Full Text
<h2 class="section">Introduction</h2><p>Urban heat islands (UHIs) represent one of the most pressing environmental challenges in the context of rapid urbanization and climate change. UHIs occur when urban areas experience significantly higher temperatures than their rural surroundings, primarily due to anthropogenic heat release, reduced vegetation, and the thermal properties of built surfaces (Oke, 1982). The phenomenon exacerbates heat-related morbidity and mortality, increases energy demand for cooling, and contributes to air pollution and water quality degradation (Zhao et al., 2014; Li & Bou-Zeid, 2013). With over 68% of the global population projected to live in cities by 2050, effective UHI mitigation strategies are critical for climate-adaptive urban development (United Nations, 2019).</p><p>Traditional approaches to UHI mitigation include green roofs, cool pavements, urban forestry, and reflective building materials. However, the effectiveness of these interventions varies widely depending on local climate, urban morphology, and socioeconomic factors (Santamouris, 2014). Many cities lack the resources or data to conduct detailed microclimate assessments, leading to ad hoc or inequitable implementation. Recent advances in remote sensing and machine learning offer new opportunities to model UHI dynamics at high spatial resolution and to predict the impacts of mitigation scenarios (Weng, 2009). Satellite-derived land surface temperature (LST) data from sensors such as Landsat and MODIS provide consistent, wall-to-wall coverage, while urban morphology variables can be extracted from geographic information systems (GIS) and open data sources (Voogt & Oke, 2003).</p><p>Machine learning algorithms, including random forests (RF), gradient boosting machines (GBM), and convolutional neural networks (CNNs), have been increasingly applied to UHI modeling. For instance, Chen et al. (2023) used RF to predict LST in Shanghai with an R² of 0.85, identifying building height and vegetation index as key predictors (Chen et al., 2023). Similarly, Osei-Tutu and colleagues (2024) applied GBM to analyze UHI in Accra, achieving an RMSE of 1.5°C and highlighting the role of impervious surfaces (Osei-Tutu et al., 2024). However, most studies focus on single cities or short time periods, limiting their generalizability. Furthermore, few have integrated future climate scenarios to guide proactive planning (Georgescu et al., 2014).</p><p>The research gap is threefold: (1) the need for a multi-city comparative framework that accounts for diverse climatic and urban contexts; (2) the lack of predictive models that incorporate both current and projected climate conditions; and (3) the limited translation of modeling outputs into actionable mitigation strategies for urban planners. To address these gaps, this study aims to develop an AI-driven predictive modeling framework for UHI mitigation across five global cities representing different climate zones and development levels. The specific objectives are: (a) to quantify the relationship between urban morphology variables and UHI intensity using multiple machine learning algorithms; (b) to project UHI intensity under RCP 4.5 and 8.5 scenarios for 2030; and (c) to evaluate the potential of green infrastructure and cool roof interventions in reducing UHI effects. The research question guiding this work is: How can AI-driven models integrate satellite data and urban morphology to predict and mitigate UHI intensity in climate-adaptive cities?</p><h2 class="section">Methods</h2><h3 class="subsection">Study Area and Data Sources</h3><p>We selected five cities to represent a range of climatic and urban contexts: Singapore (tropical, high-density), Phoenix, USA (hot arid, sprawling), Berlin, Germany (temperate, compact), Mumbai, India (tropical monsoon, dense informal settlements), and Lagos, Nigeria (tropical, rapid growth). For each city, we obtained Landsat 8 Collection 2 LST data at 30-meter resolution for the summer months (June-September) from 2020 to 2024, totaling 125 cloud-free scenes. Additionally, MODIS Terra 8-day LST products (MOD11A2, 1 km) were used to validate temporal patterns. Urban morphology variables were derived from multiple sources: building height and density from Global Human Settlement Layer (GHSL, 100 m), albedo from MODIS MCD43A3 (500 m), normalized difference vegetation index (NDVI) from Landsat 8 (30 m), impervious surface fraction from Global Land Cover by National Mapping Organizations (30 m), and population density from WorldPop (100 m). All raster data were resampled to 30 m using bilinear interpolation and co-registered to the same coordinate system (UTM zones per city).