Introduction

Climate change and urbanisation have resulted in urban overheating worldwide1. The effects of urban overheating are spatially uneven, as tropical areas are expected to experience amplified warming of extreme temperatures – in intensity and frequency of hot days in the twenty-first century – relative to other regions2,3. This is due to both anthropogenic climate change4 and rapid urbanisation trends5, with the urban population in Association of Southeast Asian Nations (ASEAN) states, which all experience tropical climates, set to increase by another 70 million people. This increase is more than the current population of all ASEAN capital cities combined6.

Within the region, warming is also not ubiquitous, as urban areas experience enhanced warming relative to their rural areas, a phenomenon known as the urban heat island (UHI). The UHI is caused by changes in the local-scale urban energy balance, attributed to anthropogenic heat waste (e.g., from air-conditioning and fuel combustion), and conversions of vegetated surfaces to surfaces and materials of higher thermal inertia (e.g., asphalt and concrete)7.

Unless appropriate planning interventions are deployed, urban overheating can lead to higher heat-related mortality and morbidity8,9, and declining mental- and physical-health10. To alleviate these detriments, cities look at several options including physical infrastructure, social and planning policies, and nature-based solutions (NbS)4,11. One effective NbS option is in (re)introducing greenery. Parks are popular interventions, as they alleviate the residents’ heat exposure outdoors, improve outdoor thermal comfort12,13,14,15, and even mental wellbeing16. More importantly, studies reported lower air temperature (Ta) within green spaces compared to their surrounding urban areas, up to 5.0 °C in some cases17,18.

The cooling effect of a park is commonly termed ‘Park Cool Island’ (PCI). The PCI intensity is strongest during daytime periods of clear, sunny conditions under maximum solar radiation. This is because tree shade and evapotranspirative cooling reduces the park’s Ta and mean radiant temperatures, as compared to the surrounding urban areas19. PCI is determined by Land Surface Temperature (LST) or Ta differences between park and non-park (urban) sites. LST data is collected through remote sensing techniques20. In this approach, LST and other vegetation indices, such as normalised difference vegetation index (NDVI), are commonly correlated to monitor park cooling performance on a larger scale (for instance, local or city scale)13,15. For better representation of cooling experience by the individual at all times of the day, deploying field measurements (FM) (in-situ or mobile)21,22,23 to collect empirical Ta data would offer the longitudinal and real-time monitoring of the PCI at the canopy level.

A literature search was conducted to examine the cooling performances of parks through FM, using the keywords “park” and “cool island”. The summary of the papers based on their study sites’ Köppen-Geiger climate classification, study period, FM method, PCI equation used, and PCI results are presented in Table 1.

Table 1 Summary of papers using FM to quantify PCI.
Full size table

How the PCI is quantified depends on how the thermal differences between urban and rural Ta are observed44. A direct way to understand PCI’s effectiveness is to quantify its intensity along diurnal changes7. PCI intensity is derived from the difference of Ta between the urban and park measurements. Several studies calculated their PCI using Eq. (1), where cooling differences are accounted for between the park and its immediate vicinity outside of the park or its surrounding urban spaces (Table 1). These studies adapted their calculations by referencing Turban – rural18, which enables the examination of a green area’s UHI mitigation potential across the micro-, local- or regional-scale.

$$PCI [^\circ C] ={T}_{a \,average \,(urban \,or \,outside\, park)}- {T}_{a \,average\, (inside \,park)}$$
(1)

where PCI refers to the average Ta difference between the urban and park area. \({T}_{a \,average \,(urban \,or \,outside \,park)}\) refers to the average Ta in the urban areas or just outside of the park and \({T}_{a\, average\, (inside \,park)}\) refers to the average Ta within the park.

An alternative way to quantify PCI is to calculate the park’s maximum cooling, also known as PCI intensity. It can be derived by subtracting the minimum Ta of the park site from the maximum Ta of the urban site, as shown in Eq. (2). This equation focuses on the maximum cooling of the studied park compared to its surroundings14.

$$PCI intensity [^\circ C]= {T}_{a \,maximum (urban) }- {T}_{a\, minimum (park)}$$
(2)

where PCI intensity refers to the maximum Ta difference between the urban and park area. \({T}_{a \,maximum (urban)}\) refers to the maximum Ta in the urban areas and \({T}_{a \,minimum (park)}\) refers to the minimum Ta in the park.

