Spatially-optimized urban greening for reduction of population exposure to land surface temperature extremes

Over half of the world’s population lives in cities, which occupy less than 3% of Earth’s land surface, and urban populations are projected to grow substantially by 2050. Urbanization alters local climate by replacing natural cover with impervious materials, reducing albedo and evaporative cooling, and reshaping the balance among impermeable surfaces, vegetation, and water bodies. Combined with global warming, these changes are increasing the number of people exposed to health-endangering heat, with rising risks of morbidity and mortality during heat waves. Traditional planning approaches often overlook social and environmental factors, underscoring the need for knowledge-based, city-specific adaptation strategies and tools to monitor and model urban thermal environments at appropriate spatial and temporal scales. Prior remote-sensing studies largely emphasized the Surface Urban Heat Island (SUHI)—the urban-rural LST difference—which is influenced by surrounding rural vegetation and does not directly represent urban thermal comfort. This study addresses the need for city-specific assessments by predicting absolute population exposure to LST extremes using remote sensing-based spatial modeling, focusing on how urban vegetation and proximity to water mitigate exposure, and evaluating optimized greening strategies across 200 cities worldwide.

The paper situates its contribution within extensive literature on urban heat, remote sensing of urban climates, and mitigation through urban green infrastructure. It notes that SUHI metrics can conflate urban thermal conditions with characteristics of surrounding rural landscapes, motivating a shift to direct exposure metrics. Prior work demonstrates that vegetation cools urban environments through shading and evapotranspiration and that remote sensing enables globally consistent, high-resolution thermal mapping. The study builds on guidance from IPCC and WMO for defining heat extremes (e.g., using high-percentile thresholds) and responds to calls for city-specific, spatially detailed models. It also acknowledges complementary mitigation strategies (e.g., increasing surface albedo) and research on the socio-spatial inequities of heat exposure, framing the need for targeted, equitable interventions.

Study design and scope: The analysis covers 200 cities globally across major Köppen–Geiger climate zones (Arid, Continental, Temperate, Tropical). Cities are defined using GHSL urban boundaries with a 5 km buffer. Each urban area is gridded at 1 km resolution for analysis. Data sources and preprocessing: - Land Surface Temperature (LST): MODIS MOD11A1 V6 daily LST and emissivity (nominal 1 km), day (~10:30) and night (~22:30) overpasses. For each city, day and night LST thresholds are computed as the 90th percentile of the 20-year distribution of city pixels. Days/nights exceeding thresholds are tallied over the 3 warmest months per year. - NDVI: MODIS MOD13Q1 V6 (250 m). NDVI values are averaged to the 1 km grid cells. - Population and water: From the EU GHSL datasets. Distance to water bodies is computed using the Guidos toolbox; predictor d_w = 1/D_w (normalized 0–1). - Warmest months selection: Determined per city-year using ERA5/ECMWF Copernicus climate data. All variables are aggregated in Mollweide projection at 1 km. Exposure metric: - For each 1 km pixel, compute the number of daytime (T_E^D) and nighttime (T_E^N) days exceeding the respective LST thresholds during the 3 warmest months. Total exposure T_E^T is defined as the average of day and night exposures: T_E^T = (T_E^D + T_E^N)/2. Population exposure (person-days) is the pixel exposure times the pixel population; city totals sum across pixels. Modeling approach: - Dependent variables: Y_D and Y_N = average number of days/nights over thresholds (2010–2020) per pixel. - Predictors: X1 = NDVI (fractional greenness), X2 = proximity to water (d_w). - Spatial model: Spatial Lag Model (SLM): Y = βX + ρWY + ε, where W is a row-standardized spatial weights matrix (k-nearest neighbors, k=8, equal weights) capturing spatial autocorrelation among neighboring pixels. - Training and validation: k-fold cross-validation across pixels within each city; model performance assessed via R^2 and MAE on training/validation/test splits. Ordinary Least Squares (OLS) is evaluated as a baseline. Greening simulations: - Scenario definitions for mitigation analysis using fitted SLM coefficients: - Scenario 1 (uniform): Increase NDVI homogeneously by a specified increment (e.g., +0.3) for all pixels with NDVI < 0.85, capped at NDVI = 0.85; distance-to-water remains fixed. - Scenario 2 (targeted): Increase NDVI only in the most populated (high-exposure) pixels to achieve the same total city exposure reduction as Scenario 1; report both local (per-pixel) increments and total citywide NDVI increment. - Simulations are run for Paris as an illustrative case and extended to all 200 cities. Additional sensitivity analyses examine NDVI increments from 0.1 to 0.5. Metrics include exposure reduction (ΔT_E_T %) and relative NDVI savings: NDVI(%) = 100*(NDVI_SO – NDVI_All)/NDVI_All, where NDVI_All and NDVI_SO are total post-intervention NDVI under uniform and targeted approaches, respectively.

