The potential of urban trees to reduce heat-related mortality in London

Near-surface temperate data was obtained from two different sources from 01 October 2015–30 September 2022: official Met Office weather stations and Netatmo PWS stations within a wide bounding domain that includes the GLA. This timeframe includes the historically hottest summer of 2018, and the three exceptional heatwaves that occurred during the summer of 2022 (15th–17th June, 17th–19th July, and 9th–15th August) when temperatures reached a maximum of 39 °C at London Heathrow.

Hourly temperatures, station ID, and coordinates for Met Office stations were obtained from September 2015–December 2022 from MIDAS Open [34]. This included the station at London Heathrow, commonly used for studying the London climate and which offers a comprehensive coverage of the period of interest. This was used as a reference station.

PWS station ID, and coordinates was obtained using the Patatmo Python module that utilizes the Netatmo API (https://dev.netatmo.com/). This resulted in 619 stations, of which 502 were within GLA administrative boundaries. Hourly air temperatures were then downloaded for PWS stations for the study period. Cleaning and analysis of PWS temperatures was carried out using R. First, the altitude of each station was spatially joined using a 10 m resolution Digital Elevation Model of London [35]. Height correction and outliers removal was performed using CrowdQC+ [25], a tool developed for quality control of PWS weather data. Hourly measurements were removed if they were taken by duplicate stations located at the same coordinates (Step M1); were outliers, as determined by their z-score (Step M2); if more than 20% of the measurements in the month were removed in the previous steps, indicating the PWS may be unreliable (Step M3); and finally if the measurements were not correlated to the median temperature of other measurements (Pearson r < 0.9) and are assumed to be indoors (Step M4). We then excluded days with more than three hours of data absent. This reduced the number of stations within the GLA boundaries to 490, with 520 623 sensor-days of valid PWS measurements.

The Met Office and PWS temperatures were combined into a common dataset (figure 1), except for Heathrow which was used as a reference temperature for each date. Stations located in rural areas outside the GLA were used as a baseline to estimate UHI intensity. These rural areas were identified via a spatial join to 2011 Office for National Statistics rural/urban classifications [36] for different census output areas.

Zoom In Zoom Out Reset image size

2.1.1. Environmental and population data

Tree canopy coverage data for the GLA is from the London Data Store [37] (figure 2(a)). The dataset contains 15 041 hexagons, each 350 m across, derived from high resolution (10 cm per pixel) colour infrared imagery collected in September 2016, with pixels as being either a tree canopy or not using a machine learning algorithm. Further details on how the tree canopy data was derived can be found in [38].

Zoom In Zoom Out Reset image size

Urban temperatures also vary according to local climate zones (LCZs), or local surface structure and cover. As tree canopy is assumed to be correlated with the natural component of LCZs, we derived two continuous variables to describe the height and cover of the built component. For each hexagon, a spatial join to the OS MasterMap Building Height Attribute [39] dataset was used to calculate the median height of buildings (removing buildings with low measurement confidence) (figure 2(b)) and the percent area covered by all buildings (figure 2(c)). OpenPopGrid [40], a 10 m gridded population dataset, was used to calculate the population within each hexagon (figure 2(d)). To account for broader spatial impacts, buffers around the hexagons at 200, 500, and 1000 m were drawn and the area-weighted average BC and building height (BH) re-calculated within these buffers.

Hexagon canopy coverage was then adjusted for different scenarios:

• a)  
  Current scenario: a base case with existing tree canopy coverage.
• b)  
  Grey scenario: tree canopy cover is reduced to near zero, representing a scenario where trees are reduced to the lowest possible amounts seen per BC. In this scenario, the tree coverage was set to 0.05th percentile for each hexagon.
• c)  
  London strategy: tree canopy coverage each hexagon is increased 10%, as per the London Environment Strategy goals for 2050 [41].
• d)  
  Average strategy: hexagons below the population-weighted average canopy coverage are increased to the average.
• e)  
  Green scenario: tree canopy cover is increased to an amount that represents a maximum amount, constrained by existing BC. Quantile regression was used to identify the maximum (99.5th percentile) tree coverage according to the BC within each hexagon. In cases where tree coverage exceeded the 99.5th percentile, the current coverage was maintained.

Finally, to estimate the carbon benefits of these scenarios, the estimated 2.4 million tonnes of carbon stored and 77 200 tonnes of carbon sequestered annually by London's trees [42] were recalculated given the proportional changes in tree coverage.

Analyses were performed to produce descriptive statistics on UHI intensity and examine how air temperatures vary by the amount of tree canopy coverage.