A third way to quantify PCI is to examine hourly warming/cooling rates using Eq. (3)19. This represents the different rates of near-surface- or urban canopy-level warming and cooling7.

$${\Delta T}_{a(x)}= {T}_{a (x)}- {T}_{a (x-1)}$$
(3)

where \({\Delta T}_{a(x)}\) refers to the change in Ta in a given hour, indicative of the hourly warming/cooling rate. \({T}_{a (x)}\) and \({T}_{a (x-1)}\) refer to the Ta in a given hour and the hour prior, respectively.

This study used a network of 18 fixed meteorological stations to measure, analyse, and map the PCI effect in an urban park in tropical Singapore. The PCI is presented in the following ways: 1) diurnal Ta and dew point temperature (DPT) profiles of a park to its surrounding urban areas; and 2) modifying PCI calculation method to map the PCI’s spatial extent using Geographical Information Science (GIScience). This would enable an assessment of a park’s ability to regulate temperature in a tropical city.

Study design and methodology

Study area

Singapore is an island city-state (1.20° N, 103.49° E) located in the equatorial region. It has a Tropical Rainforest (Af) climate according to the Köppen-Geiger climate classification45. In 2022, Singapore experienced a triple year La Niña and a positive Indian Ocean Dipole event. The mean annual Ta was 27.9 °C, 0.1 °C above the 30-year average 1991—2020. Diurnal Ta ranges from 23.0—25.0 °C (night) to 31.0—33.0 °C (day). The mean annual rainfall was 3012 mm, 478 mm above the 30-year annual average46. Climatically, Singapore is in a maritime continent region with generally low surface wind speeds of approximately 2.1 ms-1 across the year47. The island experiences two distinct monsoon seasons — a wetter Northeast (December – early March) and drier Southwest (June – September) — separated by two intermonsoon periods. Higher surface wind gusts are typically recorded during the monsoons. During the intermonsoon periods, wind direction is more variable, and wind speeds recorded are lower than in monsoon seasons48.

Bishan-Ang Mo Kio (BAMK) Park is one of the largest public greenspaces and the largest community park in Singapore (62 ha)49 (Fig. 1a). The park contains a variety of facilities and landscapes, such as the Inclusive Playground and Landscape Pond, which support community activities and biodiversity. It was first developed by Singapore’s Public Utilities Board in 1988 as a green buffer between the two high-density neighbourhoods with ~ 250,000 residents living along a concrete drainage viaduct channel of the Kallang River (Fig. 1b and c)50. A 3-km stretch of the channel was later re-naturalised, via water-sensitive urban design principles locally known as the Active, Beautiful, Clean Waters Programme51,52. The project’s main function is to resolve flooding, drought, and water pollution arising from rapid urbanisation53. BAMK Park also provides temperature regulatory services through shading and evapotranspiration. This linear park is split into two by a large road, to enable access between the housing estates.

Fig. 1
figure 1

Map of Singapore, BAMK Park and its urban morphology. (a) Map of Singapore, with the BAMK Park study area highlighted in red, (b) satellite image of BAMK Park study area (extracted from Google Earth Pro 7.3.6.9345)55 and (c) study area on ArcGIS Pro with HOBO sensor locations, tree locations and building heights.

Full size image

Data collection

A network of 18 Honest Observer by Onset (HOBO) sensors (MX2301A Ta/RH data loggers) was deployed with RS1 radiation shields32 in the study area from January to December 2022 (H01 – H18 in Fig. 1c). The sensors were secured onto lamp posts at heights of 1.8 – 2.0 m above ground level. While this is higher than the human core height of 1.1 – 1.3 m, the sensors were deployed higher to comply with safety concerns from the site study managers. This also corresponds with the sensor deployment guidelines established by Oke54. Within the park, 10 HOBO sensors were placed (H01 – H10), while eight were deployed in the surrounding urban areas (H11 – H18) (Fig. 1c).

