Leveraging urban AI for high-resolution urban heat mapping: Towards climate resilient cities

Study area and data collection

This study focused on urban heat mapping within the metropolitan area of Adelaide, South Australia (Supplemental Figure S1). Adelaide’s urban landscape, comprising densely populated city centres, suburban neighbourhoods, and surrounding green spaces, offers a dynamic and varied environment for analysing urban heat distribution. This diversity provides an ideal case study for assessing how different urban forms and infrastructure impact surface temperatures, making Adelaide a valuable model for understanding UHI effects in cities with similarly mixed landscapes. The selection of Adelaide allows for a detailed analysis of how densely built-up areas, green spaces, and suburban layouts each contribute uniquely to localised temperature patterns.

Data sources and preparation

High-resolution thermal and spatial data were obtained from Data SA, the South Australian Government Data Directory (SA Dataset, 2024). This dataset provides advanced satellite imagery capable of capturing detailed surface temperature variations across both urban and non-urban areas. The thermal data were analysed at a spatial resolution of 2m, where each pixel value corresponds to land surface temperature (LST). This fine spatial resolution enables the precise identification of localised thermal variations, a key requirement for understanding UHI patterns and their relationship with urban morphology and vegetation. High-resolution imagery was specifically selected to capture fine-scale variations in surface heat intensity, thereby enabling a precise examination of temperature distribution. This level of detail is essential for identifying UHI patterns, understanding how temperature gradients shift across different urban layouts, and supporting strategies for urban heat mitigation.

The dataset used in this study contained spatial and thermal information at a resolution of 1024 × 1024 pixels per image, offering a balanced combination of spatial granularity and computational feasibility, which is particularly suitable for deep learning applications. The dataset captures a single temporal snapshot during summer, with recorded temperatures ranging from a maximum of 33°C to a minimum of 16°C. While these conditions provide critical insights into peak urban heat scenarios, the static nature of the dataset limits its ability to account for seasonal variability and dynamic changes in urban heat patterns.

The dataset includes critical spatial and thermal details that allow the model to learn the impacts of urban morphology and vegetation on UHI patterns. High-density urban areas with impervious surfaces, such as roads and rooftops, retain more heat, resulting in localised hotspots. Conversely, green spaces, including parks and vegetation, mitigate urban heat through natural cooling processes such as evapotranspiration. The fine spatial resolution of 2 m enables the model to detect subtle thermal gradients and distinguish between densely built and vegetated areas. By leveraging these capabilities, the U-Net model captures complex interactions between urban structure, vegetation, and thermal dynamics, delivering actionable insights for targeted UHI mitigation strategies.

This study focused on high-resolution thermal mapping as the primary approach for urban heat assessment. Although variables such as humidity and wind influence urban heat dynamics, they were not included in this model due to constraints in data availability, computational efficiency, and the study’s focus on land surface temperature mapping.

Data normalisation

As a preparatory step, the values of each pixel in the satellite imagery were normalised to a continuous range [0, 1]. Normalising pixel values is particularly beneficial in deep learning, as it helps prevent issues such as vanishing or exploding gradients, thereby promoting stable training and faster convergence. This step is essential for achieving consistent results and improving the overall effectiveness of the model training process.

Ground truth heat maps

Urban heat maps representing ground-truth heat intensity were converted into grayscale for single-channel processing, where pixel intensity represents relative heat levels across the urban landscape. This simplified format focuses the model’s learning on heat intensity distribution without colour channel complexities. Following prediction, the grayscale output was colour-mapped through a predefined lookup table (LUT) for enhanced interpretability. This colourisation aligns heat intensities with visually distinct colour bands, which is crucial for visual assessments.

Training-validation split

The dataset was split into training (80%) and validation (20%) subsets to evaluate the model’s generalisation capability. This ratio was selected to maximise the training set size, ensuring the model could learn robust spatial and thermal patterns while retaining sufficient validation data to assess performance on unseen samples.

To ensure the model generalised effectively, validation error was used as the primary indicator. The objective was to minimise training errors without compromising validation performance. If validation error began increasing while training error continued decreasing, it indicated overfitting, where the model memorised training data rather than learning meaningful patterns. To mitigate this, validation loss was continuously monitored, and early stopping was applied, halting training when validation performance began to deteriorate.