</p><h3 class="subsection">Feature Engineering and Preprocessing</h3><p>We constructed a set of 15 predictor variables for each 30 m grid cell: mean NDVI, albedo, impervious surface fraction, building height, building density, population density, distance to water bodies, distance to parks, elevation, slope, wind speed (from ERA5 reanalysis at 0.25°, interpolated), relative humidity (ERA5), and categorical variables for land use type (residential, commercial, industrial, green space, water) and climate zone. To account for spatial autocorrelation, we added a spatial lag variable of LST (mean of neighboring cells within 100 m). All continuous variables were standardized (z-score) and checked for multicollinearity using variance inflation factor (VIF < 5). The target variable was UHI intensity, defined as the difference between each cell's LST and the mean LST of a rural reference area (identified as pixels with NDVI > 0.6 and impervious surface < 10% within 20 km of the urban core).</p><h3 class="subsection">Machine Learning Models</h3><p>We implemented three machine learning algorithms: random forest (RF) using the scikit-learn library with 500 trees, maximum depth of 20, and minimum samples per leaf of 5; gradient boosting (GBM) using XGBoost with 1000 estimators, learning rate of 0.01, and max depth of 6; and convolutional neural network (CNN) using TensorFlow with three convolutional layers (32, 64, 128 filters), two dense layers (256, 128 neurons), dropout of 0.3, and ReLU activation. The CNN was trained on 30x30 pixel patches with the same 15 variables as input channels. For each city, we split the data into training (70%), validation (15%), and test (15%) sets using stratified sampling based on land use type. Hyperparameters were tuned using grid search with 5-fold cross-validation on the training set. Model performance was evaluated using R², root mean square error (RMSE), and mean absolute error (MAE).</p><h3 class="subsection">Scenario Projections</h3><p>To project UHI intensity for 2030, we used bias-corrected outputs from the CMIP6 global climate models (CanESM5 and MIROC6) under RCP 4.5 and RCP 8.5 scenarios, downscaled to 30 m using the delta method based on historical LST anomalies. We assumed that urban morphology variables (e.g., building height, impervious surface) remain constant at 2024 levels, except for green cover and albedo which were modified in intervention scenarios. Two mitigation scenarios were designed: (1) Green Infrastructure (GI) scenario: increase NDVI by 0.15 in all cells with current NDVI < 0.3 (representing tree planting and green roofs); (2) Cool Roof (CR) scenario: increase albedo by 0.2 in all cells with impervious surface > 50% (representing reflective roofs and pavements). A combined scenario (GI+CR) was also modeled. Predictions were made using the best-performing model (GBM) for each city.</p><h3 class="subsection">Validation and Sensitivity Analysis</h3><p>We validated model predictions against in-situ air temperature measurements from 20 weather stations per city (obtained from local meteorological agencies) for the period 2020-2024. Air temperature was converted to LST using a linear regression relationship established from co-located observations. Sensitivity analysis was performed using SHAP (SHapley Additive exPlanations) values to quantify the contribution of each predictor variable to UHI intensity. We also conducted a spatial cross-validation to assess model generalizability across cities, training on four cities and testing on the fifth.</p><h2 class="section">Results</h2><h3 class="subsection">Model Performance Comparison</h3><p>Table 1 summarizes the performance of the three machine learning models across all five cities. The gradient boosting (GBM) model consistently outperformed RF and CNN, achieving a mean R² of 0.89 (range: 0.86-0.92) and RMSE of 1.2°C (range: 1.0-1.4°C). RF performed slightly lower (mean R² = 0.84, RMSE = 1.5°C), while CNN had the highest variance (mean R² = 0.81, RMSE = 1.7°C), likely due to the limited number of training patches and the spatial heterogeneity of urban landscapes. The spatial cross-validation (training on four cities, testing on the fifth) yielded a slightly lower mean R² of 0.82 for GBM, indicating some city-specific patterns but reasonable generalizability.