Figure 2 presents the fisheye images at each site and their corresponding Sky View Factor (SVF) values. Fisheye images were taken using a Canon EOS 5DS R, with a Canon EF 8-15 mm F4L USM Fisheye lens at ground level. The image taken provides a 360° field of view of each site. SVF, ranging from 0 to 1, measures the openness of the sky to radiative transport, where 0 (complete obstruction) means that all outgoing radiation will be intercepted by obstacles and 1 (no obstruction) means that all radiation will reflect freely to the sky55,56.

Fig. 2
figure 2figure 2

Sky View Factor photos and values of deployed HOBO sensors across the study area.

Full size image

The HOBO sensors were deployed with RS1 radiation shields and measured point data of Ta and RH at a one-minute interval regardless of wet-weather conditions. Prior to the deployment, all 18 sensors were tested simultaneously in the same environment. The variance of Ta and RH measurements across all sensors was ± 0.2 °C, which is within the accuracy specifications (Table 2).

Table 2 Specifications of HOBO sensors used with RS1 Shield57.
Full size table

Although the HOBO measures RH directly, our study analysed dew point temperature (DPT) instead as a measure of atmospheric moisture content (Eqs. 4a, 4b and 4c). RH is dependent on Ta, and the relatively small variation in Ta (as compared to other climatic regions) inhibits our full understanding of the changes in air moisture content58. On the other hand, DPT, which is the Ta at which a fixed volume of air needs to be cooled at constant pressure to reach full saturation7, is dependent on air vapour moisture and has residual correlation with Ta58. The minimal elevation difference across the study area, and hence negligible differences in vapour pressure, allows DPT to more accurately reflect the actual changes in the atmospheric moisture compared to RH. This makes DPT a more suitable measurement of atmospheric moisture content for this study.

$$DPT [^\circ C]=\frac{-430.22+(237.7 \times ln \left(e\right) )}{\left(-1 \times ln \left(e\right)\right)+ 19.08}$$
(4a)
$$e [\%]=\frac{RH \times es}{100.0}$$
(4b)
$$es= 6.078\times {10}^{\frac{{T}_{a} \times 7.5}{{T}_{a} + 237.3}}$$
(4c)

where \(DPT\) refers to dew point temperature, \(e\) refers to vapour pressure. \(RH\) refers to relative humidity and is calculated using \(es\). \(es\) refers to saturation vapour pressure coefficient. If \(RH=100\), \(DPT\) = \({T}_{a}\). If \(RH\) < 0.000001, \(DPT= -273.0\). The above Eqs. (4) were supplied by ONSET32.

Data analysis

Ta and DPT profile analysis

The “dplyr” and “ggplot2” packages of RStudio Version 4.2.2 were used to calculate hourly averages of Ta and DPT for the study period. Each hourly average is derived from the arithmetic mean of 60 one-minute records at each sensor for each day of the study period. Hourly warming/cooling rate intensities were calculated for every day of the study period using Eq. (3), then aggregated to present an average diurnal profile of 2022. PCI intensity was calculated using Eq. (2), then plotted to present an hourly PCI profile of the study area for the study period.

The dataset was tested for statistical differences between the park (H01—H10) and urban (H11—H18) data. This was first done by conducting a Levene’s test for homogeneity of variances on RStudio using the “car” package. If variances are equal, a Student’s t-test59 was conducted to test for statistically significant differences. Conversely, when variances are different, Welch’s t-test60 was conducted to examine if the two datasets are significantly different. All statistical tests were conducted at 95% confidence interval.

Spatial interpolation mapping

The study’s novelty lies in its modification of Eq. (2) to leverage the network of sensors deployed in the study area. Traditionally, the PCI intensity equation is only able to output one PCI value per study area, where PCI intensity represents the arithmetic mean of maximum temperature differences between the entire park against the surrounding urban area (Table 1). The modification, Eq. (5), calculates the maximum cooling measured at each site relative to the arithmetic mean of maximum urban Ta, which preserves the spatial granularity of the sensor network. The values at each site were then used to visualise the spatial extent of PCI intensity and radiative cooling (RC) in the study site. This was done using ArcGIS Pro Version 3.0.3 (ESRI Inc.), a Geographical Information Systems software. RC in this study refers to the outgoing longwave radiative flux from the surface 61.

