Optimized spatial planning offers a dual solution for managing urban heat and air pollution in humid subtropical climates

Abstract

As cities expand, urban heat islands and air pollution have merged, especially in humid subtropical cities. Yet a dual solution remains lacking to simultaneously mitigate these two issues due to complex meteorology-chemistry interactions. Based on numerical simulations and a multi-objective optimization algorithm, we develop an optimized spatial planning that considers the intricate interplay between urban heat and pollution. Optimized development prioritizes urban locations upwind of the summer monsoon, and suggests developing medium-density urban (where the built-up fraction ranges from 0.5 to 0.8), as well as increasing urban green space fractions from 0.15 to 0.27. Such an urban layout yields the greatest reductions in urban warming and pollution. Both regression analysis and numerical simulations confirm the effectiveness of the framework. Enhanced ventilation improves air quality, while strong southerly winds mitigate urban heat. This study underscores the necessity of addressing urban heat and air pollution simultaneously.

Urban heat and air pollution often coexist as cities develop. This study proposes an urban development optimization framework that addresses both issues simultaneously by considering the complex interactions between the thermal environment and air pollution.

Introduction

Urbanization is an inevitable part of societal progress in China. During this process, natural vegetation is replaced with impervious surfaces, which decreases surface albedo and increases solar energy absorption¹. These changes lead to higher temperatures in urban areas that can exacerbate heat-related illness and affect local ecosystems^2–5. In parallel, human activities in urban areas contribute over 70% of greenhouse gas emissions⁶ and generate the majority of air pollutants^7,8, resulting in deteriorated air quality.

Urbanization intensifies temperature differences between urban and rural areas, resulting in the urban heat island (UHI) and degrading the urban thermal environment^9,10. Although clean air actions in China since 2013 have effectively reduced emissions¹¹, haze and ozone pollution continue to pose major challenges for urban air quality^12–14. As primary emissions have declined, secondary organic aerosol (SOA) has become increasingly important in particulate matter formation¹³. Many urban areas are Volatile Organic Compounds (VOC)-limited for ozone formation during warmer months, meaning that changes in VOCs concentrations strongly impact ozone levels¹⁵. Elevated temperatures over urban areas enhance biogenic and anthropogenic emissions of VOCs and accelerate chemical reaction rates, often leading to higher ozone and SOA^16–18. Consequently, compound heat and air pollution events are increasingly observed in cities^19–22, presenting substantial health risks and economic burdens^23,24.

To enhance urban resilience, cities have implemented various green (vegetation-based), blue (water-based), and grey (engineered) infrastructures (GBGI)^25–29. These measures have proven effective in mitigating urban warming^30–33 or improving air quality^34–36 to varying degrees. However, they are typically applied to existing urban areas and often require substantial financial investment. An alternative strategy is to optimize future urban expansion to achieve long-term balanced development^37,38. By refining urban land use locations^39–41, building density^42,43, location of green spaces^44–47, and other factors, urban planners can minimize adverse effects while maximizing social and environmental benefits.

Most studies on optimizing urban layout for sustainable development consider either urban heat^48–50 or air pollution^43,51 separately, neglecting their coupled processes. Recent work demonstrates two-way feedback between urban heat and air quality^52,53. Urban heat affects air quality through changes in atmospheric dynamics and chemistry. Warming strengthens atmospheric turbulence and reshapes urban–rural circulations, altering pollutant transport, mixing, and deposition, whereas cooler conditions stabilize the boundary layer and limit dispersion^54–58. Elevated temperatures also increase emissions of gaseous precursors and accelerate photochemical reactions, enhancing ozone formation and degrading air quality^17,59,60. Air pollutants in the boundary layer, especially aerosols, alter the surface energy balance through their radiative effects^61,62. By scattering and absorbing incoming solar radiation, particulate matter reduces shortwave radiation reaching the surface, lowering surface temperatures^63–66, yet absorption warms the lower atmospheres^67–69. Therefore, efforts to mitigate urban heat and air quality issues simultaneously can lead to synergies or trade-offs, and actions focused on one issue may unintentionally worsen the other.