U-net model architecture

U-Net CNN architecture was employed owing to its strengths in pixel-wise image-segmentation tasks. Originally developed for biomedical image segmentation, U-Net’s encoder–decoder structure has proven effective in diverse applications where both fine and broad spatial details are essential to evaluate the output (Ronneberger et al., 2015). Its architecture allows the model to retain the spatial details of the input image, making it highly suitable for mapping tasks (Supplemental Figure S2). The U-Net model facilitates the detection of urban heat patterns by learning both local and global features, effectively balancing spatial accuracy with computational efficiency.

Model training and optimisation

The U-Net model was trained using a GPU-based environment, which is necessary for handling large-scale data and complex architecture efficiently. Model training was guided by a range of hyperparameters optimised to ensure effective convergence and robust performance on unseen data.

Model performance and training metrics

The U-Net model was trained on high-resolution satellite imagery of urban areas, utilising an 85/15 split between training and validation data. The model’s training process captured both fine-scale and broad thermal patterns within urban heat distribution. Over 400 epochs, the model’s training performance was monitored using MSE as the primary loss function (Supplemental Figure S3). The training loss consistently decreased from an initial MSE of approximately 0.0013 to a final value of 0.0007, indicating effective convergence. The validation loss followed a similar decreasing trend, confirming the ability of the model to generalise to unseen data.

The model was optimised using the Adam algorithm with an initial learning rate of 1e−4. A learning rate scheduler dynamically adjusted the learning rate by a factor of 0.1 when the validation loss plateaued for eight consecutive epochs, refining the convergence process. A batch size of 2 was selected to balance memory requirements for processing 1024 × 1024 resolution images while capturing urban thermal patterns effectively. This combination of extended training, dynamic learning rate adjustments, and optimal batch size supported the model in learning complex spatial dependencies, ensuring robust performance across various urban environments.

Visual analysis of predicted heat maps

A comparison of satellite images, ground truth heat maps, and U-Net model’s predicted heat maps for four distinct urban areas (A, B, C, and D) is presented in (Supplemental Figure S4). Starting with Area A, the model successfully identified high-temperature zones within densely built-up regions, while cooler temperatures were accurately represented in nearby green spaces. This differentiation between urban infrastructure and vegetated areas closely aligned with the ground truth, highlighting the model’s sensitivity to variations in land cover.

In Area B, which features an open green space surrounded by residential buildings, the model predictions showed lower temperatures within the vegetative area and higher temperatures in the adjacent built environment. This result reflects the model’s capability to capture thermal gradients influenced by both natural cooling effects and nearby urban structures. In densely populated Area C, elevated temperatures appeared primarily along major roads and rooftops, where surface materials are known to absorb and retain heat. The predicted map closely mirrored the ground truth, demonstrating the efficacy of the model in identifying the typical thermal patterns associated with high-density urban zones.

Finally, Area D, with its complex layout of curved streets and varied building densities, further illustrates the model’s ability to generalise across intricate urban environments. Temperature gradients were accurately represented, with higher temperatures in dense building clusters and cooler values in vegetated zones. This consistency across diverse urban settings underscores the robustness and adaptability of the model in predicting thermal patterns across structured and complex landscapes.

Although the model performed consistently well, it exhibited minor inaccuracies in specific urban environments, particularly in areas with high variability in building density, surface materials, and environmental context. In these heterogeneous settings, the model showed slight deviations from the ground-truth heat maps, particularly in regions where mixed land cover introduced complex thermal patterns.

In densely built urban centres with minimal vegetation, the model sometimes underestimated the intensity of UHIs. In contrast, in highly vegetated zones, the model occasionally predicted marginally higher temperatures than observed. These discrepancies were most prominent in urban areas with a combination of concrete and vegetative surfaces, where thermal variation was complex.

Quantitative evaluation metrics

The final model performance was assessed using MSE, metrics well-suited for continuous temperature prediction. On the validation set, the model achieved an MSE of 0.00383, reflecting its ability to produce pixel-level predictions that closely matched the ground-truth heat intensities. MSE provided insight into the model’s effectiveness in minimising larger errors.

These metrics were selected to assess the accuracy of model predictions in capturing real urban heat distributions. Due to the continuous nature of the output, segmentation metrics such as Intersection over union (IoU) and pixel accuracy were not applicable, as they are intended for categorical segmentation tasks. Nonetheless, MSE directly assessed the capacity of the model to predict fine-scale temperature gradients, affirming its suitability for urban heat mapping applications. In addition to its predictive accuracy, the model demonstrated efficient processing, with each image processed in under 30 seconds. This rapid processing time is advantageous for large-scale urban heat mapping.

