimation of monthly near surface air temperature in China. Remote Sens 9(12):1278. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/rs9121278\u003c/span\u003e\u003cspan address=\"10.3390/rs9121278\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFlorio E, Lele S, Chi Chang Y, Sterner R, Glass G (2004) Integrating AVHRR satellite data and NOAA ground observations to predict surface air temperature: a statistical approach. Int J Remote Sens 25(15):2979\u0026ndash;2994. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/01431160310001624593\u003c/span\u003e\u003cspan address=\"10.1080/01431160310001624593\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhou D et al (2018) Satellite remote sensing of surface urban heat islands: Progress, challenges, and perspectives. Remote Sens 11(1):48. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/rs11010048\u003c/span\u003e\u003cspan address=\"10.3390/rs11010048\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNaserikia M, Hart MA, Nazarian N, Bechtel B (2022) Background climate modulates the impact of land cover on urban surface temperature. Sci Rep 12(1):15433. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-022-19431-x\u003c/span\u003e\u003cspan address=\"10.1038/s41598-022-19431-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMartilli A, Krayenhoff ES, Nazarian N (2020) Is the Urban Heat Island intensity relevant for heat mitigation studies? Urban Clim 31:100541. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.uclim.2019.100541\u003c/span\u003e\u003cspan address=\"10.1016/j.uclim.2019.100541\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNaserikia M, Hart MA, Nazarian N, Bechtel B, Lipson M, Nice KA (2023) Land surface and air temperature dynamics: The role of urban form and seasonality. Sci Total Environ 905:167306. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.scitotenv.2023.167306\u003c/span\u003e\u003cspan address=\"10.1016/j.scitotenv.2023.167306\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhan W et al (2025) Satellite-derived Land Surface Temperatures Strongly Mischaracterise Urban Heat Hazard. \u003cem\u003earXiv preprint arXiv:2509.16568\u003c/em\u003e, p. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2509.16568\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2509.16568\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVenter ZS, Chakraborty T, Lee X (2021) Crowdsourced air temperatures contrast satellite measures of the urban heat island and its mechanisms. Sci Adv 7(22):eabb9569. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1126/sciadv.abb9569\u003c/span\u003e\u003cspan address=\"10.1126/sciadv.abb9569\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen Y, Sun H, Li J (2016) Estimating daily maximum air temperature with MODIS data and a daytime temperature variation model in Beijing urban area. Remote Sens Lett 7(9):865\u0026ndash;874. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/2150704X.2016.1193792\u003c/span\u003e\u003cspan address=\"10.1080/2150704X.2016.1193792\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNoi PT, Kappas M, Degener J (2016) Estimating daily maximum and minimum land air surface temperature using MODIS land surface temperature data and ground truth data in Northern Vietnam. Remote Sens 8(12):1002. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/rs8121002\u003c/span\u003e\u003cspan address=\"10.3390/rs8121002\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShen H, Jiang Y, Li T, Cheng Q, Zeng C, Zhang L (2020) Deep learning-based air temperature mapping by fusing remote sensing, station, simulation and socioeconomic data. Remote Sens Environ 240:111692. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.rse.2020.111692\u003c/span\u003e\u003cspan address=\"10.1016/j.rse.2020.111692\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVenter ZS, Brousse O, Esau I, Meier F (2020) Hyperlocal mapping of urban air temperature using remote sensing and crowdsourced weather data. Remote Sens Environ 242:111791. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.rse.2020.111791\u003c/span\u003e\u003cspan address=\"10.1016/j.rse.2020.111791\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXu Y, Knudby A, Shen Y, Liu Y (2018) Mapping monthly air temperature in the Tibetan Plateau from MODIS data based on machine learning methods. IEEE J Sel Top Appl earth observations remote Sens 11(2):345\u0026ndash;354. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/JSTARS.2017.2787191\u003c/span\u003e\u003cspan address=\"10.1109/JSTARS.2017.2787191\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHo HC, Knudby A, Sirovyak P, Xu Y, Hodul M, Henderson SB (2014) Mapping maximum urban air temperature on hot summer days. Remote Sens Environ 154:38\u0026ndash;45. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.rse.2014.08.012\u003c/span\u003e\u003cspan address=\"10.1016/j.rse.2014.08.012\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSchl\u0026uuml;nzen KH, Grimmond S, Baklanov A (2023) Guidance on measuring, modelling and monitoring the canopy layer urban heat island (CL-UHI). World Meteorological Origanization\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWorld Meteorological Organization (WMO) (2021) Guide to Instruments and Methods of Observation \u003cem\u003eVol. V\u0026mdash;Quality Assurance and Management of Observing Systems\u003c/em\u003e. WMO Geneva, Switzerland\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFenner D, Meier F, Bechtel B, Otto M, Scherer D (2017) Intra and inter local climate zone variability of air temperature as observed by crowdsourced citizen weather stations in Berlin, Germany. Meteorol Z 26:525\u0026ndash;547. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1127/metz/2017/0861\u003c/span\u003e\u003cspan address=\"10.1127/metz/2017/0861\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMeier F, Fenner D, Grassmann T, Otto M, Scherer D (2017) Crowdsourcing air temperature from citizen weather stations for urban climate research. Urban Clim 19:170\u0026ndash;191. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.uclim.2017.01.006\u003c/span\u003e\u003cspan address=\"10.1016/j.uclim.2017.01.006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMuller C, Chapman L, Johnston S, Kidd C, Illingworth S, Foody G, Overeem A, Leigh R (2015) Crowdsourcing for climate and atmospheric sciences: current status and future potential. Int J Climatol 35(11):3185\u0026ndash;3203. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/joc.4210\u003c/span\u003e\u003cspan address=\"10.1002/joc.4210\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChapman L, Bell C, Bell S (2016) Can the crowdsourcing data paradigm take atmospheric science to a new level? A case study of the Urban Heat Island of London quantified using Netatmo weather stations. Int J Climatol. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/joc.4940\u003c/span\u003e\u003cspan address=\"10.1002/joc.4940\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e(ABS) (2024) A.B.o.S., \u003cem\u003eAustralian Statistical Geography Standard (ASGS)\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eArg\u0026uuml;eso D, Evans JP, Pitman AJ, Di A, Luca (2015) Effects of city expansion on heat stress under climate change conditions. PLoS ONE 10(2):e0117066. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1371/journal.pone.0117066\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0117066\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePfautsch S, Wujeska-Klause A, Walters JR (2025) Spatiotemporal variation of intra-urban heat and heatwaves across Greater Sydney, Australia. Weather Clim Extremes 47:100741. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.wace.2024.100741\u003c/span\u003e\u003cspan address=\"10.1016/j.wace.2024.100741\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSidiqui P, Huete A, Devadas R (2016) Spatio-temporal mapping and monitoring of Urban Heat Island patterns over Sydney, Australia using MODIS and Landsat-8. in \u003cem\u003e4th International Workshop on Earth observation and Remote Sensing applications (EORSA)\u003c/em\u003e. IEEE\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSantamouris M, Haddad S, Fiorito F, Osmond P, Ding L, Prasad D, Zhai X, Wang R (2017) Urban heat island and overheating characteristics in Sydney, Australia. An analysis of multiyear measurements. Sustainability 9(5):712. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/su9050712\u003c/span\u003e\u003cspan address=\"10.3390/su9050712\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNice KA et al (2022) Isolating the impacts of urban form and fabric from geography on urban heat and human thermal comfort. Build Environ 224:109502. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.buildenv.2022.109502\u003c/span\u003e\u003cspan address=\"10.1016/j.buildenv.2022.109502\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKonarska J, Holmer B, Lindberg F, Thorsson S (2016) Influence of vegetation and building geometry on the spatial variations of air temperature and cooling rates in a high-latitude city. Int J Climatol 36(5). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ep. https://doi.org/10.1002/joc.4502\u003c/span\u003e\u003cspan address=\"p. 10.1002/joc.4502\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYin C, Yuan M, Lu Y, Huang Y, Liu Y (2018) Effects of urban form on the urban heat island effect based on spatial regression model. Sci Total Environ 634:696\u0026ndash;704. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.scitotenv.2018.03.350\u003c/span\u003e\u003cspan address=\"10.1016/j.scitotenv.2018.03.350\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNaserikia M, Asadi Shamsabadi E, Rafieian M, Leal Filho W (2019) The urban heat island in an urban context: A case study of Mashhad, Iran. Int J Environ Res Public Health 16(3):313. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/ijerph16030313\u003c/span\u003e\u003cspan address=\"10.3390/ijerph16030313\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDemuzere M, Kittner J, Martilli A, Mills G, Moede C, Stewart ID, Van Vliet J, Bechtel B (2022) A global map of Local Climate Zones to support earth system modelling and urban scale environmental science. \u003cem\u003eEarth System Science Data Discussions\u003c/em\u003e, 2022: pp. 1\u0026ndash;57. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5194/essd-14-3835-2022\u003c/span\u003e\u003cspan address=\"10.5194/essd-14-3835-2022\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLipson MJ, Nazarian N, Hart MA, Nice KA, Conroy B (2022) A transformation in city-descriptive input data for urban climate models. Front Environ Sci 10:866398. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fenvs.2022.866398\u003c/span\u003e\u003cspan address=\"10.3389/fenvs.2022.866398\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGorelick N, Hancher M, Dixon M, Ilyushchenko S, Thau D, Moore R (2017) Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sens Environ 202:18\u0026ndash;27. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.rse.2017.06.031\u003c/span\u003e\u003cspan address=\"10.1016/j.rse.2017.06.031\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDuguay-Tetzlaff A et al (2015) Meteosat land surface temperature climate data record: Achievable accuracy and potential uncertainties. Remote Sens 7(10):13139\u0026ndash;13156. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/rs71013139\u003c/span\u003e\u003cspan address=\"10.3390/rs71013139\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eErmida SL, Soares P, Mantas V, G\u0026ouml;ttsche F-M, Trigo IF (2020) Google earth engine open-source code for land surface temperature estimation from the landsat series. Remote Sens 12(9):1471. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/rs12091471\u003c/span\u003e\u003cspan address=\"10.3390/rs12091471\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVarentsov M, Konstantinov P, Shartova N, Samsonov T, Kargashin P, Varentsov A, Fenner D, Meier F (2020) Urban heat island of the Moscow megacity: the long-term trends and new approaches for monitoring and research based on crowdsourcing data. in \u003cem\u003eIOP Conference Series: Earth and Environmental Science\u003c/em\u003e. IOP Publishing. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://doi.org/10.1088/1755-1315/606/1/012063\u003c/span\u003e\u003cspan address=\"10.1088/1755-1315/606/1/012063\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFenner D, Bechtel B, Demuzere M, Kittner J, Meier F (2021) CrowdQC+\u0026mdash;a quality-control for crowdsourced air-temperature observations enabling world-wide urban climate applications. Front Environ Sci 9:720747. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fenvs.2021.720747\u003c/span\u003e\u003cspan address=\"10.3389/fenvs.2021.720747\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhao X, Wang L, Zhang Y, Han X, Deveci M, Parmar M (2024) A review of convolutional neural networks in computer vision. Artif Intell Rev 57(4):99. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10462-024-10721-6\u003c/span\u003e\u003cspan address=\"10.1007/s10462-024-10721-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen G, Shi Y, Wang R, Ren C, Ng E, Fang X, Ren Z (2022) Integrating weather observations and local-climate-zone-based landscape patterns for regional hourly air temperature mapping using machine learning. Sci Total Environ 841:156737. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.scitotenv.2022.156737\u003c/span\u003e\u003cspan address=\"10.1016/j.scitotenv.2022.156737\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003edos Santos RS (2020) Estimating spatio-temporal air temperature in London (UK) using machine learning and earth observation satellite data. Int J Appl Earth Obs Geoinf 88:102066. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jag.2020.102066\u003c/span\u003e\u003cspan address=\"10.1016/j.jag.2020.102066\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHobeichi S, Nishant N, Shao Y, Abramowitz G, Pitman A, Sherwood S, Bishop C, Green S (2023) Using machine learning to cut the cost of dynamical downscaling. Earths Future 11(3). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1029/2022EF003291\u003c/span\u003e\u003cspan address=\"10.1029/2022EF003291\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. p. e2022EF003291\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNishant N, Hobeichi S, Sherwood S, Abramowitz G, Shao Y, Bishop C, Pitman A (2023) Comparison of a novel machine learning approach with dynamical downscaling for Australian precipitation. Environ Res Lett 18(9):094006. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1088/1748-9326/ace463\u003c/span\u003e\u003cspan address=\"10.1088/1748-9326/ace463\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRampal N et al (2024) Enhancing regional climate downscaling through advances in machine learning. Artif Intell Earth Syst 3(2):230066. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1175/AIES-D-23-0066.1\u003c/span\u003e\u003cspan address=\"10.1175/AIES-D-23-0066.1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePotgieter J, Nazarian N, Lipson MJ, Hart MA, Ulpiani G, Morrison W, Benjamin K (2021) Combining high-resolution land use data with crowdsourced air temperature to investigate intra-urban microclimate. Front Environ Sci 9:720323. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fenvs.2021.720323\u003c/span\u003e\u003cspan address=\"10.3389/fenvs.2021.720323\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen F et al (2011) The integrated WRF/urban modelling system: development, evaluation, and applications to urban environmental problems. Int J Climatol 31(2):273\u0026ndash;288. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/joc.2158\u003c/span\u003e\u003cspan address=\"10.1002/joc.2158\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSalamanca F, Krpo A, Martilli A, Clappier A (2010) A new building energy model coupled with an urban canopy parameterization for urban climate simulations\u0026mdash;part I. formulation, verification, and sensitivity analysis of the model. Theoret Appl Climatol 99(3):331\u0026ndash;344. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00704-009-0142-9\u003c/span\u003e\u003cspan address=\"10.1007/s00704-009-0142-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYoo C, Im J, Park S, Quackenbush LJ (2018) Estimation of daily maximum and minimum air temperatures in urban landscapes using MODIS time series satellite data. ISPRS J photogrammetry remote Sens 137:149\u0026ndash;162. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.isprsjprs.2018.01.018\u003c/span\u003e\u003cspan address=\"10.1016/j.isprsjprs.2018.01.018\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang W, Br\u0026ouml;nnimann S, Zhou J, Li S, Wang Z (2025) Near-surface air temperature estimation for areas with sparse observations based on transfer learning. ISPRS J photogrammetry remote Sens 220:712\u0026ndash;727. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.isprsjprs.2025.01.021\u003c/span\u003e\u003cspan address=\"10.1016/j.isprsjprs.2025.01.021\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSalih W et al (2025) A High-Resolution Air Temperature Estimation Using Machine Learning and Land Surface Data Across Morocco. Earth Syst Environ 9(2):1311\u0026ndash;1331. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s41748-025-00629-8\u003c/span\u003e\u003cspan address=\"10.1007/s41748-025-00629-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVancutsem C, Ceccato P, Dinku T, Connor SJ (2010) Evaluation of MODIS land surface temperature data to estimate air temperature in different ecosystems over Africa. Remote Sens Environ 114(2):449\u0026ndash;465. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.rse.2009.10.002\u003c/span\u003e\u003cspan address=\"10.1016/j.rse.2009.10.002\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMacdonald R, Griffiths R, Hall D (1998) An improved method for the estimation of surface roughness of obstacle arrays. Atmos Environ 32(11):1857\u0026ndash;1864. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/S1352-2310(97)00403-2\u003c/span\u003e\u003cspan address=\"10.1016/S1352-2310(97)00403-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKanda M, Inagaki A, Miyamoto T, Gryschka M, Raasch S (2013) A new aerodynamic parametrization for real urban surfaces. Boundary Layer Meteorol 148(2):357\u0026ndash;377. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10546-013-9818-x\u003c/span\u003e\u003cspan address=\"10.1007/s10546-013-9818-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDing X, Zhao Y, Fan Y, Li Y, Ge J (2023) Machine learning-assisted mapping of city-scale air temperature: Using sparse meteorological data for urban climate modeling and adaptation. Build Environ 234:110211. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.buildenv.2023.110211\u003c/span\u003e\u003cspan address=\"10.1016/j.buildenv.2023.110211\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"UNSW Sydney","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Air temperature mapping, Land surface temperature, Crowdsourcing, Remote sensing, Machine learning, Urban heat","lastPublishedDoi":"10.21203/rs.3.rs-7987756/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7987756/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eUrban heat is a growing concern for public health, energy demand, and urban liveability. High-resolution air temperature (T\u003csub\u003ea\u003c/sub\u003e) data is needed for developing effective, localised adaptation strategies, but obtaining such data at the city scale remains a challenge as weather stations are often sparse and unevenly distributed within cities. To address this data gap, we developed a machine learning (ML) framework that creates high-resolution, gridded T\u003csub\u003ea\u003c/sub\u003e maps using crowdsourced observations and diverse geospatial datasets describing the city. This framework uses satellite-derived Land Surface Temperature (LST), urban form and fabric datasets, and meteorological variables to train a Convolutional Neural Network (CNN) algorithm. The approach was implemented in Sydney, Australia, using multi-day observations from 2019 to 2024, producing 30 m gridded T\u003csub\u003ea\u003c/sub\u003e estimates with high accuracy (R\u0026sup2; = 0.97, RMSE\u0026thinsp;=\u0026thinsp;0.91\u0026deg;C, surpassing previously reported ML performances). We further assessed the generalisation and spatial transferability of this method, which revealed that the CNN model maintained strong predictive accuracy for unseen locations across the city (R\u0026sup2; = 0.91\u0026ndash;0.93; RMSE\u0026thinsp;=\u0026thinsp;1.27\u0026ndash;1.44\u0026deg;C). Model performance also remained stable when the number of T\u003csub\u003ea\u003c/sub\u003e stations was significantly reduced by ~\u0026thinsp;80%. Performance slightly declined for unseen days (R\u0026sup2; = 0.66\u0026ndash;0.93; RMSE\u0026thinsp;=\u0026thinsp;1.52\u0026ndash;2.60\u0026deg;C), suggesting the need for incorporating a broader range of weather conditions. We find that high-resolution city-descriptive datasets are beneficial but not essential, as comparable accuracy was achieved using only globally available predictors. These results indicate that the proposed framework is transferable to other cities, including those in data-sparse regions. The study provides an effective and scalable approach for developing city-scale air temperature maps, which are urgently needed for urban heat assessment, climate adaptation, and public health planning.\u003c/p\u003e","manuscriptTitle":"Developing gridded air temperature data over cities using machine learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-31 04:19:32","doi":"10.21203/rs.3.rs-7987756/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"fd42a03a-4489-44ba-8598-6054c1d8237a","owner":[],"postedDate":"October 31st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-10-31T04:19:32+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-31 04:19:32","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7987756","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7987756","identity":"rs-7987756","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}