Hourly temperature data was used to calculate UHI intensity for each day, defined as the maximum hourly difference between stations within the GLA boundaries (90th percentile) and the average of stations in the larger domain that are classified as rural. For the three heatwaves during the 2022 summer, the UHI intensity, hourly timeseries, and differences in 24 h, daytime, and nighttime mean, maximum, and minimum temperatures in areas with differing amounts of tree coverage were calculated. Daily (d) 24 h mean (${T_{{\text{mean}},\;d}}$) was calculated for all stations and joined by date to the corresponding value at the heathrow reference station (${T_{{\text{mean}},\;d,\;{\text{heathrow}}}}$).

GAMs were then fit to the aggregated daily data using the mgcv package in R. GAMs were used because they allow the response variable to depend on smooth functions of the predictor variables; this can present an advantage over linear regression models which contain the assumption that response is linear in the predictor variables. Another advantage is the assumption that effects of each variable are separate, enabling extrapolation to the counterfactual scenarios. The effects of the predictor variables were assumed to be independent to enable counterfactual reasoning about the effect of changing tree cover while keeping other aspects of the built environment the same. Smoothing parameters were estimated using restricted maximum likelihood (REML), with degrees of freedom limited to 9 a priori to improve transferability for the counterfactual scenario and interpolation between data points. The response variable was ${T_{{\,\text{mean}},\;d}}$ and the predictor variables were ${T_{{\,\text{mean}},\;d,\;{\text{heathrow}}}}$; tree canopy cover, BC, and BH in the hexagon which the station is located; and the BC and BH within 200 m, 500 m, or 1000 m buffers.

Models were fit for the different buffer sizes, using 5-fold cross-validation (stratified by station) to identify the buffer best able to predict temperatures at an unseen station. Model residuals were inspected, testing for spatial autocorrelation using a variogram and calculation of Morans I (for the whole dataset, summers, and each day), and examined for temporal autocorrelation using the Durban–Watson test.

The GAM model was then used to estimate temperature exposures from 2015–2022 under different tree coverage scenarios and the corresponding health impacts calculated. Rural temperatures from 2015–2022 were used as a baseline. For the same time period, the GAMS model was used to predict the daily mean temperature (${T_{{\text{mean}},\;d,k}}$) for each hexagon (k) under different canopy coverage scenarios. For each day, the population-weighted mean daily temperature for the GLA was calculated (equation (1)):

$$\begin{align}\overline {{T_{{\,\text{mean}},\;d}}} = \frac{{{{\mathop \sum \nolimits}}_{i = 1}^n\left( {{T_{{\text{mean}},\;d,k}}} \right) \times {P_k}}}{{{{\mathop \sum \nolimits}}_{i = 1}^n{P_k}}}\end{align} \tag{ 1 }$$

where ${P_k}$ is the population within a hexagon.

Health impact calculations were then carried out to estimate daily heat-related mortality. We use the heat-mortality relationship for London from Arbuthnott et al [43] (table 1), which derived a relative risk (RR) of mortality for different age groups (i) using lag 0, 1 ${T_{{\text{mean}},\;d}}$ (herein referred to as ${T_{{\text{mean}},\;d,{\text{lag}}}}$) for London. When ${T_{{\text{mean}},\;d,{\text{lag}}}}$ exceeds a temperature threshold (${T_h}$, or 18.9 °C in London) the RR of heat related mortality increases per degree C:

$$\begin{align}R{R_{i,d}}& = {e^{{c_a}\left( {{T_{{\text{mean}},\;d,\;{\text{lag}}}} - {T_h}} \right)}}\;{\text{when}}\;{T_{{\text{mean}},\;d,\;{\text{lag}}}} > {T_h} \nonumber\\ R{R_{i,d}}& = 0 \,\,\kern58pt{\text{when}}\;{T_{{\text{mean}},\;d,\;{\text{lag}}}} \unicode{x2A7D} {T_h}\end{align} \tag{ 2 }$$

Table 1. Heat-mortality relationship for London, from Arbuthnott et al [43].

Age group	Threshold, ${T_h}$	RR of mortality (95% CI), total
0–64	18.9 °C	1.022
65–74		1.024
75–		1.049

where c(a) is the natural log of the RR at 1 °C above the threshold. The attributable fraction (AF) of heat-related mortality is calculated as:

$$\begin{equation}A{F_{i,d}} = \frac{{R{R_{i,d}} - 1}}{{R{R_{i,d}}}}.\end{equation} \tag{ 3 }$$

This was used alongside the number of deaths per each day of the year for different age groups (${q_{i,d}}$) in London (2015–2022) [44, 45] to calculate heat-related mortality and summed across all days and ages to estimate attributable mortality (AM):

$$\begin{equation}AM = \sum\limits_d \sum\limits_i {q_{i,d}} \times A{F_{i,d}}.\end{equation} \tag{ 4 }$$

Heat AM under different scenarios was also examined using projected (1980–2010, 2041–2060, 2061–2080) climates for London for RCP8.5 from UKCP18 [46]. This data includes 12 ensemble members of average temperatures for each day of the year for these time periods. The GAM model was used to estimate $\overline {{T_{{\text{s}},d}}}$ for projected climates, and AM estimated using the number of deaths per day of year and age group averaged over 2015–2022.