The hourly mean maximum urban Ta is the arithmetic mean of hourly maximum Ta of each day during the study period measured by the urban sensors (H11—H18), hereafter referred to as \(mean {T}_{a \,urban\, maximum}\). The hourly mean minimum Ta is the arithmetic mean of hourly minimum Ta of each day during the study period measured at each sensor, hereafter referred to as \({mean T}_{a\, minimum (H01 \dots H18)}\). PCImodified is the difference between \(mean {T}_{a \,urban\, maximum}\) and \({meanT}_{a\, minimum \,(H01 \dots H18)}\) (Eq. (5)). PCImodified intensity was calculated at these selected hours (06:00, 07:00, 13:00, and 19:00 local time GMT + 8 (LT)) to represent the local solar patterns; minimum Ta (06:00—06:59 LT), maximum Ta at solar maximum (13:00—13:59 LT), sunrise (07:00—07:59 LT), and sunset (19:00—19:59 LT).

$${PCI}_{modified} [^\circ C]= mean {T}_{a \,urban\, maximum}- {mean T}_{a \,minimum (H01 \dots H18)}$$
(5)

where PCImodified refers to difference in Ta of each sensor \({mean T}_{a\, minimum (H01 \dots H18)}\) relative to the \(mean {T}_{a\, maximum\, (urban \,sensors\, only)}\) within the study area at a given hour of each day, then aggregated across the study period for the specific hour.

Figure 2 indicated displayed a variance in SVF (0.09 – 0.36) between park and urban sites. The site visits revealed that the predominant urban features distinguishing each site were the relative proportion of buildings and trees. Thus, the study extracted building and tree data from the Singapore’s Urban Redevelopment Authority Master Plan 201962 and National Park Board’s TreesSG63 respectively to account for the building height, tree height, and tree density within the study area. For tree data, only those managed under the National Parks Board were available; this does not account for trees managed by other authorities. A buffer was drawn around each sensor using the Buffer tool, where the mean building height (metres), tree height (metres), and tree density (count per buffer) were calculated (Fig. 3). Different buffer radii of 2 m, 5 m, 10 m, 20 m, 50 m, 75 m and 100 m were tested to determine the optimal radius for the subsequent PCImodified analysis. The 50 m buffer provided the best balance of capturing the urban heterogeneity of the study area while recognising the limits of how much the wider environment will influence sensor measurements. This is also supported by multiple studies 64,65. Jin et al.’s66 study also identified that 50 m buffer is suitable when observing variations in microclimatic in Singapore. Thus, the building and tree vector layers were rasterised to 50 m spatial resolution to match the buffer (Appendix F).

Fig. 3
figure 3

Maps presenting the 50 m buffer zone around urban sensors (H11 – H18).

Full size image

The Ordinary Least Squares (OLS) tool in ArcGIS Pro67 was used to understand the relationship between PCImodified intensity and the distinguishing urban features of building height, tree height and tree density. The OLS calculates the estimated PCImodified intensity and its residuals (difference between measured Ta and estimated Ta), as well as the intercept, and coefficients of building height, tree height, and tree density. The coefficients represent a weightage of each environmental parameter.

The coefficients, intercepts, and OLS model as a whole were tested for statistical significance on SPSS68 to determine the viability of the model; whether urban morphology has an impact on the spatial distribution of PCImodified intensity. Each coefficient was tested with a partial t-test to determine whether each variable had a significant effect on the PCImodified in the context of the entire model. An overall F-test was run to test for the significance of the entire model. The residuals were tested for spatial autocorrelation using the Global Moran’s I tool69,70 in ArcGIS Pro. All residuals were not spatially autocorrelated.

Although Geographically Weighted Regression and Universal Kriging produce accurate interpolation plots71,72, our sample size, as represented by the 18 sensor locations, is too small to utilise these tools. For this case, the Empirical Bayesian Kriging tool73 was used on the OLS residuals to conduct regression kriging to interpolate PCImodified variability74 to produce a kriged residuals layer.