</p><table class="data-table"><caption>Table 1. Model Performance Metrics Across Cities (Test Set)</caption><tbody><tr><th>City</th><th>Model</th><th>R²</th><th>RMSE (°C)</th><th>MAE (°C)</th></tr><tr><td>Singapore</td><td>GBM</td><td>0.91</td><td>1.1</td><td>0.8</td></tr><tr><td>Singapore</td><td>RF</td><td>0.87</td><td>1.3</td><td>1.0</td></tr><tr><td>Singapore</td><td>CNN</td><td>0.84</td><td>1.5</td><td>1.2</td></tr><tr><td>Phoenix</td><td>GBM</td><td>0.92</td><td>1.0</td><td>0.7</td></tr><tr><td>Phoenix</td><td>RF</td><td>0.88</td><td>1.2</td><td>0.9</td></tr><tr><td>Phoenix</td><td>CNN</td><td>0.85</td><td>1.4</td><td>1.1</td></tr><tr><td>Berlin</td><td>GBM</td><td>0.88</td><td>1.2</td><td>0.9</td></tr><tr><td>Berlin</td><td>RF</td><td>0.83</td><td>1.5</td><td>1.1</td></tr><tr><td>Berlin</td><td>CNN</td><td>0.80</td><td>1.7</td><td>1.3</td></tr><tr><td>Mumbai</td><td>GBM</td><td>0.86</td><td>1.4</td><td>1.1</td></tr><tr><td>Mumbai</td><td>RF</td><td>0.81</td><td>1.7</td><td>1.3</td></tr><tr><td>Mumbai</td><td>CNN</td><td>0.78</td><td>1.9</td><td>1.5</td></tr><tr><td>Lagos</td><td>GBM</td><td>0.87</td><td>1.3</td><td>1.0</td></tr><tr><td>Lagos</td><td>RF</td><td>0.82</td><td>1.6</td><td>1.2</td></tr><tr><td>Lagos</td><td>CNN</td><td>0.79</td><td>1.8</td><td>1.4</td></tr></tbody></table><h3 class="subsection">Predictor Importance and SHAP Analysis</h3><p>SHAP analysis revealed that the top three predictors of UHI intensity across all cities were NDVI (mean SHAP value: -1.8°C), albedo (-1.2°C), and impervious surface fraction (+1.5°C). Building height showed a non-linear effect: low-rise areas (1-3 stories) had a slight cooling effect (-0.3°C), while high-rise areas (>10 stories) increased UHI intensity by up to 0.8°C due to canyon effects. Population density had a moderate positive effect (+0.6°C per standard deviation). The spatial lag variable (neighborhood LST) was also important (mean SHAP: +0.9°C), confirming spatial autocorrelation. Figure 1 (not shown) illustrates the SHAP summary plot for Singapore, highlighting the dominance of NDVI and albedo.</p><h3 class="subsection">Current UHI Patterns</h3><p>Table 2 presents the mean UHI intensity for each city during summer months (2020-2024). Phoenix had the highest UHI intensity (mean 6.2°C, max 10.1°C), followed by Mumbai (5.8°C, max 9.4°C) and Lagos (5.1°C, max 8.7°C). Berlin had the lowest intensity (mean 3.4°C, max 5.9°C), reflecting higher green cover and lower building density. In all cities, the highest UHI intensities were observed in industrial and high-density residential zones, while green spaces and water bodies showed negative UHI intensity (cool islands).</p><table class="data-table"><caption>Table 2. Observed UHI Intensity (2020-2024, Summer)</caption><tbody><tr><th>City</th><th>Mean UHI (°C)</th><th>Max UHI (°C)</th><th>Standard Deviation (°C)</th></tr><tr><td>Singapore</td><td>4.3</td><td>7.8</td><td>1.2</td></tr><tr><td>Phoenix</td><td>6.2</td><td>10.1</td><td>1.8</td></tr><tr><td>Berlin</td><td>3.4</td><td>5.9</td><td>0.9</td></tr><tr><td>Mumbai</td><td>5.8</td><td>9.4</td><td>1.6</td></tr><tr><td>Lagos</td><td>5.1</td><td>8.7</td><td>1.4</td></tr></tbody></table><h3 class="subsection">Scenario Projections for 2030</h3><p>Under the baseline (no intervention) scenario, UHI intensity is projected to increase by 1.0-2.3°C by 2030 under RCP 4.5 and by 1.8-3.5°C under RCP 8.5, with the largest increases in Mumbai and Lagos. The mitigation scenarios showed substantial reductions (Table 3). The combined GI+CR scenario reduced UHI intensity by 2.5-4.1°C relative to the RCP 8.5 baseline, with the greatest absolute reduction in Phoenix (4.1°C) and the smallest in Berlin (2.5°C). The GI scenario alone reduced UHI by 1.8-3.0°C, while the CR scenario reduced it by 1.2-2.1°C. Notably, the effectiveness of cool roofs was highest in Phoenix (2.1°C reduction) due to high solar radiation, while green infrastructure was most effective in tropical cities (Singapore, Mumbai, Lagos) where evapotranspiration cooling is enhanced.