In this study, we develop a multi-objective optimization framework for urban development in the Great Bay Area (GBA), one of the most densely populated and developed urban clusters in China. This framework integrates urban thermal comfort and improved air quality. Recognizing that urban green spaces can enhance thermal comfort while degrading air quality^17,18,70, we also incorporate green space allocation within the proposed framework. We employ geographically weighted regression (GWR) to link land use decisions to objective functions^71–74. Multiple objectives are solved using the Non-dominated Sorting Genetic Algorithm III (NSGA-III)⁷⁵. We further advance the framework by evaluating candidate plans using a climate and chemistry coupled model (WRF–GC)^76,77, which enables two-way meteorology–chemistry feedbacks and provides physically grounded mechanistic interpretation. By applying the proposed optimization framework during a compound heatwave and ozone episode, we demonstrate that co-mitigation of urban heat and air pollution is achievable, despite conflicting geophysical mechanisms. The framework has the potential to improve living conditions, decrease energy consumption, and reduce health risks in global urban areas.

Results

Robustness of estimated relationships between land use types and environmental variables

Robust relationships between environmental variables and land use types are crucial for constructing objective functions in optimization. The environmental variables considered include 2-meter air temperature (T₂), fine particulate matter (PM_2.5) concentrations, and ground-level ozone (O₃) concentrations. We first grouped the detailed land use types used in the urban climate-chemistry model into eight categories, namely tree, shrub, grass, wetland, crop, urban, barren, and water (Supplementary Table 1). The particulars for developing estimated relationships can be found in the Methods section. Evaluation metrics, including the coefficient of determination (R²), mean absolute error (MAE), and root square mean error (RMSE) were used. As shown in Fig. 1, GWR models demonstrate reliable predictability due to their consideration of spatial variability, with all models exhibiting mean biases smaller than 0.5%. The T₂ and land use relationship exhibits good fitting, with an R² value of 0.80, comparable to the findings of Duan et al.⁷⁴. (R² = 0.73). Although there is an overestimation of more than 1.5 K (Kelvin) for T₂ values smaller than 298 K, these deviations are rare (~ 1.63%) and exert negligible impacts on subsequent analyses. PM_2.5 and ozone predictions display superior performances, with R² = 0.82 and R² = 0.84, respectively, outperforming the models proposed by Huang et al.⁷⁸. (R² = 0.72 for PM_2.5 and R² = 0.65 for ozone). In summary, the established relationships demonstrate high robustness and can effectively serve as objective functions in the optimization process.

Fig. 1. Scattered density of predictions from geographically weighted regression (GWR) models and true values from WRF-GC simulations.

A Temperature at 2 m (T₂, unit: K); B Particulate Matter with aerodynamic diameter less than 2.5 micrometers (PM_2.5, unit: μg m⁻³); and C Ozone (unit: ppb). The Y-axis represents the predictions from the GWR models, and the X-axis represents the true values from the WRF-GC simulations. The red line shows the fitted regression line between the predicted and true values, while the black line represents the 1:1 reference line. The number of samples for validation is 9967. Source data are provided as a Source Data file.