The Raster Calculator tool in ArcGIS Pro was used to produce the Spatial Interpolation Maps (SIMs) based on Eq. (6). Each environmental parameter’s coefficient, applied to their respective raster layers (i.e. building height, tree height, and tree density), together with the intercept and kriged residual layer were input into Eq. (6).

$$SIM= {\beta }_{1}{x}_{1}+ {\beta }_{2}{x}_{2}+ {\beta }_{3}{x}_{3}+c+kriged \,residuals$$
(6)

where \({x}_{1}\), \({x}_{2}\), \({x}_{3}\) represents building height, tree density, tree height respectively, \(\beta\) refers to the coefficient of each environmental parameter, \(c\) refers to the intercept, and kriged residuals represents the difference between actual and estimated values. \(\beta\) and \(c\) were derived from the OLS model, while kriged residuals were derived from inputting the residuals into the Empirical Bayesian Kriging tool.

SIMs for PCImodified, specifically for Ta, at the four selected hours were presented. DPT was not presented, as the small differences (0.01—0.17 °C) do not have real-life implications despite statistically significant differences in urban and park DPT (Appendix D.2).

Results

Air temperature and humidity profiles

Hourly air temperature profiles

The maximum and minimum hourly Ta for the park area is 34.96 °C and 22.44 °C, and 35.70 °C and 22.75 °C for urban areas (Fig. 4; Appendix A). Levene’s test for homogeneity of variances indicated the variances between measured Ta values were not equal (p < 0.001) (Table 3). Hence, a Welch’s t-test was conducted, and results indicated statistically significant differences between Ta in park and urban areas within each hour (p < 0.001) (Appendix D.1).

Fig. 4
figure 4

Boxplots of hourly Ta profiles. The bottom and top edges of the box represent the first and third quartile values, extended vertical lines show values beyond interquartile range, whereas diamonds at the end of each line represent outliers.

Full size image
Table 3 Levene’s Test for Homogeneity of Variances for Ta, DPT and Hourly Warming/Cooling Rate.
Full size table

Figure 4 displayed a similar sinusoidal trend in parks and urban areas. Maximum Ta occurred between 12:00 to 14:00 LT during solar maximum (i.e. at 13:00 LT), where incoming solar radiation is at its maximum. Ta steadily dropped through the night, reaching minimum Ta before sunrise at 06:00 LT. A converse trend was observed from 07:00 LT; Ta increased till solar maximum (Fig. 4).

Hourly warming/cooling rates

The hourly warming/cooling rates were derived for every day of the study period using Eq. (3), then aggregated to the specific hour to present an average diurnal profile of 2022. The diurnal profile exhibited a sinusoidal trend (Fig. 5). Positive rates indicate warming, where an increase in mean Ta relative to the previous hour is observed. Conversely, negative rates indicate cooling, where a decrease in mean Ta is observed. Levene’s test for homogeneity of variances of warming/cooling rates was conducted, and the variances were equal (p = 0.28) (Table 3). Student’s t-test was then used to determine statistical difference between the park and urban areas in each hour. Mean Ta for both areas in most hours did not exhibit any statistically significant differences (p > 0.05); only 19:00 LT exhibited significant difference (p = 0.02*). (Appendix D.3).

Fig. 5
figure 5

Notched boxplots for hourly warming/cooling rates of mean Ta for sensors situated in park and urban areas. The bottom and top edges of the box represent the first and third quartile values, extended vertical lines show values beyond interquartile range, whereas diamonds at the end of each line represent outliers.

Full size image

Median warming/cooling rates for all sites are positive starting from 07:00 LT and peaked at 09:00 LT (Fig. 5; Appendix C). Peak median warming rates were greater in the park (1.21 °C) compared to the urban areas (1.13 °C). The urban areas’ median warming rate were consistently higher than the park’s from 11:00 LT to 14:00 LT. From 16:00 LT, cooling is observed for both areas, with park areas experiencing a greater magnitude of cooling from 18:00 LT. Maximum cooling magnitude observed at 19:00 LT; park (-0.61 °C) compared to urban (-0.50 °C) (Fig. 5).

Hourly dew point temperature profiles

Hourly DPT were higher in the park than urban areas, with more pronounced differences observed during daytime hours (07:00 to 19:00 LT). Both areas reached a maximum median DPT of 24.93 °C (park) and 24.75 °C (urban) at 11:00 LT (Fig. 6; Appendix B). Levene’s test for homogeneity of variances indicated the variances between measured DPT values were not equal (p < 0.001) (Table 3). Welch’s t-test was then conducted and indicated that there were statistically significant differences in park and urban DPTs for each hour (p < 0.001) (Appendix D.2).