To further evaluate the visual and structural quality of the generated heat maps, we also calculated Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). A PSNR of approximately 18.7 dB indicates that the predicted heat maps are visually consistent with the ground-truth references, offering sufficient detail for spatial interpretation. Additionally, the SSIM value of up to 0.378 demonstrates the model’s ability to preserve spatial thermal patterns across different urban morphologies. Together, these complementary metrics confirm the model’s robustness in both numerical prediction and visual fidelity – attributes that are critical for practical urban analysis and planning applications.

Model performance and Implications

As an Urban AI solution, the U-Net model delivers high-resolution outputs that are particularly valuable for identifying microscale temperature variations in urban environments. This granular spatial detail enables urban planners to pinpoint heat-prone zones within dense infrastructure, facilitating targeted interventions such as green infrastructure expansion and the use of reflective materials, both of which are proven to reduce UHI intensity (Bowler et al., 2010; Gago et al., 2013; Shaamala et al., 2024). Furthermore, the model’s rapid processing capability aligns with operational requirements for extreme weather events, offering real-time decision-making support for mitigating heat risks (Chen et al., 2006; Johnson et al., 2012).

The U-Net model’s robust performance across varied urban morphologies highlights its adaptability. It accurately identified high-temperature zones in densely built areas, such as roads and rooftops, where impervious surfaces retain heat (Rizwan et al., 2008; Voogt and Oke, 2003). In suburban areas, it effectively captured lower temperatures associated with vegetation, reflecting greenery’s cooling effects (Bowler et al., 2010; Weng, 2009). However, discrepancies arose in mixed land cover areas, where complex interactions between concrete and vegetation create challenging thermal patterns. These findings emphasise the importance of integrating contextual environmental variables, such as shading and material reflectivity, to further enhance accuracy in heterogeneous urban environments (Gago et al., 2013; Middel et al., 2014).

Compared to traditional methods, the U-Net model exemplifies the transformative potential of Urban AI. Traditional techniques often rely on moderate-resolution satellite data (e.g. MODIS and Landsat) with extensive preprocessing, limiting their effectiveness for precise UHI analysis (Shi et al., 2021). In contrast, the U-Net’s encoder–decoder architecture, combined with skip connections, preserves critical pixel-level spatial features, enabling highly accurate identification of small-scale thermal gradients (Briegel et al., 2023; Ronneberger et al., 2015). This capability is indispensable for managing heterogeneous urban environments, where targeted interventions are vital for mitigating localised heat challenges (Diem et al., 2024; Elmarakby and Elkadi, 2024).

The U-Net model’s operational efficiency further cements its role as a foundational Urban AI agent. Its ability to rapidly assess thermal patterns supports timely decision-making during extreme heat events, addressing climate resilience needs effectively. Unlike traditional LST approaches that face significant delays (Tomlinson et al., 2011; Zhou et al., 2018), the U-Net model ensures prompt, data-driven urban planning and climate-adaptive design (Almashhour et al., 2024; Kumar et al., 2024).

Future advancements could include integrating additional environmental variables, such as humidity, wind patterns, and solar radiation, to better contextualise temperature patterns and enhance sensitivity to complex urban environments (Ghorbany et al., 2024; Liu and Morawska, 2020; Tehrani et al., 2024). Expanding the training data to encompass cities with diverse climates and urban morphologies would strengthen the model’s adaptability and robustness, ensuring its usability in varied geographic contexts. Further exploration of hybrid models combining U-Net with temporal data analysis, such as recurrent neural networks, could refine its predictive accuracy for continuous seasonal assessments (Demiss and Elsaigh, 2024; Nazir et al., 2024).

Evaluating the U-Net model across multiple cities would validate its scalability and establish a framework for Urban AI-enabled heat mapping. Such a framework could provide tailored insights into diverse thermal profiles, enabling planners to design interventions for location-specific UHI challenges (W. Zhou et al., 2011; Li et al., 2017). By aligning with global urban resilience goals, this Urban AI-driven approach fosters sustainable urban development, improves liveability, and enhances climate adaptability (Degirmenci et al., 2021; Ramyar et al., 2021).

Urban heat resilience and vulnerability in Adelaide

This study highlights significant variations in urban heat resilience across metropolitan Adelaide, shaped by land use, urban morphology, and vegetation cover. High-resolution mapping using the U-Net model reveals that densely built areas, including the CBD, industrial precincts, and transport corridors, record the highest land surface temperatures (LSTs) due to impervious surfaces and heat-retaining materials (Supplemental Figure S5).