Major features of optimized urban plans

We carried out spatial optimization with multi-objectives based on an assumption of 10% urban development (details in the Methods section). We then generated competitive urban plans through optimization (details are listed in Supplementary Note 1). These urban plans are all considered optimal on the Pareto front, meaning here that no single urban plan outperforms the others across all objective functions (Supplementary Fig. 1). This flexibility empowers decision-makers to customize their selections to specific requirements. Figure 2 delineates the spatial characteristics of optimized versus unplanned urban development. Figure 2A and B show the frequency among all urban plans that a given grid cell is developed as urban. Unplanned urban expansion exhibits a dispersed pattern across the area, with grid cells sharing similar probabilities for development. Conversely, optimal urban plans demonstrate an uneven likelihood of urbanization, with higher development potential in the western of the GBA urban clusters and more centrally along the coastline (as shown in red rectangles in Fig.2B), which corresponds to the upwind areas of the summer monsoon. Urban density, representing the average built-up (impervious) area fraction in an “urban” land use type grid cell, was categorized into low (<0.5), medium (0.5–0.8), and high (>0.8) density following the WRF-Noah urban canopy model. The average urban density for unplanned and optimized urban development is 0.77 and 0.64, respectively. Optimized urban land use plans favor medium-density urban areas, making up 75.68% of newly developed urban grid cells. In contrast, unplanned urban development tends to produce medium and high-density areas equally, accounting for 52.74% and 46.04%, respectively. High-density urban cells are predominantly covered by roofs and paved areas. A shift from high-density urban to medium-density urban indicates a decrease of impervious spaces, which can lead to better ventilation and reduced heat retention within the urban environment. We averaged green space fractions across all newly developed urban grid cells. Figure 2E displays that unplanned urban development exhibits low vegetation cover with a 0.15 areal share, whereas optimized urban development plans exhibit more green spaces with a share of 0.27 (Fig.2F). The optimized plans promote a shift from impervious surfaces to more contiguous green spaces, thereby enhancing shading and evapotranspiration, improving thermal comfort, and increasing stormwater retention. Overall, optimized urban development, prioritizing sites in the western GBA and coastline and featuring a medium-density urban form with increased urban green spaces, improves urban thermal comfort and air quality.

Fig. 2. Urban spatial comparisons of unplanned (A, C, E) and optimized (B, D, F) urban development.

A the frequency that grid cells are developed as urban calculated from unplanned urban development plans; B the same with (A) but calculated from optimized urban development plans. Red rectangles are two dominant developed urban areas in optimized urban plans; C averaged built-up density in a developed urban grid cell from unplanned urban development plans; D the same with C but from optimized urban development plans; E averaged urban green spaces fraction within a developed urban grid cell from unplanned urban development plans. F the same with (E) but calculated from optimized urban development plans. Source data are provided as a Source Data file.

Urban heat and air quality benefits of the optimized urban plans

We used regression to evaluate the effectiveness of the urban development optimization framework, which was established by analyzing the relationships between environmental factors and land use patterns. As shown in Fig. 3, unplanned urban development leads to a 0.04 K rise in T₂ over the whole domain, whereas optimized urban development yields a 0.01 K increase. Although the 0.03 ± 0.001 K reduction from optimization is small, it is statistically significant at the 95% significant level based on a t-test analysis. Because only 1.10% of grid cells were changed in the 10% urban development scenario, the domain-mean signal is diluted by the large share of unchanged areas. Over the changed urban cells, average mitigated warming is 0.20 K, with maximum local cooling of 1.42 K. Kang et al.⁷⁹ indicated that each 1 K rise in urban temperature is linked to a 6% increase in cardiovascular risk and a 7% increase in coronary heart disease risk. Given GBA’s dense and aging population, the 0.2 K cooling could lead to measurable health benefits. Unplanned urban development is associated with increased PM_2.5 levels and ozone concentrations, reflecting that higher anthropogenic emissions is conducive to pollutant formation over urban areas. In the optimized cases, both regional means and changed urban area means show less degraded or even better air quality, indicating that strategic siting can reduce exposure in the areas where development occurs. ESV also declines under urban expansion, but the loss is smaller with optimization because the framework favors preserving vegetation and ecological connectivity.

Fig. 3. Assessment of urban development optimization results based on estimated relationship of environmental variables and land use.

A the averaged impacts of T₂ (unit: K), PM_2.5 (unit: μg m⁻³), Ozone (unit: ppb), and ecosystem service value (ESV) on all grid cells; B the average impacts of T₂, PM_2.5, and Ozone on grid cells with changed land use. For each categorial pair, the left box is from unplanned urban development, and the right box is from optimized urban development. Y label indicates the difference between the urban development scenario and the current urban status (baseline). In each boxplot, the center line indicates the mean value, the box edges represent the first and third quartiles, and the whiskers denote the minimum and maximum values within 1.5 times the interquartile range from the lower and upper quartiles, respectively. Scatters represent the values for all plans (number = 400). Source data are provided as a Source Data file.