Fig. 6
figure 6

Boxplots of hourly DPT profiles. The bottom and top edges of the box represent the first and third quartile values, extended vertical lines show values beyond interquartile range, whereas diamonds at the end of each line represent outliers.

Full size image

PCI intensity and spatial interpolation maps

PCI intensity

The hourly time-series of mean PCI intensities were calculated by subtracting the mean maximum urban Ta by the mean minimum park Ta (Eq. (2)). Mean PCI intensities and their corresponding standard error means (s.e.m.) were plotted across the study area (Fig. 7; Appendix E).

Fig. 7
figure 7

Mean hourly profile of PCI intensity. Sunrise and sunset are marked with a vertical dotted line at 07:00 and 19:00 LT respectively. Red error bars represent the standard error means per hour.

Full size image

The study area observed a mean daytime PCI intensity of 2.21 °C (07:00 – 18:00 LT) and a mean nighttime PCI intensity of 1.69 °C (19:00 – 06:00 LT). PCI intensity increased from 06:00 LT (1.63 \(\pm\) 0.033 °C; mean PCI intensity \(\pm\) s.e.m.), peaking at 08:00 LT (2.54 \(\pm\) 0.051 °C) before declining till 22:00 LT (1.68 \(\pm\) 0.036 °C) with some fluctuations in between. PCI intensity was lower at night till sunrise (19:00 – 06:00 LT). Minimum PCI intensity was observed both at 05:00 and 06:00 LT (1.63 \(\pm\) 0.033 °C) (Fig. 7; Appendix E).

Spatial interpolation maps

SIMs were generated to observe the PCImodified at selected times of the day along the solar patterns (06:00, 07:00, 13:00 and 19:00 LT), based on Eq. (4). When mapping PCImodified onto the study area, we can infer the difference between RC of park and urban surroundings. PCImodified values close to zero indicate low PCI intensity with weak RC differences between the urban and park areas.

At 06:00 LT, the PCImodified is close to zero in most of the study area, particularly in urban sites and eastern parts of BAMK Park (Fig. 8a), indicating that the difference in RC across the study area is minimal. At sunrise (07:00 LT), the sun starts to heat the surfaces throughout the study area. As a result, the RC is higher, especially in the park, as seen in PCImodified being higher than its surrounding urban areas (Fig. 8b). At solar maximum (13:00 LT), PCImodified is at its maximum due to maximum thermal mass differences between the urban and park areas (Fig. 8c). However, the PCImodified intensity is consistent both in the park and urban areas. At nighttime (19:00 LT), the PCImodified is lower compared to solar maximum (13:00 LT) (Fig. 8d). However, PCImodified intensity is not consistent across the study area, with park areas exhibiting higher PCImodified, thus higher RC, than urban areas.

Fig. 8
figure 8

Spatial Interpolation Maps representing hourly PCImodified for (a) 06:00, (b) 07:00, (c) 13:00 and (d) 19:00 LT. Scales are kept consistent across all maps in the figure.

Full size image

PCImodified was created based on OLS regression that considered the following parameters (building height, tree density and tree height) within the buffer of 50 m. The coefficients (Table 4) derived from the OLS models show the weightage of each environmental parameter in determining PCImodified. The positive relationship of the tree density and tree height against PCImodified remains relatively consistent throughout the day. The building height had a negative linear relationship with PCImodified throughout the day.

Table 4 Intercepts, coefficients and relevant statistics for environmental parameters against PCImodified.
Full size table

All environmental parameters did not show statistical significance in their individual partial t-test results (p > 0.05). The overall F-test still showed that OLS models that incorporated the three environmental parameters for 06:00, 07:00, and 19:00 LT were statistically significant (p < 0.05) (Table 4). This showed that considering the three environmental parameters together would have a significant impact on PCImodified distribution. Conversely, both the partial t-test and overall F-test for 13:00 LT was not statistically significant (Table 4). The non-statistically significant results for the 13:00 LT SIM indicated the urban morphological parameters had negligible impact on PCImodified distribution.

