The mortality impacts of greening Italy Authors: Orazio Valerio Giannico, Rodolfo Sardone, Lucia Bisceglia, Francesco Addabbo, Francesco Pirotti, Sante Minerba, Antonia Mincuzzi Published: 2024 ABSTRACT Green spaces have been shown to be beneficial to humans, but quantifying these benefits is a challenge for epidemiology. In this health impact assessment study, we exploit satellites to estimate for the whole of Italy the number of deaths that could be prevented in the 49 million adult population by greening residential areas. The exposure was assessed by calculating the normalized difference vegetation index at 10-m resolution within a 300-m distance from homes in 7904 municipalities. In this study we estimate, by achieving nationwide the level of residential greenness currently reached by the 25% of the population, a total of 28,433 (95% confidence interval: 21,400–42,350) preventable deaths and 279,324 (210,247–415,980) preventable years of life lost in Italy in 2022, representing the 5% of the total mortality burden. More green means fewer deaths, thus strong action is needed to increase the amount and accessibility of green spaces in all human settlements. Introduction Green spaces, such as parks, playgrounds, and residential greenery, have been shown to have beneficial effects on people. They can promote mental and physical health and reduce morbidity and mortality among citizens1–12. The normalized difference vegetation index (NDVI) is a widely used satellite-derived metric that can quantify the vegetation in human settlements (e.g., street trees or general vegetation in public and private spaces)11,13. Several studies have highlighted the inverse relationship between NDVI and mortality2–10. Based on this evidence, a meta-analysis has estimated that the pooled relative risk (RR) for all-cause mortality per 0.1 increase in NDVI within a buffer of 500 m or less from a participant’s home was 0.96 (95% CI 0.94–0.97)11. Several hypotheses have been considered to explain these findings. It has been proposed that physical activity is an important health determinant associated with green spaces. Green spaces can be places for recreational physical activity. In addition, green areas may encourage walking and cycling as forms of active transport1,7,11,12,14,15. However, a mediation analysis of one of the studies4 included in the meta-analysis showed that physical activity accounted for 2.1% (95% CI 0.2;19.3%) of the association between green space and mortality11. Greenness can also have a number of beneficial health effects through ecosystem services, and has been shown to have a protective effect by reducing air pollution, noise, and the heat island effect11,12,16–18. Air pollution was included as a covariate in some of the studies2,3,6 included in the meta-analysis, but the resulting RRs are not significantly different from those of studies that did not consider air pollution. After including air pollution in a mediation analysis, James and colleagues4 found that PM2.5 (fine particulate matter) could explain 4.4% (95% CI 2.4;7.7%) of the association between greenness and mortality, while Vienneu et colleagues6 estimated a mediation of 2.4% (−0.2;5.5%) by PM1011. Other proposed mechanisms to explain the health benefits of green spaces include stress reduction, enhanced relaxation, and restoration. One theory that explains the benefits of being in green spaces is the psychosomatic stress reduction theory. According to this idea, exposure to natural environments, such as views of them, could help people who are stressed out by putting them in a more positive emotional state1,19,20. In another study conducted in Dutch cities, the relationship between urban greenery and perceived general health was shown to be most strongly mediated by stress and social cohesion21. Finally, there is evidence linking green spaces to immunological function22. Li and colleagues23,24 found an association between forest visits and enhanced immune responses, including the expression of anti-cancer proteins (such as granzymes A and B, granulysin, and perforin). It has also been suggested that exposure to a variety of microbes in their natural habitats may regulate immunity25. In terms of policy, the World Health Organization (WHO) suggests that a minimum of 0.5 hectares of green space should be available within a 300-m linear distance from homes1. The 3–30–300 rule for urban greening was proposed by urban forester Cecil Konijnendijk. It states that every home, school, and workplace should have at least three well-established trees in sight; every neighbourhood should have at least 30% tree cover; and every residence should be no more than 300 m from the nearest large public green space26. One way to link epidemiological evidence to public health policy is through health impact assessment (HIA), which combines meta-analytic effect measures with exposure and outcome data to estimate the health impacts under a counterfactual scenario11,27–43. Using the exposure-response function from Rojas-Rueda et colleagues11, two large-scale health impact assessment studies were conducted in urban areas in the United States and in Europe to estimate the impacts of greenness on mortality27,28. According to the American HIA study27, an increase of 0.1 NDVI units at the census tract level in the 35 most populous metropolitan areas could have prevented an estimated 38,187 deaths of people aged 65 and older in 2019. According to the European HIA study28, which analysed 978 selected cities and 49 greater cities using 250-m resolution NDVI data, achieving a greenness level estimated to be in line with WHO recommendations1 for access to green space could have reduced deaths by 42,968 in 2015. However, the health benefits of exposure to greenness are likely to occur in all human settlements, not just selected cities or the more urbanised areas4,7,11. In addition, the use of different counterfactuals for each city (0.1 NDVI increase for all cities in the American study27 and city-specific modelled targets in the European study28) may limit the comparability of the areas evaluated. Finally, there is uncertainty in the choice of target exposures due to the lack of specific recommendations for NDVI values1,27,28. Therefore, it may be worth exploring different methodological strategies and benefiting from updated and high-resolution data to assess the health impacts of greening interventions in all human settlements. To ensure comparability, it may be desirable to conduct these assessments using a single counterfactual exposure for all areas. Ideally, this NDVI target exposure should be as realistically achievable as possible, population-based, and possibly related to the existing green space recommendations. Finally, it would be desirable for this assessment to be replicable as far as possible around the world, using available software and data. In the present study we want to meet all these needs by capitalising on satellite-derived residential exposure data at 10-m resolution for the whole of Italy in 2022. The aim of this health impact assessment is to estimate the total adult mortality burden that could be prevented in all Italian municipalities by greening residential areas up to the level of greenness currently achieved by the 25% of the population. Results Population and exposure The total adult population (≥20 years) in all 7,904 Italian municipalities in 2022 is 48,628,328. All exposures to greenest period NDVI (2022) and to specific land cover classes (2021) at 10 m resolution were calculated within 300 m of homes (population-buffers). Maps of greenest period NDVI, population-weighted exposure (PWE) to greenest period NDVI, and population counts are shown in Figs. 1 and 2 and Supplementary Fig. 1, respectively. Population-weighted percentiles of exposures to greenest period NDVI, exposures to land cover classes, and adult population counts are reported in Supplementary Table 1. The 25th, 50th, 75th and 95th population-weighted percentiles of PWEs to greenest period NDVI in all the Italian municipalities are 0.30, 0.36, 0.43 and 0.54, respectively. The 25th, 50th, 75th and 95th population-weighted percentiles of greenest period NDVI in all the Italian 300 m population buffers are 0.27, 0.36, 0.46 and 0.60, respectively. The unweighted percentiles of these measures and the weighted percentiles of these measures when excluding sparsely populated areas in the exposure assessment are reported in Supplementary Tables 2 and 3, respectively.Fig. 1Greenest period NDVI from Sentinel-2. Mean per pixel. Italy with a 5 km buffer around the boundaries, including water bodies, April-June 2022.World Mollweide (ESRI:54009). NDVI: normalized difference vegetation index. Map created with QGIS version 3.28.4. NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’).Fig. 2Greenest period NDVI from Sentinel-2. PWE per municipality. Italy, April-June 2022.WGS 84 / UTM zone 32N (EPSG:32632). NDVI: normalized difference vegetation index. PWE: population-weighted exposure. Map created with QGIS version 3.28.4. Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data available are at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). Greenest period NDVI from Sentinel-2. Mean per pixel. Italy with a 5 km buffer around the boundaries, including water bodies, April-June 2022. World Mollweide (ESRI:54009). NDVI: normalized difference vegetation index. Map created with QGIS version 3.28.4. NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). Greenest period NDVI from Sentinel-2. PWE per municipality. Italy, April-June 2022. WGS 84 / UTM zone 32N (EPSG:32632). NDVI: normalized difference vegetation index. PWE: population-weighted exposure. Map created with QGIS version 3.28.4. Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data available are at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). The counterfactual exposure of the health impact assessment was set at 0.46, which is the 75th population-weighted percentile of the greenest period NDVI in all the 300-m radius population-buffers. This is the target level of greenness which is currently achieved by the 25% of the Italian population. In the study area, a PWE to greenest period NDVI of 0.46 within 300 m from the residence approximately corresponds to a PWE to tree cover proportion of 30% within the same distance. Health impact assessment The results of the HIA are reported in Tables 1–3 and in Supplementary Figs. 2 and 3 for all the 39,803,860 adult inhabitants of the 3720 Italian municipalities that have a PWE to greenest period NDVI below the counterfactual exposure of 0.46 in 2022. Considering the whole of Italy, the PWEs to tree cover proportion, green area proportion and greenest period NDVI are 21.2%, 32.3% and 0.33, respectively. Table 1Population and exposure to greenness in the municipalities included in the HIA when using as counterfactual exposure (k) the 75th population-weighted percentile of greenest period NDVI in all the 300 m population-buffers. Italy, 2021 (land cover) and 2022 (NDVI and population)N = 3720 municipalities with a PWE to greenest period NDVI < 75th percentile (k = 0.46)PopulationExposure to greennessNumberPWE≥20 yearsThree cover proportion [%]Green area proportion [%]Greenest period NDVIDifference between k and the greenest period NDVIItaly (n = 3720)39,803,86021.232.30.330.13Provincial capitals (n = 105)14,315,14622.529.10.310.15<45,000 inhabitants (n = 27)900,42128.538.30.380.0845–75,000 inhabitants (n = 25)1,425,79524.333.80.330.1375–120,000 inhabitants (n = 26)2,356,21720.629.70.330.13≥120,000 inhabitants (n = 27)9,632,71322.127.50.300.16Non provincial capitals (n = 3615)25,488,71420.534.10.350.11<1000 inhabitants (n = 454)286,15025.645.00.410.051–2000 inhabitants (n = 592)869,36124.442.90.400.062–5000 inhabitants (n = 1049)3,469,67222.039.40.380.08≥5000 inhabitants (n = 1520)20,863,53120.032.70.340.12Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). Land cover data from WorldCover are available at: https://code.earthengine.google.com (European Space Agency. WorldCover Project. 2021. Image collection: ‘ESA/WorldCover/v200’).HIA health impact assessment, NDVI normalized difference vegetation index, PWE population-weighted exposure.Table 2Estimated preventable deaths in the municipalities included in the HIA when using as counterfactual exposure (k) the 75th population-weighted percentile of greenest period NDVI in all the 300 m population-buffers. Italy, 2022N = 3720 municipalities with a PWE to greenest period NDVI < 75th percentile (k = 0.46)Preventable deathsNumber (95% CI)Rate (95% CI) [per 100,000 inhabitants]Rate fraction (95% CI) [%]20–39 years40–59 years60–79 years≥80 years≥20 years20–39 years40–59 years60–79 years≥80 years≥20 years20–39 years40–59 years60–79 years≥80 years≥20 yearsItaly (n = 3720)216 (162–321)1671 (1258–2487)8148 (6133–12,131)18,400 (13,847–27,411)28,433 (21,400–42,350)2 (2–3)11 (9–17)73 (55–108)504 (379–750)71 (54–106)5.2 (3.9–7.8)5.2 (3.9–7.7)5.1 (3.9–7.6)4.9 (3.7–7.3)5.0 (3.7–7.4)Provincial capitals (n = 105)84 (63–125)671 (505–999)3413 (2569–5079)8210 (6179–12,226)12,377 (9317–18,428)2 (2–3)13 (10–19)85 (64–126)572 (430–851)86 (65–129)6.0 (4.5–8.9)6.0 (4.5–8.9)6.0 (4.5–8.9)5.7 (4.3–8.5)5.8 (4.4–8.6)<45,000 inhabitants (n = 27)3 (3–5)24 (18–37)128 (96–191)333 (250–499)488 (366–732)2 (1–2)8 (6–12)47 (35–71)348 (261–521)54 (41–81)3.5 (2.6–5.2)3.4 (2.6–5.1)3.4 (2.5–5.1)3.3 (2.5–5.0)3.3 (2.5–5.0)45–75,000 inhabitants (n = 25)8 (6–12)59 (45–88)309 (233–460)714 (538–1063)1091 (821–1623)2 (2–3)12 (9–17)74 (56–111)517 (389–770)76 (58–114)5.1 (3.8–7.5)5.3 (4.0–7.8)5.3 (4.0–7.8)5.0 (3.7–7.4)5.1 (3.8–7.5)75–120,000 inhabitants (n = 26)14 (11–21)102 (77–152)494 (372–735)1202 (904–1791)1812 (1364–2699)2 (2–4)12 (9–18)73 (55–109)506 (381–754)77 (58–115)5.7 (4.3–8.5)5.5 (4.1–8.2)5.3 (4.0–7.9)5.0 (3.8–7.4)5.1 (3.8–7.6)≥120,000 inhabitants (n = 27)58 (44–87)485 (365–721)2482 (1869–3692)5961 (4488–8873)8986 (6766–13,373)2 (2–4)14 (10–21)93 (70–138)618 (465–919)93 (70–139)6.5 (4.9–9.6)6.5 (4.9–9.6)6.5 (4.9–9.7)6.3 (4.7–9.3)6.3 (4.8–9.4)Non provincial capitals (n = 3615)132 (99–196)999 (752–1488)4735 (3564–7053)10,190 (7667–15,186)16,056 (12,083–23,923)2 (2–3)11 (8–16)66 (50–98)460 (346–685)63 (47–94)4.8 (3.6–7.2)4.7 (3.6–7.0)4.7 (3.5–6.9)4.4 (3.3–6.6)4.5 (3.4–6.7)<1000 inhabitants (n = 454)1 (1–1)6 (4–9)31 (24–47)79 (59–118)117 (87–175)1 (1–2)6 (4–9)35 (26–53)250 (188–376)41 (30–61)2.4 (1.8–3.6)2.3 (1.7–3.4)2.2 (1.6–3.3)2.1 (1.6–3.2)2.1 (1.6–3.2)1–2000 inhabitants (n = 592)2 (2–4)20 (15–30)100 (75–150)262 (196–393)385 (289–577)1 (1–2)6 (5–10)39 (29–58)302 (227–454)44 (33–66)2.6 (1.9–3.9)2.5 (1.9–3.8)2.6 (2.0–3.9)2.6 (2.0–3.9)2.6 (2.0–3.9)2–5000 inhabitants (n = 1049)13 (10–20)98 (73–146)469 (352–702)1089 (818–1629)1669 (1254–2496)2 (1–2)8 (6–12)47 (35–70)347 (260–518)48 (36–72)3.4 (2.6–5.1)3.4 (2.5–5.0)3.3 (2.5–4.9)3.2 (2.4–4.8)3.3 (2.4–4.9)≥5000 inhabitants (n = 1520)115 (87–172)876 (659–1303)4134 (3113–6153)8760 (6594–13,046)13,886 (10,453–20,675)2 (2–3)11 (8–17)71 (54–106)491 (370–731)67 (50–99)5.2 (3.9–7.8)5.1 (3.8–7.6)5.0 (3.8–7.5)4.8 (3.6–7.1)4.9 (3.7–7.2)The exposure-response function per increment of 0.1 NDVI is a RR for all-cause mortality of 0.96 (95% CI 0.94–0.97)11. Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’).CI confidence interval, HIA health impact assessment, NDVI normalized difference vegetation index, PWE population-weighted exposure, RR relative risk.Table 3Estimated preventable years of life lost in the municipalities included in the HIA when using as counterfactual exposure (k) the 75th population-weighted percentile of greenest period NDVI in all the 300 m population-buffers. Italy, 2022N = 3720 municipalities with a PWE to greenest period NDVI < 75th percentile (k = 0.46)Preventable years of life lostNumber (95% CI)Rate (95% CI) [per 100,000 inhabitants]Rate fraction (95% CI) [%]20–39 years40–59 years60–79 years≥80 years≥20 years20–39 years40–59 years60–79 years≥80 years≥20 years20–39 years40–59 years60–79 years≥80 years≥20 yearsItaly (n = 3720)11,103 (8359–16,526)52,266 (39,345–77,819)121,903 (91,762–181,518)94,052 (70,780–140,116)279,324 (210,247–415,980)107 (81–160)357 (269–532)1090 (820–1623)2575 (1938–3836)702 (528–1045)5.2 (3.9–7.7)5.1 (3.9–7.7)5.1 (3.8–7.6)4.9 (3.7–7.3)5.0 (3.8–7.5)Provincial capitals (n = 105)4322 (3254–6432)21,042 (15,844–31,316)51,132 (38,498–76,102)41,761 (31,434–62,190)118,257 (89,030–176,040)118 (89–175)407 (306–605)1269 (956–1889)2907 (2188–4329)826 (622–1230)5.9 (4.5–8.8)6.0 (4.5–8.9)5.9 (4.5–8.9)5.7 (4.3–8.5)5.9 (4.4–8.7)<45,000 inhabitants (n = 27)180 (135–270)764 (573–1145)1901 (1426–2849)1643 (1232–2463)4487 (3366–6727)82 (62–124)242 (182–363)702 (527–1053)1716 (1287–2572)498 (374–747)3.5 (2.6–5.2)3.4 (2.5–5.1)3.4 (2.5–5.1)3.3 (2.5–5.0)3.4 (2.5–5.1)45–75,000 inhabitants (n = 25)405 (305–603)1847 (1391–2749)4571 (3442–6802)3557 (2678–5295)10,380 (7816–15,448)110 (83–165)365 (275–543)1101 (829–1638)2576 (1940–3835)728 (548–1084)5.0 (3.8–7.5)5.2 (3.9–7.8)5.2 (3.9–7.8)5.0 (3.8–7.4)5.1 (3.9–7.6)75–120,000 inhabitants (n = 26)726 (547–1080)3212 (2418–4780)7354 (5536–10,951)6062 (4562–9034)17,355 (13,063–25,845)121 (91–180)379 (286–565)1094 (824–1629)2553 (1921–3804)736 (554–1097)5.7 (4.3–8.5)5.5 (4.1–8.2)5.3 (4.0–7.8)5.0 (3.8–7.5)5.2 (3.9–7.8)≥120,000 inhabitants (n = 27)3011 (2267–4479)15,219 (11,461–22,643)37,305 (28,094–55,500)30,500 (22,963–45,398)86,035 (64,785–128,020)121 (91–180)434 (327–646)1397 (1052–2078)3160 (2379–4703)893 (673–1329)6.5 (4.9–9.6)6.4 (4.9–9.6)6.5 (4.9–9.6)6.3 (4.7–9.3)6.4 (4.8–9.5)Non provincial capitals (n = 3615)6781 (5105–10,094)31,224 (23,501–46,503)70,771 (53,265–105,417)52,290 (39,346–77,927)161,066 (121,217–239,940)102 (77–151)330 (249–492)989 (744–1473)2359 (1775–3516)632 (476–941)4.8 (3.6–7.2)4.7 (3.5–7.0)4.6 (3.5–6.9)4.4 (3.3–6.6)4.6 (3.4–6.8)<1000 inhabitants (n = 454)43 (32–65)177 (133–266)474 (355–713)382 (287–575)1077 (807–1619)66 (49–99)178 (134–268)530 (397–797)1220 (914–1833)376 (282–566)2.4 (1.8–3.6)2.3 (1.7–3.4)2.2 (1.6–3.3)2.1 (1.6–3.2)2.2 (1.6–3.3)1–2000 inhabitants (n = 592)126 (95–189)622 (466–932)1510 (1133–2264)1309 (982–1963)3567 (2676–5349)59 (44–88)201 (151–301)583 (437–874)1512 (1134–2267)410 (308–615)2.6 (1.9–3.9)2.5 (1.9–3.8)2.6 (1.9–3.9)2.6 (2.0–3.9)2.6 (1.9–3.9)2–5000 inhabitants (n = 1049)673 (506–1006)3056 (2296–4570)7076 (5315–10,582)5510 (4139–8240)16,315 (12,255–24,398)76 (57–114)240 (180–359)708 (532–1060)1754 (1317–2622)470 (353–703)3.3 (2.5–5.0)3.4 (2.5–5.0)3.3 (2.4–4.9)3.2 (2.4–4.8)3.3 (2.5–4.9)≥5000 inhabitants (n = 1520)5938 (4472–8833)27,369 (20,607–40,735)61,712 (46,462–91,858)45,089 (33,939–67,149)140,108 (105,479–208,574)108 (81–160)352 (265–525)1062 (800–1581)2527 (1902–3763)672 (506–1000)5.2 (3.9–7.8)5.1 (3.8–7.5)5.0 (3.8–7.4)4.7 (3.6–7.1)4.9 (3.7–7.3)The exposure-response function per increment of 0.1 NDVI is a RR for all-cause mortality of 0.96 (95% CI 0.94–0.97)11. Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’).CI confidence interval, HIA health impact assessment, NDVI normalized difference vegetation index, PWE population-weighted exposure, RR relative risk. Table 1Population and exposure to greenness in the municipalities included in the HIA when using as counterfactual exposure (k) the 75th population-weighted percentile of greenest period NDVI in all the 300 m population-buffers. Italy, 2021 (land cover) and 2022 (NDVI and population)N = 3720 municipalities with a PWE to greenest period NDVI < 75th percentile (k = 0.46)PopulationExposure to greennessNumberPWE≥20 yearsThree cover proportion [%]Green area proportion [%]Greenest period NDVIDifference between k and the greenest period NDVIItaly (n = 3720)39,803,86021.232.30.330.13Provincial capitals (n = 105)14,315,14622.529.10.310.15<45,000 inhabitants (n = 27)900,42128.538.30.380.0845–75,000 inhabitants (n = 25)1,425,79524.333.80.330.1375–120,000 inhabitants (n = 26)2,356,21720.629.70.330.13≥120,000 inhabitants (n = 27)9,632,71322.127.50.300.16Non provincial capitals (n = 3615)25,488,71420.534.10.350.11<1000 inhabitants (n = 454)286,15025.645.00.410.051–2000 inhabitants (n = 592)869,36124.442.90.400.062–5000 inhabitants (n = 1049)3,469,67222.039.40.380.08≥5000 inhabitants (n = 1520)20,863,53120.032.70.340.12Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). Land cover data from WorldCover are available at: https://code.earthengine.google.com (European Space Agency. WorldCover Project. 2021. Image collection: ‘ESA/WorldCover/v200’).HIA health impact assessment, NDVI normalized difference vegetation index, PWE population-weighted exposure. Population and exposure to greenness in the municipalities included in the HIA when using as counterfactual exposure (k) the 75th population-weighted percentile of greenest period NDVI in all the 300 m population-buffers. Italy, 2021 (land cover) and 2022 (NDVI and population) Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). Land cover data from WorldCover are available at: https://code.earthengine.google.com (European Space Agency. WorldCover Project. 