Simple regressions between land use and environmental variables seldom incorporate realistic feedback between meteorology and pollutants. Therefore, we performed the simulations based on the WRF-GC model to further assess the effectiveness of optimization strategies. Given that simulations for all generated urban land use plans are resource-intensive, we selected ten representative urban plans using the ranking of weighted scores from normalized objective function values (see methods). Five of these urban plans were derived from unplanned urban development scenarios, while the other five were taken from optimized scenarios. The objective function values distribution of these ten urban plans is summarized in Supplementary Fig. 2. The distribution of newly developed urban grid cells and urban green spaces fraction in each simulated urban plan is shown in Supplementary Fig.3 and Supplementary Fig. 4. The selected urban plans can present the general characteristics of optimal urban plans.

Model simulations indicate that optimized land use plans consistently outperform scenarios of unplanned urban expansion across all environmental objectives (Fig. 4). Unplanned development leads to increases in ozone concentrations across the entire domain and within newly developed urban areas by more than 2ppb, whereas optimized plans suppress O₃ at both the regional scale (−1.60 ppb) and within urban areas (−2.67 ppb) (Fig.4A). For PM_2.5, both scenarios result in substantial reductions; however, the optimized plans deliver greater benefits, reducing PM_2.5 concentrations by 20.80% in the changed urban cells. To investigate the drivers behind these improvements, we focused on two principal urban development areas identified in the optimized plans (red rectangles in Fig. 2B). In Location 2, optimization strengthens coastal near-surface winds and increases planetary boundary layer height (PBLH), while in Location 1, both wind speed and PBLH are enhanced (Supplementary Fig. 5). These changes collectively improve urban ventilation through more effective horizontal dispersion and enhanced vertical mixing. Consistent with this mechanism, vertical profiles reveal pronounced decreases in PM_2.5 and O₃ throughout the 0–5 km layer after optimization (Supplementary Fig. 6), indicating that enhanced mixing and stronger dispersion clean the entire column, rather than merely redistributing pollutants near the surface. Near-surface air temperature increases under both scenarios; however, the optimized plans consistently moderate the warming relative to unplanned expansion (Fig.4B). The spatial pattern of T₂ under the optimized plans indicates that cooling is concentrated along the coastline (Supplementary Fig. 7), coinciding with the areas where urban development is preferentially allocated. By aligning urban development with the upwind of the summer monsoon regime, optimized coastal urban areas are exposed to more frequent and stronger southernly winds (Supplementary Fig. 8), which contribute to additional cooling benefits.

Fig. 4. Comparisons of urbanization environmental impacts between unplanned urban development and optimized urban development based on WRF-GC simulations.

A Ozone (unit: ppb); B T₂ (unit: K); and C PM_2.5 (unit: μg m⁻³). The Y-axis represents the differences between urban development plans and the baseline (current urban development status). Source data are provided as a Source Data file.

Discussion

The expansion of urban areas has led to significant environmental challenges that require urgent mitigation efforts. In this study, we introduced a viable method for optimizing urban development with a focus on environmentally conscious planning. By acknowledging the conflicts between urban warming and the deterioration of air quality, our proposed framework integrates four trade-off objectives within the optimization process. We utilized the NSGA-III algorithm to produce a diverse array of urban plans for stakeholders to select. When applied to a case study in the GBA, this approach revealed several important insights. Both regression analyses and numerical modeling simulations highlight the effectiveness of optimized urban development, with benefits extending beyond urban areas into the surrounding areas. The simulations attribute air quality improvements primarily to enhanced ventilation, and heat mitigation to southward summer monsoon flows. By incorporating environmental considerations into urban development optimization, it is possible to co-mitigate urban heat and air pollution, both of which are associated with dense population and rising energy consumption in urban areas.