The tree canopy coverage varied significantly, with a population-weighted average of 17.5% (range of 0%–99.9%), and a total of 336 km² of tree canopy with the GLA. The results of the quantile regression showed a correlation between the 99.5th percentile tree canopy and BC (figure 3) (1.5% reduction in tree canopy per 1% increase in building area, p = 0.00). PWS stations were relatively representative of the distribution of populated hexagons, although the maximum tree coverage was 64%. Population-weighted tree canopy coverage increased to 58% under the greening scenario (1242 km²) and 21% (395 km²) under the average scenario. Met stations were in areas with lower canopy coverage.

Zoom In Zoom Out Reset image size

Over the 2015–2022 period, the average UHI intensity was 3.3 °C, greatest during spring (mean of 3.8 °C, max of 7.6 °C) and lowest during winter (mean of 2.9 °C, max of 7.1 °C). An hourly timeseries of temperatures during the 2022 heatwaves (figure 4) shows the average of all PWS within the GLA, the average of those within the top decile (40.2%–99.9% tree canopy coverage), those in the bottom decile (0%–5.4% canopy coverage), and in the rural areas surrounding the GLA administrative boundaries. During these heatwaves, areas in the top decile tree canopy coverage were, on average, 1.2 °C cooler that those in the lowest decile tree canopy coverage. Differences were most apparent at night, when nighttime minimum temperatures were around 2.0 °C lower in the areas with the most tree canopy, while maximum daytime temperatures were on average 0.8 °C lower. The maximum UHI intensities tended to be in the very early morning hours and averaged 4.3 °C across all heatwave days.

Zoom In Zoom Out Reset image size

Differences in daily mean, minimum, and maximum temperatures for stations with different levels of tree canopy coverage can be seen over the long-term data (figures 5 and 6). These figures show mean, minimum, and maximum anomalies (or differences between the stations and the average of all stations) as temperatures increase. Temperatures are, on average, 0.8 °C higher in deciles with the least tree coverage compared to the most, or around −0.019 °C (−0.0192 to −0.0187) per % increase in tree canopy. The temperature differences between the areas with the greatest and least amount of tree canopy coverage was greatest at lower temperatures (average minimum difference 1.1 °C), and the least at higher temperatures (average maximum differences 0.6 °C).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

As expected, data was noisy and with unexplained variance in the model. The GAM with the 1000 m buffer showed the best performance as judged by chi-squared tests between models (table S1 in the appendix). Negligible amounts of spatial autocorrelation were present in the residuals of this model (an average Morans I of 0.000 41 during summers) compared to the total variation. A cyclical temporal correlation was apparent with peaks during summer; closer examination alongside additional Met Office data indicates that these may be due to exposure to solar radiation. As a result, both spatial and temporal autocorrelation were decided to be unimportant, but we acknowledge that errors may be underestimated because daily samples may not be independent. The final model was able to predict daily temperatures with an R2 of 98.3% and a root mean square error (RMSE) of 0.808 for stations and 0.761 for observations on held-out data. Partial dependence plots can be seen in the appendix.

Health impact calculation results can be seen in figures 7 and 8. Between 2015–2022, we estimate that the current tree coverage avoided 153 heat attributable deaths. This represents a 16% reduction of the UHI-related mortality that would have occurred under the 'grey scenario'. Relative to the current tree canopy coverage, increasing tree coverage to average levels from 2017–2022 would have further reduced UHI mortality by 8%, the London strategy by 10%, and the green scenario by 55%. During the heatwave events of 2022, we estimate that the current tree coverage helped avoid around 16 heat attributable deaths (14% of UHI-associated deaths). This would increase to 23 (21%), 25 (22%), and 67 (61%) under the average, London strategy, and green scenarios, respectively.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Heat mortality is projected to increase around 8-fold by 2061–2080 relative to 1981–2000 as a result of the rising temperatures and the more frequent exceedance of ${T_h}$ under RCP8.5 (figure 8). By 2061–2080, we estimate that the current tree coverage would help avoid around 42 heat-attributable deaths a year. Increased coverage from the London strategy would help avoid an additional 23 deaths a year relative to current levels, and the green scenario 131 deaths a year. The changes also lead to substantial benefits for carbon storage and sequestration, with an additional 1.1 million tonnes of carbon stored and 36 500 tonnes sequestered annually under the London strategy, rising to 5.6 million tonnes stored and 181 000 tonnes sequestered annually under the green scenario.