2021. Image collection: ‘ESA/WorldCover/v200’). HIA health impact assessment, NDVI normalized difference vegetation index, PWE population-weighted exposure. Table 2Estimated preventable deaths in the municipalities included in the HIA when using as counterfactual exposure (k) the 75th population-weighted percentile of greenest period NDVI in all the 300 m population-buffers. Italy, 2022N = 3720 municipalities with a PWE to greenest period NDVI < 75th percentile (k = 0.46)Preventable deathsNumber (95% CI)Rate (95% CI) [per 100,000 inhabitants]Rate fraction (95% CI) [%]20–39 years40–59 years60–79 years≥80 years≥20 years20–39 years40–59 years60–79 years≥80 years≥20 years20–39 years40–59 years60–79 years≥80 years≥20 yearsItaly (n = 3720)216 (162–321)1671 (1258–2487)8148 (6133–12,131)18,400 (13,847–27,411)28,433 (21,400–42,350)2 (2–3)11 (9–17)73 (55–108)504 (379–750)71 (54–106)5.2 (3.9–7.8)5.2 (3.9–7.7)5.1 (3.9–7.6)4.9 (3.7–7.3)5.0 (3.7–7.4)Provincial capitals (n = 105)84 (63–125)671 (505–999)3413 (2569–5079)8210 (6179–12,226)12,377 (9317–18,428)2 (2–3)13 (10–19)85 (64–126)572 (430–851)86 (65–129)6.0 (4.5–8.9)6.0 (4.5–8.9)6.0 (4.5–8.9)5.7 (4.3–8.5)5.8 (4.4–8.6)<45,000 inhabitants (n = 27)3 (3–5)24 (18–37)128 (96–191)333 (250–499)488 (366–732)2 (1–2)8 (6–12)47 (35–71)348 (261–521)54 (41–81)3.5 (2.6–5.2)3.4 (2.6–5.1)3.4 (2.5–5.1)3.3 (2.5–5.0)3.3 (2.5–5.0)45–75,000 inhabitants (n = 25)8 (6–12)59 (45–88)309 (233–460)714 (538–1063)1091 (821–1623)2 (2–3)12 (9–17)74 (56–111)517 (389–770)76 (58–114)5.1 (3.8–7.5)5.3 (4.0–7.8)5.3 (4.0–7.8)5.0 (3.7–7.4)5.1 (3.8–7.5)75–120,000 inhabitants (n = 26)14 (11–21)102 (77–152)494 (372–735)1202 (904–1791)1812 (1364–2699)2 (2–4)12 (9–18)73 (55–109)506 (381–754)77 (58–115)5.7 (4.3–8.5)5.5 (4.1–8.2)5.3 (4.0–7.9)5.0 (3.8–7.4)5.1 (3.8–7.6)≥120,000 inhabitants (n = 27)58 (44–87)485 (365–721)2482 (1869–3692)5961 (4488–8873)8986 (6766–13,373)2 (2–4)14 (10–21)93 (70–138)618 (465–919)93 (70–139)6.5 (4.9–9.6)6.5 (4.9–9.6)6.5 (4.9–9.7)6.3 (4.7–9.3)6.3 (4.8–9.4)Non provincial capitals (n = 3615)132 (99–196)999 (752–1488)4735 (3564–7053)10,190 (7667–15,186)16,056 (12,083–23,923)2 (2–3)11 (8–16)66 (50–98)460 (346–685)63 (47–94)4.8 (3.6–7.2)4.7 (3.6–7.0)4.7 (3.5–6.9)4.4 (3.3–6.6)4.5 (3.4–6.7)<1000 inhabitants (n = 454)1 (1–1)6 (4–9)31 (24–47)79 (59–118)117 (87–175)1 (1–2)6 (4–9)35 (26–53)250 (188–376)41 (30–61)2.4 (1.8–3.6)2.3 (1.7–3.4)2.2 (1.6–3.3)2.1 (1.6–3.2)2.1 (1.6–3.2)1–2000 inhabitants (n = 592)2 (2–4)20 (15–30)100 (75–150)262 (196–393)385 (289–577)1 (1–2)6 (5–10)39 (29–58)302 (227–454)44 (33–66)2.6 (1.9–3.9)2.5 (1.9–3.8)2.6 (2.0–3.9)2.6 (2.0–3.9)2.6 (2.0–3.9)2–5000 inhabitants (n = 1049)13 (10–20)98 (73–146)469 (352–702)1089 (818–1629)1669 (1254–2496)2 (1–2)8 (6–12)47 (35–70)347 (260–518)48 (36–72)3.4 (2.6–5.1)3.4 (2.5–5.0)3.3 (2.5–4.9)3.2 (2.4–4.8)3.3 (2.4–4.9)≥5000 inhabitants (n = 1520)115 (87–172)876 (659–1303)4134 (3113–6153)8760 (6594–13,046)13,886 (10,453–20,675)2 (2–3)11 (8–17)71 (54–106)491 (370–731)67 (50–99)5.2 (3.9–7.8)5.1 (3.8–7.6)5.0 (3.8–7.5)4.8 (3.6–7.1)4.9 (3.7–7.2)The exposure-response function per increment of 0.1 NDVI is a RR for all-cause mortality of 0.96 (95% CI 0.94–0.97)11. Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’).CI confidence interval, HIA health impact assessment, NDVI normalized difference vegetation index, PWE population-weighted exposure, RR relative risk. Estimated preventable deaths in the municipalities included in the HIA when using as counterfactual exposure (k) the 75th population-weighted percentile of greenest period NDVI in all the 300 m population-buffers. Italy, 2022 The exposure-response function per increment of 0.1 NDVI is a RR for all-cause mortality of 0.96 (95% CI 0.94–0.97)11. Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). CI confidence interval, HIA health impact assessment, NDVI normalized difference vegetation index, PWE population-weighted exposure, RR relative risk. Table 3Estimated preventable years of life lost in the municipalities included in the HIA when using as counterfactual exposure (k) the 75th population-weighted percentile of greenest period NDVI in all the 300 m population-buffers. Italy, 2022N = 3720 municipalities with a PWE to greenest period NDVI < 75th percentile (k = 0.46)Preventable years of life lostNumber (95% CI)Rate (95% CI) [per 100,000 inhabitants]Rate fraction (95% CI) [%]20–39 years40–59 years60–79 years≥80 years≥20 years20–39 years40–59 years60–79 years≥80 years≥20 years20–39 years40–59 years60–79 years≥80 years≥20 yearsItaly (n = 3720)11,103 (8359–16,526)52,266 (39,345–77,819)121,903 (91,762–181,518)94,052 (70,780–140,116)279,324 (210,247–415,980)107 (81–160)357 (269–532)1090 (820–1623)2575 (1938–3836)702 (528–1045)5.2 (3.9–7.7)5.1 (3.9–7.7)5.1 (3.8–7.6)4.9 (3.7–7.3)5.0 (3.8–7.5)Provincial capitals (n = 105)4322 (3254–6432)21,042 (15,844–31,316)51,132 (38,498–76,102)41,761 (31,434–62,190)118,257 (89,030–176,040)118 (89–175)407 (306–605)1269 (956–1889)2907 (2188–4329)826 (622–1230)5.9 (4.5–8.8)6.0 (4.5–8.9)5.9 (4.5–8.9)5.7 (4.3–8.5)5.9 (4.4–8.7)<45,000 inhabitants (n = 27)180 (135–270)764 (573–1145)1901 (1426–2849)1643 (1232–2463)4487 (3366–6727)82 (62–124)242 (182–363)702 (527–1053)1716 (1287–2572)498 (374–747)3.5 (2.6–5.2)3.4 (2.5–5.1)3.4 (2.5–5.1)3.3 (2.5–5.0)3.4 (2.5–5.1)45–75,000 inhabitants (n = 25)405 (305–603)1847 (1391–2749)4571 (3442–6802)3557 (2678–5295)10,380 (7816–15,448)110 (83–165)365 (275–543)1101 (829–1638)2576 (1940–3835)728 (548–1084)5.0 (3.8–7.5)5.2 (3.9–7.8)5.2 (3.9–7.8)5.0 (3.8–7.4)5.1 (3.9–7.6)75–120,000 inhabitants (n = 26)726 (547–1080)3212 (2418–4780)7354 (5536–10,951)6062 (4562–9034)17,355 (13,063–25,845)121 (91–180)379 (286–565)1094 (824–1629)2553 (1921–3804)736 (554–1097)5.7 (4.3–8.5)5.5 (4.1–8.2)5.3 (4.0–7.8)5.0 (3.8–7.5)5.2 (3.9–7.8)≥120,000 inhabitants (n = 27)3011 (2267–4479)15,219 (11,461–22,643)37,305 (28,094–55,500)30,500 (22,963–45,398)86,035 (64,785–128,020)121 (91–180)434 (327–646)1397 (1052–2078)3160 (2379–4703)893 (673–1329)6.5 (4.9–9.6)6.4 (4.9–9.6)6.5 (4.9–9.6)6.3 (4.7–9.3)6.4 (4.8–9.5)Non provincial capitals (n = 3615)6781 (5105–10,094)31,224 (23,501–46,503)70,771 (53,265–105,417)52,290 (39,346–77,927)161,066 (121,217–239,940)102 (77–151)330 (249–492)989 (744–1473)2359 (1775–3516)632 (476–941)4.8 (3.6–7.2)4.7 (3.5–7.0)4.6 (3.5–6.9)4.4 (3.3–6.6)4.6 (3.4–6.8)<1000 inhabitants (n = 454)43 (32–65)177 (133–266)474 (355–713)382 (287–575)1077 (807–1619)66 (49–99)178 (134–268)530 (397–797)1220 (914–1833)376 (282–566)2.4 (1.8–3.6)2.3 (1.7–3.4)2.2 (1.6–3.3)2.1 (1.6–3.2)2.2 (1.6–3.3)1–2000 inhabitants (n = 592)126 (95–189)622 (466–932)1510 (1133–2264)1309 (982–1963)3567 (2676–5349)59 (44–88)201 (151–301)583 (437–874)1512 (1134–2267)410 (308–615)2.6 (1.9–3.9)2.5 (1.9–3.8)2.6 (1.9–3.9)2.6 (2.0–3.9)2.6 (1.9–3.9)2–5000 inhabitants (n = 1049)673 (506–1006)3056 (2296–4570)7076 (5315–10,582)5510 (4139–8240)16,315 (12,255–24,398)76 (57–114)240 (180–359)708 (532–1060)1754 (1317–2622)470 (353–703)3.3 (2.5–5.0)3.4 (2.5–5.0)3.3 (2.4–4.9)3.2 (2.4–4.8)3.3 (2.5–4.9)≥5000 inhabitants (n = 1520)5938 (4472–8833)27,369 (20,607–40,735)61,712 (46,462–91,858)45,089 (33,939–67,149)140,108 (105,479–208,574)108 (81–160)352 (265–525)1062 (800–1581)2527 (1902–3763)672 (506–1000)5.2 (3.9–7.8)5.1 (3.8–7.5)5.0 (3.8–7.4)4.7 (3.6–7.1)4.9 (3.7–7.3)The exposure-response function per increment of 0.1 NDVI is a RR for all-cause mortality of 0.96 (95% CI 0.94–0.97)11. Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’).CI confidence interval, HIA health impact assessment, NDVI normalized difference vegetation index, PWE population-weighted exposure, RR relative risk. Estimated preventable years of life lost in the municipalities included in the HIA when using as counterfactual exposure (k) the 75th population-weighted percentile of greenest period NDVI in all the 300 m population-buffers. Italy, 2022 The exposure-response function per increment of 0.1 NDVI is a RR for all-cause mortality of 0.96 (95% CI 0.94–0.97)11. Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). CI confidence interval, HIA health impact assessment, NDVI normalized difference vegetation index, PWE population-weighted exposure, RR relative risk. A total of 28,433 (95% CI 21,400–42,350) annual preventable deaths, 71 (95% CI 54–106) annual preventable death rate (per 100,000 inhabitants) and 5.0% (95% CI 3.7–7.4%) annual preventable death rate fraction were estimated for Italy in 2022. The estimated number of preventable deaths per year is 12,377 (95% CI 9,317–18,428) in provincial capitals and 16,056 (95% CI 12,083–23,923) in non provincial capitals. The higher preventable death rate was estimated for provincial capitals (86 deaths per 100,000) and for provincial capitals with a population of 120,000 or more (93 deaths per 100,000). The higher preventable death rate fraction was estimated for provincial capitals (5.8% of deaths) and for provincial capitals with a population of 120,000 or more (6.3% of deaths). Considering all municipalities, the higher preventable death rate was estimated for the population aged ≥80 years (504 deaths per 100,000) and the higher preventable death rate fraction for the population aged <80 years (5.1–5.2% of deaths). A total of 279,324 (95% CI 210,247–415,980) annual preventable years of life lost (YLL), 702 (95% CI 528–1,045) annual preventable YLL rate (per 100,000 inhabitants) and 5.0% (95% CI 3.8–7.5%) annual preventable YLL rate fraction were estimated for Italy in 2022. The estimated number of preventable YLL per year is 118,257 (95% CI 89,030–176,040) in provincial capitals and 161,066 (95% CI 121,217–239,940) in non provincial capitals. The higher preventable YLL rate was estimated for provincial capitals (826 YLL per 100,000) and for provincial capitals with a population of 120,000 or more (893 YLL per 100,000). The higher preventable YLL rate fraction was estimated for provincial capitals (5.9% of YLL) and for provincial capitals with a population of 120,000 or more (6.4% of YLL). Considering all municipalities, the higher preventable YLL rate was estimated for the population aged ≥80 years (2575 YLL per 100,000) and the higher preventable YLL rate fraction for the population aged <80 years (5.1–5.2% of YLL). The findings of the uncertainty and sensitivity analyses for the HIA are reported in