Our optimization framework determined the locations and the fractions of built-up and green spaces in modified urban grid cells on a regional scale. Although the spatial resolution is coarser than that used in typical land use optimization studies at the street or block level^38,45,48,80, regional scale optimization more effectively captures atmospheric circulation, pollutant transportation, boundary layer evolution, and meteorology-chemistry interactions. In China, urbanization is characterized by clusters of cities⁸¹, and its impacts extend beyond local urban cores into the surrounding regions. It is therefore crucial to understand the regional impacts of urbanization and to find effective ways to mitigate its negative effects. Performing urban land use optimization at this larger scale not only highlights the regional impacts of urbanization but also enhances our mechanistic understanding of how optimized land use configurations can achieve co-mitigation benefits.

Green roof and cool roof are widely used strategies in megacities to mitigate UHI, particularly during the summer time. We implemented WRF-GC-based simulations to evaluate these two retrofits against the proposed optimized planning strategy, assuming a 10% urban expansion relative to the current quo (see Methods). Although green roofs and cool roofs result in a greater decrease in near-surface temperatures compared to the optimization strategy, they also degrade air quality, leading to increased concentrations of ozone and PM_2.5 (Supplementary Fig. 9 and Supplementary Table 2). Green roofs emit additional Biogenic Volatile Organic Compounds (BVOC) emissions, which promote the formation of SOA¹⁸. Cool roofs lower the surface temperature and boundary layer height, thereby inhibiting the dispersion of PM_2.5 and ozone⁵⁶. This comparison highlights that optimized urban development offers a different perspective to achieve co-benefits of thermal comfort and improved air quality without incurring substantial retrofit costs.

Despite these advances, several limitations remain and should be acknowledged. First, in the WRF urban canopy model, the fraction of urban green spaces is prescribed uniformly across all urban grid cells. This configuration does not account for the spatial heterogeneity introduced by the optimized plans, nor does it capture the enhanced evapotranspiration associated with increased greenery, which may underestimate the reduction in T₂ under the optimized urban development scenarios¹⁸. Future work should incorporate spatially varying urban vegetation parameters, as well as explicit representations of urban trees and irrigated surfaces, to improve the realism of thermal response. Second, the WRF-GC modeling system involves uncertainties from multiple sources, including initial and boundary conditions, emission inventories, chemical mechanisms, physical parameterizations, and numerical schemes. Due to computational constraints, it is not feasible to conduct a systematic uncertainty analysis, such as large ensembles or comprehensive sensitivity experiments, which limits our ability to robustly quantify the uncertainty of our results. Future research should aim to further explore numerical modeling techniques to improve the accuracy of models and emissions estimates based on altered land use patterns. The optimization results are influenced by several factors, including local meteorological characteristics, air pollutant transport pathways, and the regional atmospheric circulation background. Consequently, when implementing the proposed method in other cities, it is essential to carefully evaluate and account for these site-specific variables in order to ensure the effectiveness and applicability of the approach.

Methods

Coupled urban climate-chemistry modelling

We used WRF-GC v2.0⁷⁶ to perform the urban climate and air quality simulations. WRF-GC is an online coupling of the Weather and Research and Forecasting meteorological (WRF) model and the GEOS-Chem atmospheric chemistry model. Previous studies have demonstrated its satisfactory performance against observations regarding the magnitudes and spatial patterns of air pollutants and meteorology variables^{76,77,82–84}. Supplementary Fig. 10 shows the characteristics of meteorological and air pollution variables during the selected event. We configured 3 domains with grid resolutions of 27 km, 9 km, and 3 km, respectively (Supplementary Fig. 11). The 3 km inner domain is convection-permitting, allowing for explicit resolution of key urban and coastal circulations in the GBA, while maintaining computational feasibility for simulations across all ten scenarios. The primary parameterization options used for physical schemes and domain settings are listed in Supplementary Table 3. The meteorological initial and boundary conditions were obtained from the National Center for Environmental Prediction Final Operational Global Analysis (NCEP FNL, 1° × 1°) with 6-hour intervals. The chemical boundary and initial conditions were taken from a standard full-chemistry simulation from GEOS-Chem v12.8.1. The monthly mean anthropogenic emissions covering China with 0.25° × 0.25° grid resolution in 2020 were obtained from the Multi-resolution Emission Inventory for China (MEIC)⁸⁵. Anthropogenic emissions for the areas outside China were obtained from a mosaic Asian anthropogenic emission inventory^86,87. Validation against observations indicates that, although biases exist, the accuracy is comparable to that reported in other studies^59,76,88. Further details regarding the validation are provided in Supplementary Note 2, Supplementary Fig. 12, and Supplementary Tables 4 and 5.