- Exposure trends: Urban population exposure to LST extremes increased from 2010 to 2020, largely due to urbanization. The exposure per capita (exposure divided by population) remained approximately constant across years and climate zones. - Model performance: The SLM accurately predicts daytime and nighttime exposure across 200 cities with test-set R^2 ≥ 0.8 in k-fold cross-validation, outperforming OLS (which showed low/negative R^2). This underscores the importance of accounting for spatial autocorrelation. - Drivers and coefficients: Vegetation (NDVI) and proximity to water (d_w) are significant and negatively associated with exposure. Average regression coefficients (mean over best fits in cross-validation): - Daytime (T_E^D): β_NDVI ≈ -4.7, β_d ≈ -1.2. - Nighttime (T_E^N): β_NDVI ≈ -2.7, β_d ≈ -0.6. Increases in greenness and proximity to water reduce the number of days/nights over thresholds. - Paris case study (3 warmest months, 2010–2020 average): - Observed total exposure: ~164 million person-days. - Scenario 1 (uniform +0.3 NDVI up to 0.85): exposure reduced to ~144 million person-days (−12%), requiring a 44% increase in total city NDVI. - Scenario 2 (targeted most populated areas to match −12%): achieved with local NDVI increases ~0.38 in those pixels and only a 14% increase in total city NDVI, demonstrating substantial efficiency gains by targeting. - Global simulations across 200 cities (NDVI +0.3 uniform): average exposure reductions by climate zone: - Temperate: ~12% - Arid: ~13% - Continental: ~16% - Tropical: ~32% - Targeted vs uniform greening efficiency: - Relative savings in total NDVI required to achieve the same exposure reduction by targeting high-exposure areas (most populated): ~90% (Arid), ~20% (Continental), ~26% (Temperate), ~33% (Tropical); up to ~70% greening saved on average across increments. - Average local NDVI increment required in targeted areas is ~0.45. Local (targeted) increments were on average ~0.15 lower than global uniform increments needed to achieve the same reductions. - Exposure reduction scales approximately linearly with NDVI increment (0.1–0.5), while relative NDVI savings from targeting remain stable across increment sizes.

The study demonstrates that a simple, globally applicable spatial lag model using only remote sensing-derived covariates (NDVI and distance to water) can reliably predict urban exposure to LST extremes at 1 km resolution, providing a practical, scalable tool for inter-city comparisons and preliminary planning. Results highlight that urban vegetation and water proximity are key determinants of exposure, and that strategic placement of greening—targeting the most populated/high-exposure pixels—substantially improves efficiency compared to uniform interventions, reducing the overall greening required to achieve a given exposure reduction. These findings address the research need for city-specific, spatially explicit assessments of heat exposure and mitigation potential. Importantly, the work reframes SUHI-focused assessments toward absolute exposure metrics more directly relevant to urban residents. While validated across diverse climates, the results should guide rather than replace local planning: greening impacts can vary by neighborhood and climate (e.g., irrigation effects in arid regions), and equity considerations suggest prioritizing historically marginalized and vulnerable communities. Moreover, outdoor exposure measured via LST does not account for indoor conditions and differential access to cooling, which are critical for health outcomes. The modeling framework offers actionable insights for prioritizing zones where greening yields the largest reductions in exposure per unit of intervention and can be integrated with complementary strategies (e.g., high-albedo surfaces, water features) for greater cumulative benefits.

This study introduces and validates a global, remote sensing-based spatial lag model that accurately predicts population exposure to LST extremes across 200 cities. It quantifies how increases in urban vegetation reduce exposure, and shows that targeting high-exposure, densely populated pixels can achieve the same exposure reductions with substantially less greening than uniform citywide interventions. The approach provides decision-makers with quantitative, spatially resolved guidance to design efficient climate adaptation strategies. Future work should: (i) incorporate additional mitigation levers (e.g., surface albedo modification, water bodies), (ii) account for socio-economic vulnerability and equity in targeting strategies, (iii) refine exposure metrics using air temperature and microclimatic factors (radiation, wind, humidity) at finer spatial/temporal scales, and (iv) integrate indoor exposure and cooling access to better capture health-relevant risk profiles.

- LST as a proxy: LST measures surface temperature, not air temperature; it may not fully represent human thermal exposure, especially at hyper-local scales. Urban canyon effects, emissivity variability across materials/vegetation, and diurnal/weather variability can introduce errors. - Temporal/spatial aggregation: Analysis averages day/night LST over the 3 warmest months at 1 km resolution and across years, not capturing hourly/daily extremes or microclimates within neighborhoods. - Exposure definition: Uses the 90th percentile LST thresholds to define extremes, aligning with guidance for extremes but potentially omitting relevant outliers. - Limited mitigation scope: The simulations consider only greening; other strategies (e.g., increased albedo, building energy efficiency, codes) are not modeled. - Practical constraints not modeled: Space availability, water use, maintenance, financial and social factors, and urban form constraints limit feasible greening and are not included. - Indoor exposure not assessed: Differences in access to cooling and indoor environments, which strongly influence health risk, are outside the scope.