Discussion

This study aimed to measure the diurnal Ta and DPT profile of park and urban areas, and to map PCI intensity and spatial extent using GIScience. Despite park and urban areas displaying similar sinusoidal trends (Fig. 4), the results show that there were statistically significant differences in Ta between both areas at every hour (Appendix D.1). As explained by other PCI studies (Table 1), this can be attributed to the differences in the study areas’ physical characteristics/morphology, namely, percentage of tree cover, thermal properties of building materials, and sky view factor.

Figure 5 also exhibits a sinusoidal trend in warming/cooling rate. The findings corroborate with the existing literature, with the maximum warming/cooling magnitude observed during 07:00 to 09:00 LT at sunrise, and from 17:00 to 19:00 LT at sunset. These differences were the highest when the sun’s zenith angle was low (i.e. sunrise and sunset)7 . Urban areas exhibited consistently higher warming rates compared to the park from 11:00 LT to 14:00 LT. From 18:00 LT, urban areas experienced a smaller magnitude of cooling. The difference in warming/cooling performance can be explained by urban materials like concrete and asphalt having higher thermal admittance, leading to greater heat absorption and slower release7,75, thus affecting near-surface Ta.

Higher DPT values were observed within park spaces compared to urban areas, due to evapotranspiration from vegetation in the park and the evaporation from park water bodies (e.g. Kallang River, Landscape Pond, and others) (Fig. 6). While initial statistical analyses revealed park and urban areas measured significantly different DPT values, the mean difference of DPT between park and urban areas only ranged from 0.01 to 0.17 °C (Appendix D.2). The higher atmospheric moisture content in parks compared to urban areas is negligible, and may not be felt by the average park goer.

Utilising FM, many studies observed a significant PCI effect from parks, making them a useful NbS for temperature regulation in urban areas (Table 1). This concurs with the findings of this study, where PCI was observed throughout the day, with a maximum PCI intensity of 2.54 \(\pm\) 0.051 °C at 08:00 LT. Diurnal variation of PCI intensity indicates its cooling intensity was most pronounced in the daytime, while cooling intensities were smaller from nighttime (1.74 \(\pm\) 0.035 °C at 20:00 LT) to before sunrise (e.g. 1.63 \(\pm\) 0.033 °C at 06:00 LT) (Fig. 7; Appendix E). Previous studies on greenspace and cooling also found Ta reductions within a park and some advection to its surrounding peripheries14,18,44,76, with PCI intensity being strongest in the hours of direct solar exposure in the daytime as opposed to nighttime24,32,39. This is due to the presence of vegetation in parks that provide evapotranspiration cooling, as well as park spaces not heating up as much as impervious urban surfaces, especially around solar maximum77.

Through PCImodified (Eq. (6)), this study introduced a novel method to visualise PCI across a study area, as shown in the SIMs (Fig. 8). In Figs. 8a,b,d, PCImodified was higher in park areas than in urban areas, indicating higher RC in the park. This result strongly illustrates the effectiveness of this NbS in regulating urban overheating during specific hours of the day. However, this same effectiveness was not observed during solar maximum at 13:00 LT (Fig. 8c). Specifically, there was little variation in RC between the park area and urban areas, as observed by the higher but less varied PCImodified across the study area. In addition, statistical tests of the OLS model for the 13:00 LT SIM showed no statistical significance (Table 4), indicating that environmental parameters (e.g. buildings, trees) have little to no impact on Ta. As such, the effectiveness of parks as a NbS in temperature regulation is limited during solar maximum.

The methodology presented in this paper leveraged the microclimatic data from all measurement points within the network of low-cost sensors to produce SIMs to visualise the intensity and spatial extent of park cooling (Fig. 8), with relevance to the environmental parameters (buildings, trees). The SIMs are based on the OLS regression model which determines the relationship between building height, tree density and tree height and PCImodified intensity. The regression coefficients (Table 4) present a negative relationship between building height and PCImodified intensity across the day. As such, it is observed that the taller the buildings, the lower observed RC. Tree height and tree density had a positive relationship with PCImodified intensity throughout the day. As such, the greater the tree density and height, the higher the PCImodified intensity, showing that the park is effective in regulating Ta. However, these relationships were only significant in the 06:00, 07:00 and 19:00 LT plots (Fig. 8a,b,d) (Table 4). As the OLS model was not significant for the 13:00 LT dataset, it can be concluded that the study area’s environmental parameters have a negligible impact on PCImodified. Figure 8c also reflects this deduction, as PCImodified distribution was mostly uniformed throughout the study area, with no distinct differences between the park and urban areas. This highlights how the OLS model and SIMs can draw statistical links between environmental parameters and cooling performance of the park at different times of the day.