Multi-objective optimization framework for urban land use development

Supplementary Fig. 13 outlines the structure of multi-objective optimization for urban development. Initially, we identified the key objectives and represented them with mathematical functions. Subsequently, optimization was conducted based on NSGA-III to yield anticipated results. The effectiveness of optimization was evaluated through regression analysis and numerical simulations.

Our primary motivation is to promote sustainable urban development by simultaneously mitigating urban warming and enhancing air quality. Urban heat is represented by T₂, which is an important indicator related to human health⁸⁹ and a widely used variable for evaluating urban heat adaptation measures³². Urban air quality is characterized by hourly concentrations of PM_2.5 and the maximum daily 8-hour average ozone (MDA8 O₃). We adopted MDA8 O₃ rather than the daily mean to capture sustained periods of high exposure, as O₃ exhibits a strong diurnal cycle driven by photochemical processes, with pronounced peaks in the afternoon. The 2019 Outline Development Plan for the GBA advances an “ecological civilization” and calls for strict protection of key ecological spaces⁹⁰. To promote resilience, we incorporated the maximization of ecosystem services value (ESV) as one of the objective functions³⁸. The ESV is calculated as:

$$ESV={\sum }_{i=1}^n{\sum }_{j=1}^8{\alpha }_j{f}_j$$ 1

Where, $i$ indexes grid cells (n is the total number of grids) and j indexes the land use type. ${\alpha }_j$ denotes the ESV coefficient per unit area for land use type $j$, obtained from Ding, et al.⁹¹ and tailored to the GBA. $f_j$ denotes the fractional coverage of land use type $j$ in grid cell $i$ (range from 0 to 1).

According to the China Statistical Yearbook for 2012 to 2023, the built-up urban area in Guangdong province increased by about 7% from 15984.1 to 17126.6 square kilometers^92,93. We therefore selected a 10% urban expansion scenario to reflect a plausible and policy-relevant decadal change. We also evaluated the urban expansion ratios of 5% and 15%, and obtained results similar to those of the 10% scenario (Supplementary Fig. 14). For consistency across experiments, all data, including land use, land use fractions, meteorological data, and air quality variables, were extracted from the simulations of the innermost domain (Supplementary Fig. 8). This domain comprises 1105 urban-type grid cells, each with a horizontal spatial resolution of 3 km x 3 km. Notably, a cell is classified as urban if at least one-third of its area has urban land use, while the remainder can have other land uses such as water bodies, forests, shrubs, and crops. Several constraints were applied to ensure that the selection of newly developed urban cells is consistent with regulatory and geographical realities in the region: (1) Urban development should not encroach upon croplands or water bodies, in order to protect these vital resources; (2) The elevation for new urban cells must not exceed 95 m, which corresponds to the 95th percentile of current urban elevations; (3) Selected urban cells should have water on no more than 50% of their borders to exclude isolated islands.

Application of these constraints yielded 1164 grid cells being identified as candidate developed urban sites. Then the optimization issue becomes determining which grid cells should be developed, as well as how to allocate the urban fraction and green spaces fraction within each selected cell to achieve optimal co-mitigation of urban warming and degraded air quality. The distribution of potential urban grid cells and urban land use grid cells is illustrated in Supplementary Fig. 15.