the Supplementary Notes 1–2 and in Supplementary Table 4. There were no relevant changes from using a different greenest period or satellite, excluding sparsely populated areas, applying a 20% reduction of the effect in non provincial capitals, or using an alternative approach to calculating the preventable fractions. Lower estimates were obtained by using as counterfactual exposures the 75th population-weighted percentile of the municipality-level PWEs to greenest period NDVI or the 50th population-weighted percentiles. Higher estimates were obtained by using as counterfactual exposures the 75th unweighted percentiles or the 95th population-weighted percentiles, or by using other exposure-response functions. Analysis of land cover The results of Spearman correlations and generalised additive mixed effect models (GAMM) between the PWEs to relative land cover measures and the PWE to greenest period NDVI are shown in Table 4 (Spearman correlation, single-predictor models and multiple-predictor models), Fig. 3 (single-predictor models) and Fig. 4 (multiple-predictor models). The plots are helpful in observing the estimated functional forms, while the statistics reported in the table identify the most influential measures.Table 4Relationship between the PWEs to relative land cover measures and the PWE to greenest period NDVI. Spearman rank correlations and GAMMs. Italy, 2021 (land cover) and 2022 (NDVI)N = 7904 municipalitiesSpearman rank correlationsGAMMs with cubic regression splines, province random effects, and population weightsExposures [log of PWE]ρ95% CIEDFAdj. R2Deviance explainedP-valueTree cover, proportion0.770.76; 0.787.850.820.84<0.001Shrubland, proportion−0.02−0.04; 0.00––––Grassland, proportion0.150.13; 0.176.020.630.66<0.001Green area, proportion0.860.85; 0.877.200.910.93<0.001Tree cover, proportion;Grassland, proportion––8.50; 6.320.920.93<0.001; <0.001Built-up, proportion−0.87−0.88; −0.877.940.930.94<0.001Tree cover, ratio to built-up0.870.86; 0.877.830.890.91<0.001Shrubland, ratio to built-up0.070.05; 0.09––––Grassland, ratio to built-up0.570.56; 0.595.460.760.79<0.001Green area, ratio to built-up0.890.89; 0.907.630.940.95<0.001Tree cover, ratio to built-up;Grassland, ratio to built-up––7.57; 7.200.930.94<0.001; <0.001Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). Land cover data from WorldCover are available at: https://code.earthengine.google.com (European Space Agency. WorldCover Project. 2021. Image collection: ‘ESA/WorldCover/v200’).Adj. R2 adjusted R2, CI confidence interval, EDF estimated degrees of freedom, GAMMs generalised additive mixed effects models, Log natural logarithm, NDVI normalized difference vegetation index, PWE population-weighted exposure.Fig. 3Relationship between the PWEs to relative land cover measures and the PWE to greenest period NDVI. The plots show the smooth functions as estimated by the single-predictor GAMMs. The shaded areas show confidence bands at 2 standard errors above and below the estimates. Rug plots show the distribution of the data. Italy, 2021 (land cover) and 2022 (NDVI).a, c and e: one model for each proportion measure. b, d and f: one model for each ratio-to-built-up (odds) measure. GAMMs: generalised additive mixed effects models. Log: natural logarithm. NDVI: normalized difference vegetation index. PWE: population-weighted exposure. Plot created with R version 4.2.3. Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). Land cover data from WorldCover are available at: https://code.earthengine.google.com (European Space Agency. WorldCover Project. 2021. Image collection: ‘ESA/WorldCover/v200’).Fig. 4Relationship between the PWEs to relative land cover measures and the PWE to greenest period NDVI. The plots show the smooth functions as estimated by the multiple-predictor GAMMs. The shaded areas show confidence bands at 2 standard errors above and below the estimates. Rug plots show the distribution of the data. Italy, 2021 (land cover) and 2022 (NDVI).a and c: one model for multiple proportion measures. b and d: one model for multiple ratio-to-built-up (odds) measures. GAMMs: generalised additive mixed effects models. Log: natural logarithm. NDVI: normalized difference vegetation index. PWE: population-weighted exposure. Plot created with R version 4.2.3. Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). Land cover data from WorldCover are available at: https://code.earthengine.google.com (European Space Agency. WorldCover Project. 2021. Image collection: ‘ESA/WorldCover/v200’). Table 4Relationship between the PWEs to relative land cover measures and the PWE to greenest period NDVI. Spearman rank correlations and GAMMs. Italy, 2021 (land cover) and 2022 (NDVI)N = 7904 municipalitiesSpearman rank correlationsGAMMs with cubic regression splines, province random effects, and population weightsExposures [log of PWE]ρ95% CIEDFAdj. R2Deviance explainedP-valueTree cover, proportion0.770.76; 0.787.850.820.84<0.001Shrubland, proportion−0.02−0.04; 0.00––––Grassland, proportion0.150.13; 0.176.020.630.66<0.001Green area, proportion0.860.85; 0.877.200.910.93<0.001Tree cover, proportion;Grassland, proportion––8.50; 6.320.920.93<0.001; <0.001Built-up, proportion−0.87−0.88; −0.877.940.930.94<0.001Tree cover, ratio to built-up0.870.86; 0.877.830.890.91<0.001Shrubland, ratio to built-up0.070.05; 0.09––––Grassland, ratio to built-up0.570.56; 0.595.460.760.79<0.001Green area, ratio to built-up0.890.89; 0.907.630.940.95<0.001Tree cover, ratio to built-up;Grassland, ratio to built-up––7.57; 7.200.930.94<0.001; <0.001Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). Land cover data from WorldCover are available at: https://code.earthengine.google.com (European Space Agency. WorldCover Project. 2021. Image collection: ‘ESA/WorldCover/v200’).Adj. R2 adjusted R2, CI confidence interval, EDF estimated degrees of freedom, GAMMs generalised additive mixed effects models, Log natural logarithm, NDVI normalized difference vegetation index, PWE population-weighted exposure. Relationship between the PWEs to relative land cover measures and the PWE to greenest period NDVI. Spearman rank correlations and GAMMs. Italy, 2021 (land cover) and 2022 (NDVI) Tree cover, proportion; Grassland, proportion Tree cover, ratio to built-up; Grassland, ratio to built-up Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). Land cover data from WorldCover are available at: https://code.earthengine.google.com (European Space Agency. WorldCover Project. 2021. Image collection: ‘ESA/WorldCover/v200’). Adj. R2 adjusted R2, CI confidence interval, EDF estimated degrees of freedom, GAMMs generalised additive mixed effects models, Log natural logarithm, NDVI normalized difference vegetation index, PWE population-weighted exposure. Relationship between the PWEs to relative land cover measures and the PWE to greenest period NDVI. The plots show the smooth functions as estimated by the single-predictor GAMMs. The shaded areas show confidence bands at 2 standard errors above and below the estimates. Rug plots show the distribution of the data. Italy, 2021 (land cover) and 2022 (NDVI). a, c and e: one model for each proportion measure. b, d and f: one model for each ratio-to-built-up (odds) measure. GAMMs: generalised additive mixed effects models. Log: natural logarithm. NDVI: normalized difference vegetation index. PWE: population-weighted exposure. Plot created with R version 4.2.3. Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). Land cover data from WorldCover are available at: https://code.earthengine.google.com (European Space Agency. WorldCover Project. 2021. Image collection: ‘ESA/WorldCover/v200’). Relationship between the PWEs to relative land cover measures and the PWE to greenest period NDVI. The plots show the smooth functions as estimated by the multiple-predictor GAMMs. The shaded areas show confidence bands at 2 standard errors above and below the estimates. Rug plots show the distribution of the data. Italy, 2021 (land cover) and 2022 (NDVI). a and c: one model for multiple proportion measures. b and d: one model for multiple ratio-to-built-up (odds) measures. GAMMs: generalised additive mixed effects models. Log: natural logarithm. NDVI: normalized difference vegetation index. PWE: population-weighted exposure. Plot created with R version 4.2.3. Population data from GHSL are available at: https://ghsl.jrc.ec.europa.eu/download.php?ds=pop (Global Human Settlement Layer. GHS population grid (R2023). Product: GHS-POP, epoch: 2020, resolution: 100 m, coordinate system: Mollweide). NDVI data from Sentinel-2 are available at: https://code.earthengine.google.com (European Union. Copernicus Programme. Sentinel-2 mission. Image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). Land cover data from WorldCover are available at: https://code.earthengine.google.com (European Space Agency. WorldCover Project. 