A strength of this study is the large amounts of temperature data with dense coverage. A tradeoff, however, is in the accuracy of the stations compared to official meteorological stations. To minimize errors, we use data cleaning methods to remove outliers, stations with unreliable data, and indoor stations, while all PWS stations are from the same manufacturer and are assumed to have similar instrumental errors. Five official stations are used to supplement coverage, meaning a relatively small number of different instruments are included in the modelling dataset, but the data from these will be more with accurate than those from PWS. Our reference for UHI intensity are stations outside the GLA located in rural areas, and acknowledge that the trends and amplitude of the UHI we report are sensitive to these reference temperatures [52].

By using quantile regression to identify the maximum amount of local tree coverage according to building density, we aimed to estimate a maximal amount of tree canopy that can be added, while the London Strategy and average scenarios present more achievable targets; these scenarios are an advance over prior studies that offered a percent increase in tree coverage unsupported by evidence or policy. Temperatures will also depend on LCZs, but due to the correlation of the natural component with tree coverage we opted to derive from a buildings dataset to reflect the built component. However, tree canopy coverage and buildings are just one factor which contributes to urban temperature differences and may be confounded by other parameters which we have not included here. The GAM model does not include prevailing wind conditions, nor any advective cooling from the river Thames or the sea. The use of hexagons with a relatively small area helps to separate large areas of tree canopy (such as parks) from areas with residential populations. The modelled temperature changes are associative and cross-sectional and do not demonstrate causality, although this is equally a problem with all other studies of this type, e.g [16].

The tree canopy data offers high resolution coverage derived using machine learning processed aerial imagery. The report on the generation of the dataset acknowledges various limitations. These include the data coming from single time point (September 2016) and a time of the year when leaf structure may be breaking down. The model occasionally misclassifies trees as bushes and can miss sections of canopy altogether. However, the model has an accuracy rate of 94% and produces estimates that are at minimum comparable to estimates from traditional survey methods when aggregated across larger areas. Estimates of the stored carbon within London's trees is based on figures from a 2015 report [42], and may in fact underestimate the amount of carbon stored [53].

The heat-mortality estimates use a relationship between outdoor air temperature and mortality derived for 1996–2013. This model does not include relative humidity or air pollution as confounders, although previous research has not found these to be significant factors for heat mortality in the UK [9, 43]. This heat-mortality relationship uses daily mean temperature, however, there is evidence that nighttime temperatures are particularly important for health outcomes [54]. Given that the greatest effect of tree coverage was seen at night, this could mean that the health benefits of the tree canopy are underestimated. We have not looked at cold. For the future projections, we do not model population social or technological adaptation to heat, such as heatwave plans or use of air conditioning, nor do we model population demographic changes such as population aging. The baseline mortality rate for 2020–2022 will include the effects of Covid-19.

The heat-mortality-relationship used is based on epidemiological analysis of mortality data which partially pre-dates the development of many of the UK heatwave public health measures. Our predicted mortality for the heat periods of the summer of 2022 is both higher than other real-time models and the observed excess mortality [8], likely due to different sources of temperature data, as well as different underlying methods in calculating heat-AM. During these heatwaves, the UK Met Office issued heat-health alert and red EH weather warnings to the public for the first time. It is possible that the warnings and advice from the Met Office and Public Health agencies had a role in reducing the actual mortality below what was expected. The estimated mortality in this study should therefore be interpreted as a theoretical quantification of the role of trees rather than an absolute estimate of heat-related mortality during this period.

The methods used here can be applied to other locations and with different land covers wherever PWS data are available. While the results of this study can provide an indicative value for similar cities, caution should be applied when generalizing as city-specific characteristics like climate, the amount of potential greening, and the built environment will influence estimates and a recent PWS study shows that increased tree cover in many European cities does not lead to reduced air temperatures [15]. The types of trees, such as their leaf area index, and the albedo of the built environment are also important considerations [55].

In addition to benefits for reducing heat exposure, trees offer a number of benefits, including biodiversity, stormwater management, air pollution removal, and carbon storage and sequestration [42]. Green spaces have been shown to have a positive influence on physical activity, stress, social contacts, and restoration [56], and increased urban street tree density is associated with reduced mental health issues [57]. In the UK, all-cause mortality was found to be 6% lower in the quintile with the most greenspace compared to the quintile with the lowest [58].

Trees also have some disadvantages. Under certain circumstances, trees can risk negatively impacting air quality by preventing dispersion of polluted air in narrow streets for certain wind directions [59, 60], while the pollen produced by certain tree species can exacerbate allergies [61]. Our results show areas with more trees are colder at lower temperatures, with possible increased in space heating energy demand [62] if the necessary energy efficient retrofits are not carried out. Types of trees are likely to be important, with deciduous trees offering beneficial shading during the summer. Future research could compare the estimated costs associated with additional tree coverage with the potential savings provided by reduced health care cost and examine the impacts on indoor temperatures and heating or cooling demand.