Considering that highly urbanized areas in the GBA are densely concentrated around the Pearl River estuary, we imposed a compactness⁴⁵ constraint on each candidate urban plan. Compactness quantifies the similarity among neighboring land uses. We calculated local compactness by counting the number of urban cells within the 3×3 neighborhood surrounding each focal urban cell. Based on the landcover dataset from the Moderate Resolution Imaging Spectroradiometer (MODIS)⁹⁴, we identified a positive trend in compactness of 0.016 per year in the GBA from 2001 to 2024, indicating progressively more compact urban development (Supplementary Fig. 16). Accordingly, we set threshold for the mean urban compactness at 5.4 for each urban plan, corresponding to the average value observed over the most recent five years. This constraint ensures alignment with the contemporary urban growth pattern in this region.

Given the potential trade-offs among different objectives, no singular optimal solution is attainable. We therefore employed the widely used Non-dominated Sorting Genetic Algorithm (NSGA-III)^95–97 to generate a diverse set of Pareto-optimal urban plans for decision-makers to consider. NSGA-III is a state-of-the-art multi-objective evolutionary algorithm specifically designed to address problems involving more than three objectives. A reference point-based scheme⁹⁸ is introduced in this algorithm to maintain solution diversity in high-dimensional objective spaces by guiding the search and preserving a well-distributed set of solutions along the Pareto front. Detailed technical particulars for NSGA-III are provided in Supplementary Note 1.

Relationship between land use types and T₂, ozone, and PM_2.5

Considering the spatial heterogeneity across the GBA area, we utilized the GWR model to estimate the environmental variables. This approach allowed us to determine the autocorrelation between land use types and environmental factors. The formulation is represented as follows:

$$y_i={\beta }_0\left(u_i,v_i\right)+{\sum }_{k=1}^N{\beta }_k\left(u_i,v_i\right)x_{ik}+{\epsilon }_i$$ 2

where N is the number of predictor variables$;$ $\left(u_i,v_i\right)$ denotes the coordinates pair for location $i$; ${\beta }_k$ signifies the varying weight; and ${\beta }_0$ represents the changing intercept based on locations. Thus, GWR extends traditional global regression techniques to accommodate non-stationary spatial relationships⁷². Finally, $y$ is the dependent environmental variable, in particular T₂ to represent climate and hourly PM_2.5 and ozone concentrations to represent air quality.

We took both the effects of each land use type at cell $i$ and all adjacent cells into account to predict the dependent variables. Land use types in WRF-GC simulations were compiled into a smaller number of classes (Supplementary Table 1). The effects from adjacent cells were calculated by summing the inverse distance weighted (IDW) with a window size of 5 × 5.

To ensure precision in the regression relationships, we also employed the Least Squares method and Random Forest regression for comparison with the GWR method. The results indicated that GWR provided superior accuracy (Supplementary Figs. 17 and 18). Although T₂ and PM_2.5 variations show diurnal patterns³⁷, our analysis of GWR regression coefficients for different land use types during the day and night revealed consistent characteristics (Supplementary Fig. 19). Therefore, we chose not to include diurnal variations in the T₂ and PM_2.5 regression models to simplify the number of objective functions.

Assessment of spatially-optimized planning strategies

Spatially optimized urban planning strategies were evaluated by integrating the estimated relationships with the WRF-GC model. To assess the estimates, environmental variables were calculated for both optimized and unplanned urban development scenarios.

Given the computational constraints of running the WRF-GC model for all urban plans, we simulated ten representative cases: five from the unplanned set and five from the optimized set. To identify these cases, we normalized each objective function value to a scale from 0 to 1 within each urban plan set. We then calculated a composite score for each urban plan as the weighted sum of the normalized values for the four objective functions. Weights for T₂, PM_2.5, O₃, and ESV-related objective functions were assigned as 0.3,0.3,0.3 and 0.1, respectively. We prioritized T₂, PM_2.5, and O₃ to emphasize co-benefits for urban heat mitigation and urban air quality improvement. For each set, plans were selected at the minimum score, the 25th percentile, the median, the 75th percentile, and the maximum score. These five plans cover both the extremes and bulk of the normalized objective function value distribution, yielding a set of scenarios that are representative across the full performance spectrum defined by the weights. This approach ensures coverage of diverse performance regimes, while accurately reflecting the range of outcomes prioritized by the weights. Supplementary Fig. 2 presents the distribution of objective function values from selected plans.