2021. Image collection: ‘ESA/WorldCover/v200’). The proportion measures most influential in determining the PWE to greenest period NDVI appear to be the built-up (ρ −0.87; adjusted R2 0.93; null deviance explained 0.94), the green area (ρ 0.86; adjusted R2 0.91; null deviance explained 0.93) and the tree cover (ρ 0.77; adjusted R2 0.82; null deviance explained 0.84). The model with both tree cover and grassland proportion shows similar results (adjusted R2 0.92; null deviance explained 0.93). The ratio-to-built-up (odds) measures most influential in determining the PWE to greenest period NDVI appear to be the green area (ρ 0.89; adjusted R2 0.94; null deviance explained 0.95) and the tree cover (ρ 0.87; adjusted R2 0.89; null deviance explained 0.91). The model with both tree cover and grassland ratio-to-built-up shows similar results (adjusted R2 0.93; null deviance explained 0.94). The estimated functional forms of these relationships are shown in Figs. 3 and 4. The shaded areas show confidence bands at 2 standard errors above and below the estimates. Rug plots show the distribution of the data. Especially in the spaces with denser data distributions and narrower confidence intervals, the greenest period NDVI increases with increasing tree cover and green area. Applying various sensitivity analyses to the GAMMs did not lead to any significant changes in the results. The estimates from excluding sparsely populated areas, or from changing the smoothing terms are similar to the estimates from the main analysis. We have reported the results from the models without random effects and/or population weights in Supplementary Tables 5–7 and in Supplementary Figs. 4–9. As in the main analysis, in the spaces with denser data distributions and narrower confidence intervals, the greenest period NDVI increases with increasing tree cover and green area. Discussion In this a nationwide health impact assessment study, we estimate the number of deaths that could be prevented in Italy by greening residential areas up to the level of greenness currently achieved by the 25% of the population. Exposure was assessed within a 300-m distance from homes capitalising on satellite images at 10-m resolution. The estimated total mortality burden in the 49 million adult population varies from 2.1% in the less populated not provincial capitals to 6.3% in the more populated provincial capitals. This findings are also consistent with the estimated proportions of tree cover and green area for each group of municipalities. Evaluation of less ambitious NDVI targets resulted in lower but still relevant mortality estimates. Green space exposure depends not only on the amount of green space, but also on its spatial distribution in relation to the population. Areas that are green but far from homes do not contribute to the residential exposure. Unlike previous HIA studies27,28,43 which set different targets for each city, we experimented with using a single population-based target level of greenest period NDVI to ensure comparability of estimates. In addition, Barboza et colleagues28 stated that a greater number of deaths could be prevented by providing more green space than recommended by the WHO1. Our main exposure target of 0.46 was based on measured data, namely the 75th population-weighted percentile of the greenest period NDVI calculated for the whole country within 300-m distance of homes. This approach is consistent with the WHO and Konijnendijk recommendations for the 300-m distance1,26. Furthermore, in the study area, a PWE to greenest period NDVI of 0.46 within 300 m from residence approximately corresponds to a PWE to tree cover proportion of 30% within the same distance. The only element of the 3–30–300 rule that we could not directly assess on a large scale was the presence of three trees visible from the house, which would require data with a higher spatial resolution. We also couldn’t assess the quality or actual accessibility of the green space within the 300-m distance. Taking these limitations into account, the population-based counterfactual target of 0.46 on average seems to correspond roughly to the 30–300 part of the 3–30–300 rule in Italy. In the future, it would be interesting to assess the population-weighted NDVI exposure in specific Italian areas or in other countries, and how local NDVI measures relate to the land cover measures and to the Konijnendijk’s rule. The analysis of land cover data in the present study suggests the essential role of trees in residential areas in Italy, followed by grassland. The ratio-to-built-up measures are more strongly associated with the greenest period NDVI, and the ratio-to-built-up of green area and tree cover explain much of the variability in NDVI. In Italy, on average, the single most influential positive NDVI determinant appears to be the tree cover, while the most influential negative NDVI determinant appears to be the human built environment. Thus, the ratio-to-built-up measures seem to be more specific indicators of human exposure to green spaces than the proportion measures. Therefore, it is of utmost importance to integrate urban green spaces into urban centres. This will disassociate the built environment from the absence of greenness and is likely to reduce mortality. These findings are consistent with published research on the beneficial effects of trees on human health12,44–56. In general, trees have been associated with reducing harms (excessive heat, air pollution, noise, ultraviolet radiation, crime), restoring capacities (effects on cognition and attention, mental health, mood, anxiety, psychophysiological stress and clinical outcomes), and building capacities (impacts on birth outcomes, immune system, active living, weight status, cardiovascular function, social cohesion)12,44,45. Urban trees have been shown to have specific benefits in the areas of health and social well‐being (they reduce pollution, improve physical and mental health, strengthen community bonds, increase physical activity, decrease aggression and violence, reduce crime), cognitive development and education (they improve student performance, reduce stress, increase concentration, improve attention and self‐discipline), economy and resources (high return‐on‐investment, support tourism, increase houses prices and rents, reduce energy use and bills, promote food sustainability, provide resources and firewood), climate change mitigation and habitat (they reduce the urban heat island effect, store and sequester carbon, provide critical habitat), and green infrastructure (they manage stormwater and protect aquatic and terrestrial life)12,45,46. Natural capital is the term used to describe components of the environment that provide value to people, either directly or indirectly. Many countries are currently in the process of identifying the location, quality, and ecosystem services that make up their natural capital assets and contribute to human well-being, not only in rural areas but also in urban areas. Urban trees are an example of this type of natural capital47. When Kardan and colleagues used high-quality public health and demographic data to compare neighbourhoods in Toronto with different densities of street trees, they found a correlation between higher tree density and lower incidence of heart and metabolic disease, as well as better health perceptions. According to the authors, planting just ten or more trees per city block can save a household from paying over $10,000 in medical costs. This amount far exceeds the projected costs of planting and maintaining those for ten additional trees48. Similarly, Beyer et al. examined a range of urban and rural settings in Wisconsin, USA, and found that having more trees in a neighbourhood — measured as a higher percentage of tree canopy — was associated with better mental health, especially for people aged 55 and older, after adjusting for a wide range of confounders49. Similarly, in a cross-sectional study conducted in London, UK, Taylor and colleagues found that street tree density was lower in areas where smoking and antidepressant prescription rates were higher. Prescriptions for antidepressants were correlated with smoking levels, but the relationship between the number of trees and depression prescriptions persisted even after confounding variables were taken into account50. Strong evidence for the benefits of trees to human health can also be found in Donovan and colleagues’ analysis of the consequences of city tree removal. The study analysed health data before and after the emerald ash borer infestation that resulted in the loss of 100 million ash trees in 1296 U.S. counties between 1990 and 2007. They discovered that the loss of trees was linked to statistically significant increases in mortality from lower respiratory tract and cardiovascular diseases51. More neighbourhood tree cover was linked to better overall health, independent of access to green space, according to a study by Ulmer and colleagues using LIDAR (Light Detection and Ranging) data in California. This association was primarily mediated by lower overweight/obesity and better social cohesion, and to a lesser extent by lower type 2 diabetes, high blood pressure, and asthma. According to these results, nature and trees have a significant impact on improving overall population health in urban environments52. According to Astell-Burt and Feng, people with ≥30% tree cover have a decreased risk of heart disease, hypertension, and diabetes than those with 0–9% tree canopy. In addition, research revealed that areas with ≥30% tree canopy has decreased prevalence of diabetes, hypertension, and heart disease than areas with 0–9% tree cover53. Schwaab et al. show that urban trees in most European cities have lower temperatures than the urban fabric during summer and heat waves. When compared to continuous urban fabric, land surface temperatures recorded for urban trees are, on average, 0–4 K lower in Southern European regions and 8–12 K lower in Central European regions. The overall effectiveness of treeless urban green spaces in reducing temperatures is lower, and their cooling impact is roughly 2–4 times less than that of urban trees54. Trees not only block the sun’s rays from reaching people by providing shade, but they also limit the amount of heat that impermeable materials with high heat capacity and thermal conductivity, such as concrete, can absorb from the sun. Vegetation can increase urban albedo relative to black asphalt surfaces, and when compared to impermeable surfaces of the same albedo, vegetated surfaces have lower radiative temperatures45,55–57. Examples of greening interventions include the regeneration of urban areas (turning former industrial areas into urban parks), the increase of nature-based solutions in existing buildings (green roofs and vertical gardens), the reorganisation of traffic and the reallocation of road and parking space to green and natural areas (green belt and corridors), and the general greening of the city with more street trees, green corridors, and pocket parks28. Our results show discrepancies between classes of municipalities, with higher impacts in provincial capitals with 120,000 inhabitants or more (6.3% of mortality). The groups of municipalities showing lower (<3.5% of mortality) health impacts were non provincial capitals with less than 5000 inhabitants and provincial capitals with less than 45,000 inhabitants. It is important to carry out assessments for all municipalities and inhabitants, not only for selected or larger cities. In this work, we have carried out a national assessment by using large-scale data and setting a statistical, nationwide population-based target. This also ensures comparability of the estimates. However, these aspects, which are the main strengths of the study, could also be a source of concern as we lack local information and data on the practical applicability of the NDVI target in each municipality in terms of interventions and