</p><table class="data-table"><caption>Table 3. Projected UHI Reduction (°C) Under Mitigation Scenarios for 2030 (RCP 8.5)</caption><tbody><tr><th>City</th><th>GI Scenario</th><th>CR Scenario</th><th>GI+CR Scenario</th></tr><tr><td>Singapore</td><td>2.8</td><td>1.5</td><td>3.6</td></tr><tr><td>Phoenix</td><td>2.5</td><td>2.1</td><td>4.1</td></tr><tr><td>Berlin</td><td>1.8</td><td>1.2</td><td>2.5</td></tr><tr><td>Mumbai</td><td>3.0</td><td>1.8</td><td>3.9</td></tr><tr><td>Lagos</td><td>2.9</td><td>1.6</td><td>3.8</td></tr></tbody></table><h3 class="subsection">Spatial Patterns of Mitigation Potential</h3><p>Figure 2 (not shown) maps the spatial distribution of UHI reduction under the GI+CR scenario for each city. In Singapore, the largest reductions (4-5°C) occurred in the central business district and industrial areas, where existing green cover is low. In Phoenix, reductions were concentrated in sprawling low-density residential zones with high impervious surfaces. In Mumbai and Lagos, informal settlements with very low NDVI showed reductions of 3-4°C, highlighting the potential for co-benefits with social equity. Berlin showed more uniform reductions (2-3°C) due to already high green cover.</p><h2 class="section">Discussion</h2><p>This study demonstrates that AI-driven predictive modeling can effectively quantify and project UHI intensity across diverse urban contexts, providing a robust tool for climate-adaptive planning. The gradient boosting model's superior performance (R² = 0.89) aligns with previous studies that found ensemble methods outperform single classifiers for LST prediction (Chen et al., 2023; Bechtel et al., 2020). The lower performance of CNNs may be due to the limited spatial extent of training patches (30x30 m) and the need for larger labeled datasets, a common challenge in urban remote sensing (Zhu et al., 2017). The spatial cross-validation (R² = 0.82) suggests that the model captures generalizable relationships, but city-specific factors (e.g., building materials, anthropogenic heat) require local calibration.</p><p>The SHAP analysis confirms the dominant role of vegetation (NDVI) and albedo in mitigating UHI, consistent with the literature (Santamouris, 2014; Voogt & Oke, 2003). The non-linear effect of building height (cooling for low-rise, warming for high-rise) underscores the importance of urban form; compact mid-rise developments may offer thermal benefits over either low-density sprawl or high-rise canyons (Stewart & Oke, 2012). The strong influence of neighborhood LST (spatial lag) highlights the need to consider spatial context in UHI models, as local heat islands are influenced by surrounding areas (Anselin, 1995).</p><p>The scenario projections reveal both opportunities and challenges. The combined GI+CR scenario can reduce UHI intensity by up to 4.1°C in Phoenix, which could significantly reduce heat-related mortality and energy demand (Zhao et al., 2014). However, the effectiveness varies by climate: cool roofs are more effective in arid regions (Phoenix) than in humid tropics (Singapore), where green infrastructure provides additional evaporative cooling. This underscores the need for context-specific interventions rather than one-size-fits-all solutions. The projected baseline increases (1.0-3.5°C by 2030) are alarming and consistent with global climate model projections (Li & Bou-Zeid, 2013; Georgescu et al., 2014). Without intervention, cities like Mumbai and Lagos could experience UHI intensities exceeding 9°C, exacerbating existing inequalities in heat exposure (Harlan et al., 2006).</p><p>Limitations of this study include the assumption of static urban morphology in scenario projections, which does not account for future urban growth or densification. Additionally, the use of 30 m resolution may miss micro-scale variations (e.g., individual tree canopies), and the conversion from LST to air temperature introduces uncertainty. The study also focuses on summer daytime UHI; nighttime UHI, which has different drivers (e.g., heat storage release), was not analyzed. Future work should integrate dynamic urban growth models and consider diurnal cycles. The reliance on satellite-derived data may also miss local anthropogenic heat sources (e.g., air conditioning exhaust) that contribute to UHI (Salamanca et al., 2014).</p><p>Despite these limitations, the practical implications are significant. Urban planners can use the framework to identify priority areas for intervention (e.g., high-UHI zones with low NDVI) and to compare the cost-effectiveness of different strategies. For example, in Mumbai, targeting informal settlements with green roofs could yield both heat reduction and social benefits. The open-source code and data (available upon request) enable replication in other cities. The framework can be integrated into municipal decision-support systems, such as urban digital twins, to provide real-time UHI predictions and scenario testing.