Updated land use, land use fraction, leaf area index, and vegetation type fraction data were obtained to align with expanded urban land use scenarios before being integrated into the model. Adjustments to anthropogenic emissions were also necessary in response to changes in land use, given the distinctive characteristics of various land types. Emission reallocations were implemented following the procedure detailed in Supplementary Note 3. The total emissions after adjustment were consistent with those during the case study. Supplementary Fig. 20 illustrates the differences in PM_2.5 emissions for one of the optimized urban plan sets. Increased PM_2.5 emissions from power, industry, residential, and transport can be seen in the newly developed urban areas.

All model physics, chemistry, and boundary conditions were kept identical across the simulations of the selected urban plans, with only land use inputs and the corresponding anthropogenic emissions varied. We specifically focused on differences between the unplanned and optimized scenarios; thus, systematic biases are minimized under this experiment design, allowing us to attribute scenario differences primarily to optimization. Therefore, model uncertainties may not significantly affect the results of this study⁸³.

We employed the WRF–GC modeling system to quantify the impacts of an optimized urban expansion strategy relative to two widely implemented retrofit measures, namely green roof and cool roof, on urban heat and air quality. For comparability, the urban development plan for the retrofit scenarios was from the initial plan set as the member at the 25th percentile of the objective function ranking, and the optimized strategy was analogously selected as the 25th percentile within the ensemble of optimized plans. For the green roof scenario, we enabled the green roof option and set the green roof fraction to 0.5 in urban grid cells. In addition, BVOC emissions from green roofs were calculated using the Model of Emissions of Gases and Aerosols from Nature (MEGAN). In this way, both the evapotranspiration effect and BVOC emissions from green roofs were taken into account. For the cool roof scenario, 50% of urban roofs were assigned a high-albedo material (albedo = 0.8), while the remaining 50% retained the baseline albedo of 0.2, resulting in an effective roof albedo of 0.5 at the urban grid-cell level. All other model configurations (including domain, period, physics, chemistry, boundary conditions, land use, and anthropogenic emissions) were held constant across all scenarios to ensure that inter-scenario differences reflect only the specified modifications.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

M.G. designed the research. L.Z. performed the development of the optimization framework and data analysis. F.W. carried out the WRF-GC model simulations and model evaluations. C.N. contributed the ideas about the urban development. G.C. offered insights about the interactions between air quality and urban heat. L.Z. and M.G. wrote the manuscript.

Peer review

Peer review information

Nature Communications thanks Lili Ji and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

National Natural Science Foundation of China (no. 42322902), the Research Grants Council of the Hong Kong Special Administrative Region, China (Project Nos. C2002-22Y, HKBU12201023 and HKBU12202824), and the Center for Ocean Research in Hong Kong and Macau (CORE), a joint research center between the Laoshan Laboratory and HKUST.

Data availability

NCEP FNL data are accessible at https://rda.ucar.edu/datasets/ds083.2/. The Multi-resolution Emission Inventory for China (MEIC) can be obtained from http://meicmodel.org.cn/. The observed hourly near-surface temperature and relative humidity at 2 m for validation of the WRF-GC simulation are sourced from the China National Meteorological Information Center (NMIC, https://data.cma.cn/dataService/cdcindex/datacode/A.0012.0001). Other data are available in the main text, methods, or the supplementary materials. Source data generated in this study are provided in the Source Data file. Source data are provided with this paper.

Code availability

The used climate and chemistry coupled model (WRF–GC) can be downloaded from https://github.com/WRF-GC/wrf-gc-release.git. Code to reproduce the findings is available at https://github.com/LiuhuaZhu1/urban_planning_optimization.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-73855-x.