characteristics of greenness. The counterfactual exposure to greenness is statistically determined at the national level and represents a generic target that may be too high for some municipalities and too low for others. Therefore, it may be desirable in the future to also carry out assessments that are more adapted to the local context. These specific studies and targeted interventions are necessary to achieve the highest possible level of greenery in each area, compatible with the local geographical, climatic, floristic, urban, architectural, and cultural characteristics. Adaptation to climate change and urban green spaces are closely linked. Heat-related morbidity in cities is a major public health concern. The urban heat island effect can pose serious health risks during heat waves and extreme heat events. Urban green spaces, such as parks, street trees, and green roofs, can mitigate these effects with an estimated cooling effect of 1°. Trees can provide shade in warmer weather and reduce the need for air conditioning. In warmer countries, they can also provide comfortable outdoor spaces and help people avoid heat-related stress. Mitigating the urban heat island may partly explain the benefits of green spaces on mortality. Therefore, future avoidable deaths from the implementation of urban greening could potentially be even greater if we take into account future, increased baseline temperature-related deaths1,11,28,58–60. It would therefore be interesting to extend this research study in the future with scenarios of attributable deaths due to climate change58,59. Strengths of the present study are the inclusion of all municipalities and population of a country, the use of ultra-fine resolution satellite data for exposure, the use of the WHO and Konijnendijk criteria of 300 m for assessing the presence of residential greenness, the implementation of the population-weighted approach to account for the differences in the spatial distribution of the population in different areas, the use of recent exposure, population and mortality data, the use of city-specific and age-specific mortality data, the stratified analysis by municipality groups and age groups, the use of the same counterfactual exposure based on real data in all municipalities for comparability purposes, the approximate compliance of this counterfactual with the 30–300 part of the 3–30–300 rule, the analyses of land cover, and the application of several sensitivity analyses for populations, targets, functions, formulas and statistical models. In addition, our method is easily replicable without the need for georeferencing the local population. The PWE approach allows exposure to be estimated at the municipal level using public exposure and population data, which are freely available worldwide at high resolution. The required information on mortality and population are aggregated at the municipal level. Using the most recent statistical data, the HIA can be replicated worldwide. Furthermore, the implementation of various sensitivity analyses allows the comparison of impacts assessed using different methods and scenarios. On the flip side, there are also limitations that are essentially shared with the other published studies on NDVI and mortality2–11,27,28. The first is the limitation, common to most of environmental epidemiology studies, of using residential exposure as a proxy for overall individual exposure, even though people spend only part of their time at home. However, the 300-m approach employed in the present HIA can somewhat extend the exposure observation to areas around the house within a radius of 300 m. Therefore, the PWE method using these buffers could also consider the average exposure to greenness near homes1,11,26,28. A second limitation is that the assessment does not take into account the quality or use of green spaces, nor the specific characteristics of the areas analysed. Basically, it is assumed that all green spaces measured by the NDVI are qualitatively homogeneous and equally accessible to all residents in the buffer centroid. Instead, urban green spaces have different components that influence their quality and shape the patterns of use, time spent, and interactions with these spaces among different population subgroups28. Key attributes of green spaces, such as safety, aesthetics, amenities and maintenance are crucial in promoting outdoor physical activity. Negative aspects such as safety concerns, graffiti, vandalism, litter, noise, pollution, and dog fouling discourage park use and physical activity. Studies suggest that access to attractive, large public open spaces is associated with higher levels of walking61–65. The quality of green spaces, characterised by accessibility, maintenance, absence of litter, and safety, is positively associated with general health. The ability of green spaces to provide relaxation and recreation has been identified as crucial to mental well-being. Eight sensory dimensions of urban parks have been identified: serene, space, nature, rich in species, refuge, culture, prospect, and social. Refuge and nature are strongly linked to reduced stress. Refuge is a safe, enclosed area for play and observation, while nature provides a sense of being in the wild. Access to serene spaces significantly reduce the risk of mental illness. In addition, views of trees and grass from homes prevent aggression and mental fatigue compared with barren views66–74. In general, the methodology of this health impact assessment is focused on the quantitative analysis of the greenest period NDVI rather than on the qualitative assessment of green spaces. However, conducting studies on entire nations has these challenging limitations. To go into more detail and consider the quality of green spaces, it may be necessary to restrict the study to a limited number of municipalities and include additional local-level variables. Furthermore, translating these analyses into health impacts requires specific exposure-response functions that are not currently available. Future research could benefit from incorporating methods that explicitly categorise and qualify different types of green space to better understand their unique roles and impacts in different settings. A third limitation is the assumption of homogeneity in the effect of urban greenery on mortality across different age groups, sexes, socio-economic indicators, and types of human settlements according to the degree of urbanisation. In fact, we assume that the relative risk for given NDVI increments is the same in all these categories. However, we do not have solid evidence that the exposure-response function between NDVI and mortality is not homogeneous to use this information in a reliable assessment. With regard to age, one of the studies included in the meta-analysis reported a lower RR in older people compared with the pooled estimate from the meta-analysis in all age groups7,11. Conversely, another study in the meta-analysis seemed to show a more protective effect in younger age groups2. Therefore, we have reported two different sensitivity analyses for the two scenarios. We have also performed a sensitivity analysis for possible differences according to the degree of urbanisation. Further research is therefore needed to investigate the possible heterogeneity of the effect in more detail. A fourth limitation is the assumption that the exposure-response function is linear. This is consistent with the assumption of linearity of the exposure-response function reported in the meta-analysis11. Although all cohort studies included in the meta-analysis2–10 reported RRs for continuous green space exposure, we are not sure that the exposure-response function is strictly linear11. However, most of the studies included in the meta-analysis had also examined the functional forms and reported an essentially linear relationship2,4–6,10. Some of these studies also showed the plotted functional forms of the estimated relationship between greenest period NDVI and mortality2,6. Our study examined the health impacts of positive greenest period NDVI values up to a counterfactual of 0.46, a relatively limited range in which the linearity assumption appears to hold according to the plots. Finally, other studies have also reported a linear relationship between green space and human health75–77. However, we hope that future research can further investigate possible non-linear relationships between greenness and mortality. The assumptions about homogeneity and linearity are also consistent with the approach used in this HIA, which is based on the guidance provided in various scientific papers and official documents28,33–43,78–86. For example, the same framework (use of metanalytic RRs to estimate health impacts) and assumptions (linearity of the exposure-response function and homogeneity of the effect) were reported by the WHO also in the officially provided tools for air quality HIA (AirQ+)85. The WHO HRAPIE (Health risks of air pollution in Europe) document on exposure-response functions for air pollutants86 recommended the same approach and basically provided a list of available metanalytic RRs to be used in HIA, based on the same assumptions mentioned above. Therefore, this methodological part of the work (HIA design and assumptions in the use of the exposure-response function) is basically a standard and well referenced approach in HIA28,33–43,78–86. However, further research is essential and recommended for the future to explore these associations in more detail and to provide more accurate and tailored risk functions to enable more accurate assessments. With these perspectives in mind, the exposure-response function used in the present study11 is currently the most recent and robust estimate available in the literature for conducting a reliable health impact assessment. In conclusion, we estimate in this study that 5% of all deaths and years of life lost in Italy could be prevented by greening residential areas up to the level of greenness currently achieved by the 25% of the population. More green means fewer deaths, thus strong action is needed to increase the amount and accessibility of green spaces in all human settlements. Methods Design, analysis, and study area This is a quantitative health impact assessment study. Data analysis and visualisation were carried out using Google Earth Engine, R version 4.2.3 and QGIS version 3.28.4. The study area corresponds