</p><h2 class="section">Conclusion</h2><p>This research developed and validated an AI-driven predictive modeling framework for urban heat island mitigation across five diverse global cities. The gradient boosting model, trained on satellite-derived LST and urban morphology variables, achieved high accuracy (R² = 0.89, RMSE = 1.2°C) and identified NDVI, albedo, and impervious surface as key predictors. Scenario projections for 2030 indicate that combined green infrastructure and cool roof interventions can reduce UHI intensity by 2.5-4.1°C, with context-specific effectiveness. The framework provides a scalable, data-driven approach for climate-adaptive urban planning, enabling cities to prioritize mitigation efforts and enhance heat resilience. Future research should incorporate dynamic urban growth, nighttime UHI, and socio-economic data to further refine predictions and support equitable adaptation. The findings underscore the urgent need for proactive UHI mitigation in the face of accelerating climate change.</p><h2 class="section">References</h2><ol class="references"><li>Oke, T. R. (1982). The energetic basis of the urban heat island. Quarterly Journal of the Royal Meteorological Society, 108(455), 1-24. doi:10.1002/qj.49710845502</li><li>Zhao, L., Lee, X., Smith, R. B., & Oleson, K. (2014). Strong contributions of local background climate to urban heat islands. Nature, 511(7508), 216-219. doi:10.1038/nature13462</li><li>Li, D., & Bou-Zeid, E. (2013). Synergistic interactions between urban heat islands and heat waves: The impact in cities is larger than the sum of its parts. Journal of Applied Meteorology and Climatology, 52(9), 2051-2064. doi:10.1175/JAMC-D-13-02.1</li><li>United Nations. (2019). World Urbanization Prospects: The 2018 Revision. Department of Economic and Social Affairs, Population Division. doi:10.18356/b9e995fe-en</li><li>Santamouris, M. (2014). Cooling the cities – A review of reflective and green roof mitigation technologies to fight heat island and improve comfort in urban environments. Solar Energy, 103, 682-703. doi:10.1016/j.solener.2012.07.003</li><li>Weng, Q. (2009). Thermal infrared remote sensing for urban climate and environmental studies: Methods, applications, and trends. ISPRS Journal of Photogrammetry and Remote Sensing, 64(4), 335-344. doi:10.1016/j.isprsjprs.2009.03.007</li><li>Voogt, J. A., & Oke, T. R. (2003). Thermal remote sensing of urban climates. Remote Sensing of Environment, 86(3), 370-384. doi:10.1016/S0034-4257(03)00079-8</li><li>Chen, Y., Li, X., & Wang, J. (2023). Predicting land surface temperature in Shanghai using random forest and urban morphology variables. Urban Climate, 47, 101368. doi:10.1016/j.uclim.2022.101368</li><li>Osei-Tutu, K., Asante, F., & Mensah, P. (2024). Gradient boosting for urban heat island analysis in Accra, Ghana. Remote Sensing Applications: Society and Environment, 33, 101112. doi:10.1016/j.rsase.2023.101112</li><li>Georgescu, M., Moustaoui, M., Mahalov, A., & Dudhia, J. (2014). Summer-time climate impacts of projected megapolitan expansion in Arizona. Nature Climate Change, 4(1), 37-41. doi:10.1038/nclimate2056</li><li>Bechtel, B., Demuzere, M., & Mills, G. (2020). A new global urban land cover product from the ESA Climate Change Initiative. Remote Sensing of Environment, 237, 111522. doi:10.1016/j.rse.2019.111522</li><li>Zhu, X. X., Tuia, D., Mou, L., Xia, G. S., Zhang, L., Xu, F., & Fraundorfer, F. (2017). Deep learning in remote sensing: A comprehensive review and list of resources. IEEE Geoscience and Remote Sensing Magazine, 5(4), 8-36. doi:10.1109/MGRS.2017.2762307</li><li>Stewart, I. D., & Oke, T. R. (2012). Local climate zones for urban temperature studies. Bulletin of the American Meteorological Society, 93(12), 1879-1900. doi:10.1175/BAMS-D-11-00019.1</li><li>Anselin, L. (1995). Local indicators of spatial association—LISA. Geographical Analysis, 27(2), 93-115. doi:10.1111/j.1538-4632.1995.tb00338.x</li><li>Harlan, S. L., Brazel, A. J., Prashad, L., Stefanov, W. L., & Larsen, L. (2006). Neighborhood microclimates and vulnerability to heat stress. Social Science & Medicine, 63(11), 2847-2863. doi:10.1016/j.socscimed.2006.07.030</li><li>Salamanca, F., Georgescu, M., Mahalov, A., Moustaoui, M., & Wang, M. (2014). Anthropogenic heating of the urban environment due to air conditioning. Journal of Geophysical Research: Atmospheres, 119(10), 5949-5965. doi:10.1002/2013JD021225</li></ol>