The spatial extent of park cooling beyond the boundaries of the park is limited as seen in Fig. 8. It can also be observed that for the 06:00, 07:00, and 19:00 LT, that the PCImodified intensity differs between the urbanised areas surrounding the park, with the west being cooler. This could be explained by the Lower Pierce Reservoir Park, a mature secondary rainforest that surrounds a reservoir, located to the west of the study area78. The presence of the reservoir seems to also explain the spatially uneven RC in the study area (Fig. 8). It is also observed that there is little to no RC discernible in the urban areas where there are relatively taller and denser buildings, as indicated by the low PCImodified value. This is particularly so in the eastern side of the urban areas with building heights above 30 m. In all, the OLS model and SIMs can aid urban planners and decision-makers in understanding a park’s effectiveness for temperature regulation across different times of the day79,80,81. Naturally, this method is highly dependent on the data availability, data resolution, and available preprocessing tools.

This campaign did not have a full choice on the sensor sites, as the areas had high foot traffic. Hence, the study was unable to control for the characteristics of the individual sites such as sky view factor, building surface fraction, ratio of pervious to impervious surfaces, and distance to water bodies, characteristics that could affect the microclimatic conditions17,82. Indeed, based on Fig. 2, the SVF values, both within and between park and urban areas, displayed a degree of heterogeneity. As such, the scope of this paper considered the aggregation of urban (H11—H18) versus park (H01—H10) sensors rather than the comparison between individual sensors. It is also important to note that the results presented are based on aggregated data measured for the study period. This represents the synoptic conditions of 2022. Moreover, the inaccessibility to high-resolution Digital Terrain Model and Land Use/Land Cover data also inhibited the inclusion of corresponding environmental parameters into the OLS and SIMs.

Singapore’s equatorial location also hinders the use of Remote Sensing data; the low frequency of satellites passing over the island, along with Singapore’s consistently high cloud cover, prevents the acquisition of clear hourly snapshots of the study area’s surface temperature. Thus, this method presents a low-cost alternative for long-term monitoring and analysis of the microclimate, particularly for regions with limited access to high spatiotemporal-resolution microclimate remote sensing data.

Conclusion and recommendations

This study demonstrates the use of FM (HOBO sensors network) to map and analyse the hourly Ta and DPT between park and residential urban areas in the tropical city-state of Singapore. The use of HOBO sensors network allows for real-time monitoring of hourly Ta and DPT throughout the day. Based on these measurements, the warming/cooling rates of each site were analysed. These measurements were used to understand the microclimatic differences across the study area, and ascertain the diurnal effectiveness of this NbS in regulating urban temperatures for a large park.

A new method (PCImodified) was presented to measure and map the cooling performance of urban vegetation in a tropical city. PCImodified identifies differences in RC and thermal mass between the park and residential urban areas, while preserving the high spatial and temporal resolution of the measured data from the sensor network (park and urban area). Therefore, the results can be effectively visualised using Geographical Information Systems software, overcoming the inaccessibility to satellite-based remote sensing data.

The results of this study suggest that FM is useful in analysing PCI, and by extension, effectiveness of NbS in temperature regulation at different times of the day. PCImodified can be utilised in future studies by urban planners, climatologists, and policymakers to better understand and map the effect of vegetative cooling. Future studies should apply this methodology to other forms of vegetation in and beyond tropical cities. Long term FM studies should be conducted to understand the influence of seasons83 and regional climate forces (e.g. El Niño-Southern Oscillation) in affecting PCI and the effectiveness of parks in alleviating urban overheating. A larger dataset spanning across multiple years will allow for subsequent insights on the effects of seasonality and exceptionally hot or cool days on PCI.