to all the Italian municipalities, regardless of size or population, as reported by the Italian National Institute for Statistics (ISTAT). Municipalities can be classified as provincial capitals or not, and according to their population. For the year 2022, the ISTAT list includes 7,904 municipalities, 109 provincial capitals and 7,795 non provincial capitals. On 31 December 2022, the total resident adult population (aged ≥20 years) was 48,628,328: 14,530,259 in provincial capitals and 34,098,069 in non provincial capitals87–89. Internal review board approval This study is neither analytical nor individual. It is a municipality-level health impact assessment that attempts to estimate the impact of a hypothetical greening intervention using public data and functions. This is a commonly used methodology in public health and generally does not require ethical approval. However, we submitted a request to the scientific secretariat of the local ethics committee asking if their opinion was required for the study. After evaluating the manuscript, they replied that an internal review board approval is not required for this type of study, because it does not involve human subjects and relies on public mortality counts at the municipal level to assess the impact of an environmental intervention. Population and baseline outcome data The total number of inhabitants per grid cell was obtained from the Global Human Settlement (GHS) population grid for 2020 with a resolution of 100 m. This spatial raster product (‘GHS-POP_GLOBE_R2023’) depicts the distribution of residential population, expressed as the number of people per cell90–93. Pixels with no population were treated as no-data and excluded from the analysis. The population raster was converted into a vector layer of points and the information on the Italian administrative units were spatially joined from the vector of polygons of the ISTAT municipalities87. Non-Italian population points (i.e., points with no joined attributes) were excluded. This vector of population-points was used for the exposure assessment. The GHS population was used to calculate the weighted percentiles and means of the buffer-level exposures (see the section ‘Baseline exposure’). The adult (≥20 years) and age-specific (5-year age groups, 20–24, 25–29,…,95–99, ≥100) municipal ISTAT population was used to calculate the age-specific rates, to classify the municipalities by number of inhabitants, and to calculate the weighted percentiles and means of the municipal level exposures and populations88,89. The Spearman rank correlation between the GHS and ISTAT populations at the municipal level was > 0.99. The number of all-cause baseline deaths (BD) for each age group of the adult population in 2022 by municipality was obtained from a public ISTAT dataset94. The number of all-cause baseline years of life lost (BYLL) for each age group of the adult population in 2022 by municipality was estimated by multiplying the number of baseline deaths in the age group by the central age (3rd year of the age group) province-specific life expectancy reported by ISTAT95. Baseline exposure The baseline exposure is the actual level of greenness as measured by the greenest period NDVI. We calculated the NDVI at a resolution of 10 m using images from the Sentinel-2 mission of the European Union’s Copernicus Programme via Google Earth Engine (image collection: ‘COPERNICUS/S2_SR_HARMONIZED’). Sentinel-2 is a constellation of two satellites with a minimum combined revisit time of 5 days96. All images collected for Italy87 between 1 April and 30 June 2022 were included to ensure the greenest period of the year. The NDVI is not an exposure itself, but rather an index proxy of the real exposure, which is the green space. For simplicity, we have referred to it in the text as ‘PWE to greenest period NDVI’, but the full expression should be ‘PWE to greenness throughout the year, assessed using the greenest period NDVI’28. A more detailed description of this approach is reported in the Supplementary Note 2. A 5 km buffer was added to the entire study area to avoid loss of exposure information at the boundaries. Images with a granule-specific cloudy pixel percentage of 20% or more were removed (images of lower quality). The selected images were also masked using the ‘QA60’ bitmask band information to remove pixels with opaque clouds and cirrus clouds. Then, for each image time t in a pixel space s, the NDVI \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$({{{{\rm{NDVI}}}}}_{{t}_{s}})$$\end{document}(NDVIts) is estimated according to Eq. (1):1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{{\rm{NDVI}}}}}_{{t}_{s}}}=\frac{{{{{\rm{NIR}}}}_{{t}_{s}}}-{{{{\rm{Red}}}}}_{{t}_{s}}}{{{{{\rm{NIR}}}}_{{t}_{s}}}+{{{{\rm{Red}}}}_{{t}_{s}}}}$$\end{document}NDVIts=NIRts−RedtsNIRts+Redtswhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{NIR}}}}}_{{t}_{s}}$$\end{document}NIRts and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{Red}}}}}_{{t}_{s}}$$\end{document}Redts refer to the spectral reflectances of the pixel at a specific time t in a specific space s measured in the near infrared (833–835 nm) and red (665 nm) wavebands respectively. The calculation of the NDVI always results in a number that ranges from −1 to 1, with positive and higher values indicating more greenness11,13. For each pixel space s in the study region, a reduction of the NDVI measures was obtained by calculating the temporal arithmetic mean (NDVIs) of all the NDVI values of the corresponding pixels across the images in the temporal line \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$({{{{\rm{NDVI}}}}}_{{t}_{s}})$$\end{document}(NDVIts). To account for the possible beneficial effects of the blue areas on human health, the pixels corresponding to permanent water bodies (see the section ‘Analysis of land cover’) were excluded (masked out) from the exposure assessment28. In line with the WHO’s and Konijnendijk’s recommendations1,26, the exposure to NDVI value was estimated by adding a circular buffer of 300 m radius around each population-point to indicate the proximity to greenness (i.e., about 5 min walk along walkable pathways)28. For each population-buffer p, the spatial arithmetic mean (NDVIp) of the temporally averaged NDVI values of all the pixels within the population-buffer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$({{{{\rm{NDVI}}}}}_{{s}_{p}})$$\end{document}(NDVIsp) was calculated. Basically, a population-buffer was created for each inhabited point, with the point-derived (central population-point) information of population count and administrative division (municipality and province), and the circle-area-derived (surrounding population-buffer) information of temporally and spatially averaged NDVI. For each municipality-area a, the PWE to greenest period NDVI (PWEa) was calculated as a weighted mean of the temporally and spatially averaged NDVI values \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$({{{{\rm{NDVI}}}}}_{{p}_{a}})$$\end{document}(NDVIpa) of the population-buffers pa whose centroid is included in the municipality-area a. For each included population-buffer pa, the weight was the GHS population \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$({{{{\rm{POP}}}}}_{{p}_{a}})$$\end{document}(POPpa) of the centroid, as shown in Eq. (2)31,33,97.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{PWE}}}}}_{a}={\sum}_{{p}_{a}}({{{{\rm{NDVI}}}}}_{{p}_{a}}\times {{{{\rm{POP}}}}}_{{p}_{a}})/{\sum}_{{p}_{a}}({{{{\rm{POP}}}}}_{{p}_{a}})$$\end{document}PWEa=∑pa(NDVIpa×POPpa)/∑pa(POPpa) This approach allows exposure to be weighted by population and has been used in other studies on air pollution33,35,36,39–41,97. To take in consideration the influence of sparsely populated areas28, sensitivity analyses were performed including only the population-buffers with a population count greater than 2. We also reported an uncertainty analysis for the baseline exposure assessment and sensitivity analyses for the choice of the greenest period and satellite. Further details are provided in the Supplementary Notes 1 and 2, respectively. Counterfactual exposure The counterfactual exposure is the level of greenness that needs to be achieved to prevent the deaths estimated in the health impact assessment, i.e. a target level of exposure to greenness. There are no specific recommendations for NDVI values when setting the counterfactual exposure. It may be desirable to conduct these assessments using the same counterfactual exposure for all areas to assure comparability. This target exposure should be population-based, realistically achievable in some way, and possibly linked to the current green space recommendations. We proposed a population-weighted approach. Using a methodology consistent with that chosen for baseline exposure, the counterfactual level of exposure to green space was set as the 75th population-weighted percentile of the population-buffers’ greenest period NDVI. We used the 75th population-weighted percentile as a conservative measure to try to make the target realistically achievable. This means that 75% of the Italian population was exposed to a buffer-level mean NDVI in the period April-June 2022 that was lower than the counterfactual level. Or, to say in other words, that 25% of the Italian population was exposed to a buffer-level greenest period NDVI that was higher than the counterfactual level, thus achieving the target. In the study area, this counterfactual greenest period NDVI value is 0.46 and corresponds to an average of 30% tree cover within a 300-m radius buffer around the residence. Specifically, selecting only those municipalities with a PWE to greenest period NDVI in the range 0.46 ± 0.01, we found an overall PWE to tree cover proportion of 30% (see the section ‘Analysis of land cover’)1,26. This approach has several advantages. First, it could evaluate green targets that are statistically based on real exposures and populations. Second, it ensures comparability across all municipalities. Third, it is likely to be quite robust to errors in exposure assessment, as hypothetical shifts in baseline exposure measures could be propagated to the percentile-based target, making the expected difference similar. Finally, the target in the study area approximately complies with the 30–300 part of the 3–30–300 rule26. We used sensitivity analyses to explore possible changes in impact