Conversely, suburban neighbourhoods with green spaces exhibit moderate resilience, benefiting from tree cover and open spaces that help regulate temperature extremes. However, new high-density residential developments with limited vegetation contribute to localised warming, underscoring the need for climate-responsive urban design.

Coastal zones, parklands, and peri-urban areas demonstrate the greatest heat resilience, with lower LSTs due to the cooling effects of vegetation, water bodies, and open landscapes. These findings are in line with Morgan et al. (2024) and reinforce the crucial role of green and blue infrastructure in mitigating urban heat stress and enhancing urban climate resilience.

Practical Implications for urban planning

Urban planners and local governments currently rely on a combination of coarse-resolution thermal satellite imagery (e.g. MODIS or Landsat), empirical models, and in situ measurements to inform their urban heat mitigation strategies. These include tree canopy targets, zoning incentives for reflective materials, and emergency heatwave plans. However, existing approaches often lack the spatial granularity and real-time responsiveness required for street-level interventions.

The proposed U-Net model addresses these limitations by generating high-resolution, near real-time heat maps from widely available RGB imagery, requiring less than 30 seconds per image. This enables time-sensitive applications, such as identifying high-risk zones during extreme heat events and supporting proactive planning decisions. For instance, city councils could integrate these maps into GIS platforms to prioritise tree planting on heat-vulnerable streets, evaluate zoning compliance, or assess the thermal impact of new developments.

Moreover, prior studies have validated the effectiveness of RGB-based LST estimation, confirming that models trained on morphological and spatial patterns can reliably predict thermal behaviour across different urban contexts – even in the absence of thermal infrared data (Zhang et al., 2018; Zhao et al., 2023). This makes RGB-based approaches especially valuable for cities with limited access to thermal sensors or commercial satellite products.

The model also complements existing strategies by offering a scalable, low-cost solution that supports a continuous assessment of urban thermal dynamics. Nonetheless, its current reliance on land surface temperature restricts its ability to capture air temperature or thermal comfort directly. Therefore, it should be seen as a complementary decision-support tool rather than a replacement for physical sensor networks. Future integration with real-time sensing and urban analytics platforms could further enhance its value and planning impact.

Limitations and future research

While the U-Net model demonstrated strong overall performance, certain limitations emerged in areas with extreme thermal dynamics, presenting unique challenges. For example, in highly urbanised zones with minimal vegetation, the model occasionally underestimated peak temperatures. This underestimation likely stems from the limited representation of high-density UHI conditions in the training dataset. Incorporating additional data that captures extreme urban heat scenarios could improve the model’s sensitivity, ensuring more accurate predictions in such contexts. Expanding the dataset to include nuanced environmental variables would enhance the model’s capacity to perform reliably in diverse urban environments.

Another limitation lies in the model’s exclusive reliance on data from Adelaide, South Australia, which restricts its generalisability to regions with varying climates, urban layouts, and vegetation types. Although the model achieved commendable results within Adelaide, its application to other geographic areas may necessitate retraining or fine-tuning to account for unique environmental characteristics. Broadening the training dataset to include a wider range of urban forms, climates, and land-use types would significantly enhance the model’s robustness and adaptability, enabling it to deliver accurate results across diverse global landscapes.

Additionally, the static nature of the data used in this study constrains the model’s ability to account for seasonal variations in urban heat dynamics. Changes in solar radiation, atmospheric conditions, and vegetation states significantly influence surface temperature patterns, which are not captured in the current model. Incorporating seasonal variability into the dataset could provide a more dynamic perspective on urban thermal patterns, strengthening the model’s predictive power for year-round applications. Such enhancements would allow for more comprehensive analyses of heat distribution under varying environmental conditions.

Furthermore, while the model estimates land surface temperature, it does not capture air temperature or thermal comfort directly. This distinction is important, and future work could incorporate sensor-based measurements or predictive thermal comfort indices to enhance the relevance for human-centred heat risk assessment.

Addressing these limitations through data expansion, the integration of additional variables, temporal modelling, and computational optimisation would considerably improve the U-Net model’s adaptability, accuracy, and scalability. These advancements would enable the model to perform effectively across a broader range of urban settings, enhancing its role as an invaluable tool for sustainable urban planning and climate resilience. By refining its capabilities, the model could support more precise and actionable insights, empowering urban planners and policymakers to implement targeted interventions that mitigate UHI effects and promote climate-adaptive urban development.