estimates by setting alternative counterfactual exposures. Specifically, we used as alternative counterfactual exposures the 75th population-weighted percentile of the municipality-level PWEs to greenest period NDVI and the 75th unweighted percentiles of both municipality-level PWEs to greenest period NDVI and buffer-level greenest period NDVI. The first target implies that 75% of the Italian population was exposed to a municipality-level (PWE) mean NDVI in the period April-June 2022 that was lower than the counterfactual level. The interpretation of the unweighted percentiles counterfactuals is analogue to the weighted percentiles ones: it means that in the period April-June 2022 the population of 75% of the Italian municipalities or population-buffers was exposed to a municipality-level (PWE) or buffer-level mean NDVI value that was lower than the counterfactual level, respectively. We performed further sensitivity analyses using the 50th and the 95th population-weighted percentiles as counterfactual exposures. We also reported an uncertainty analysis for the counterfactual exposure assessment and sensitivity analyses for the choice of the greenest period and satellite. Further details are provided in the Supplementary Notes 1 and 2, respectively. Exposure-response function We used the exposure-response function from the systematic review and meta-analysis by Rojas-Rueda et al.11. The review included only cohort studies and estimated that the RR for all-cause mortality per 0.1 NDVI increase within a buffer of 500 m or less from homes was 0.96 (95% CI 0.94–0.97). Although this exposure-response function is the most recent and reliable estimate to date, sensitivity analyses were performed to assess how the results might vary with alternative functions. Specifically, the relative risks from the two studies with the higher7 and lower4 weight in the meta-analysis11 were used as alternative exposure-response functions. In addition, to account for the possible heterogeneity of the effects between different age groups2,7,11, two sensitivity analyses were performed, one using a different exposure-response function for the population aged ≥80 years (greater protective effect compared with the meta-analytic estimate)7,11 and the other testing a + 50% of protective effect in the population aged <80 years. Finally, to take into account the possible heterogeneity of the effect between different types of municipalities4,6,7,11, we performed a sensitivity analysis considering a 20% reduction in the protective effect in non provincial capital municipalities. This value was calculated as a weighted average of the estimated effect reduction in rural areas, using information from some of the studies in the meta-analysis4,6,7,11. The weights were the same as those used in the meta-analysis11. To be conservative, we have applied this 20% reduction to all non provincial capital, although not all of them can be considered rural87–89. Health impact assessment A quantitative health impact assessment at the municipal level was conducted to estimate the impact on all-cause mortality of increasing the exposure to greenness up to the counterfactual value11,27–43. For each municipality-area a, the exposure difference (Δa) between the counterfactual (k) and the baseline (PWEa) greenest period NDVI exposure level was calculated. For each municipality-area a, the relative risk for the exposure difference (RRa) and the preventable fraction (PFa) were estimated using the exposure-response function (RR0.1) with its confidence interval11, as shown in Eqs. (3) and (4)29,34,35:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{RR}}}}}_{a}=\exp ({{\mathrm{ln}}}({{{{\rm{RR}}}}}_{0.1})\times 10\times {\varDelta }_{a})$$\end{document}RRa=exp(ln(RR0.1)×10×Δa)4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{PF}}}}}_{a}=1-{{{{\rm{RR}}}}}_{a}$$\end{document}PFa=1−RRa For each municipality-area a and five-year age group j, the number of preventable deaths (PDa,j) and preventable years of life lost (PYLLa,j) was estimated, using the municipality and age-specific baseline deaths (BDa,j) and years of life lost (BYLLa,j), as shown in Eqs. (5) and (6)27,29–35,37,38:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{PD}}}}}_{a,j}={{{{\rm{PF}}}}}_{a}\times {{{{\rm{BD}}}}}_{a,j}$$\end{document}PDa,j=PFa×BDa,j6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{PYLL}}}}}_{a,j}={{{{\rm{PF}}}}}_{a}\times {{{{\rm{BYLL}}}}}_{a,j}$$\end{document}PYLLa,j=PFa×BYLLa,j Sensitivity analyses were performed using an alternative approach in calculating the preventable fractions. The full explanation of the formulas used in the two approaches is given in the Supplementary Note 3. Presentation of the assessments’ results Municipalities were classified according to whether they were provincial capitals or not and according to their population size. The population categories were chosen according to their rounded, unweighted quartiles specific to provincial capitals and non provincial capitals. The five-year age groups were combined into four main age groups (20–39, 40–59, 60–79, and ≥80 years). Results were reported for all the municipalities with a preventable fraction > 0, corresponding to a PWE to greenest period NDVI below the counterfactual exposure. The estimated impact measures were summed by municipality group and age group and reported as preventable numbers (PD or PYLL), preventable rates (the ratio of the preventable numbers to the ISTAT population, per 100,000) and preventable rate fractions (the ratio of the preventable rates to the baseline rates). These preventable (rate) fractions were calculated for each group using the health impact estimates and can be interpreted as the fractions of the baseline deaths or YLL that could be avoided by achieving the counterfactual exposure level29. The PWEs to greenest period NDVI, difference between counterfactual exposure and greenest period NDVI, tree cover proportion and green area proportion (see the section ‘Analysis of land cover’) were calculated for each group of municipalities as a weighted mean of the respective PWEs of the municipalities included. For each municipality, the weight was the ISTAT population. Analysis of land cover We retrieved via Google Earth Engine the latest available land cover information for Italy87 from the European Space Agency (ESA) WorldCover 10 m 2021 product (image collection: ‘ESA/WorldCover/v200’), which provides a global map for 2021 at 10 m resolution based on Sentinel-2 and Sentinel-1 data. The product includes 11 land cover classes and has been generated in the framework of the ESA WorldCover Project, part of the 5th Earth Observation Envelope Programme of the ESA98. As with the NDVI assessment, a 5 km buffer around the national boundaries was added, and the land cover pixels classified as permanent water bodies were excluded from the analysis. The classes analysed were tree cover, shrubland, grassland, green area (sum of the first three) and built-up. For each 300-m buffer, the proportion of a class was calculated as the ratio of the sum of the pixels in that class to the sum of all pixels. The PWEs to these proportion measures were calculated using the same methods described for the NDVI. Specifically, the PWE to the proportion of each green class was calculated for each municipality as the weighted mean of the proportion measure of the specific green class of the population-buffers whose centroid was included in the municipality itself. For each included population-buffer, the weight was the GHS population of the centroid. To quantify the presence of specific green classes in relation to the human built environment, the PWE to the ratio of each green class to built-up class (odds) was estimated for each municipality as the ratio of the PWE to the specific green class proportion and the PWE to the built-up proportion. To summarise, the calculated proportion measures (PWE) were the proportion of tree cover, shrubland, grassland, green area and built-up. The calculated ratio-to-built-up measures (PWE) were the ratio-to-built-up of tree cover, shrubland, grassland and green area. Spearman ρ correlation coefficients were calculated at the municipal level between the natural logarithm of the PWE to each of the relative land cover measures (proportion and ratio-to-built-up of the analysed classes) and the PWE to the greenest period NDVI. Generalised additive mixed effect models were fitted at municipal level using the PWE to greenest period NDVI as the dependent variable and the natural logarithm of the PWEs to each of the relative land cover measures as the fixed effects predictors one at a time (single-predictor models: one model for each proportion measure and one model for each ratio-to-built-up measure) or in combination (multiple-predictor models: one model for multiple proportion measures and one model for multiple ratio-to-built-up measures). A cubic regression spline penalised by the conventional integrated square second derivative cubic spline penalty was chosen as smooth term for the relative land cover measures. In all models, the province was included as a random effect to account for the differences between the provinces, specifically as a parametric term penalised by a ridge penalty, which is equivalent to assuming that the coefficients are independent and identically distributed normal random effects. The GHS population was used as weight in the models. Restricted maximum likelihood was used as the smoothness selection method99. The effect of the relative land cover measure (smooth function) on the greenest period NDVI was plotted for each model. The plotted values of greenest period NDVI represent only the estimated effect of the relative land cover measure, and not the absolute values of the NDVI. In our models, this value depends on the intercept, the random effect of province, and the land cover measures. The shaded areas show confidence bands at 2 standard errors above and below the estimates. Rug plots show the distribution of the data. Sensitivity analyses were performed by choosing different smoothing terms (thin plate regression spline, cubic or thin plate regression spline with shrinkage, cubic b-spline or p-spline) or by fitting models without random effects and/or population weights99. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Peer Review File Reporting Summary Supplementary Information Peer Review File Reporting Summary