Lack of vegetation exacerbates exposure to dangerous heat in dense settlements in a tropical African city

Both climate change and rapid urbanization accelerate exposure to heat in the city of Kampala, Uganda. From a network of low-cost temperature and humidity sensors, operational in 2018–2019, we derive the daily mean, minimum and maximum Humidex in order to quantify and explain intra-urban heat stress variation. This temperature-humidity index is shown to be heterogeneously distributed over the city, with a daily mean intra-urban Humidex Index deviation of 1.2 ∘C on average. The largest difference between the coolest and the warmest station occurs between 16:00 and 17:00 local time. Averaged over the whole observation period, this daily maximum difference is 6.4 ∘C between the warmest and coolest stations, and reaches 14.5 ∘C on the most extreme day. This heat stress heterogeneity also translates to the occurrence of extreme heat, shown in other parts of the world to put local populations at risk of great discomfort or health danger. One station in a dense settlement reports a daily maximum Humidex Index of > 40 ∘C in 68% of the observation days, a level which was never reached at the nearby campus of the Makerere University, and only a few times at the city outskirts. Large intra-urban heat stress differences are explained by satellite earth observation products. Normalized Difference Vegetation Index has the highest (75%) power to predict the intra-urban variations in daily mean heat stress, but strong collinearity is found with other variables like impervious surface fraction and population density. Our results have implications for urban planning on the one hand, highlighting the importance of urban greening, and risk management on the other hand, recommending the use of a temperature-humidity index and accounting for large intra-urban heat stress variations and heat-prone districts in urban heat action plans for tropical humid cities.


Introduction
Heat is a killer hazard with a global reach. Its exposure has been associated with both increased mortality and morbidity worldwide (Medina-Ramón and impact of extreme heat concentrates on mid-latitude, high-income countries of low to medium population density (Campbell et al 2018, Green et al 2019, thereby chronically underreporting regions that are projected to actually experience the most extreme heat in the future (Im et al 2017, Mora et al 2017, Nagendra et al 2018, Harrington and Otto 2020, Saeed et al 2021. For example, Africa is particularly vulnerable to heat stress (IPCC 2014, Singh et al 2019. A rapid increase in the intensities and frequencies of heatwaves during the past decades has been demonstrated (Ceccherini et al 2017, Amou et al 2021, while simulations project this trend to continue uninterrupted into the future (Harrington et al 2016, Russo et al 2016, Dosio et al 2018. For instance, under a higher emission scenario (SSP5-8.5), Africa's exposure to extreme heat is projected to be 7-269 times larger than it has been historically (Liu et al 2017, Asefi-Najafabady et al 2018).
Extreme heat is further amplified in cities, which are shown to be warmer than their natural surroundings, because of reduced vegetated areas, increased release of anthropogenic heat, changes in surface albedo and trapped radiation within street canyons (Oke 1982). This urban heat island effect has also been demonstrated for Sub-Saharan African cities (Nakamura 1966, Jonsson et al 2004, Roth 2007  As an example, Kampala, the capital city of Uganda, is experiencing an uncontrolled urbanization, having the fourth highest growth rate (>4% yr −1 ) of all African cities (Richmond et al 2018, Kampala Cap-ital City Authority and Uganda Bureau of Statistics 2019). Like many fast-growing cities, Kampala is expanding horizontally (Brousse et al 2019, Li et al 2021, demonstrating spatial patterns of urban sprawl (Vermeiren et al 2016, Hemerijckx et al 2020 and the formation of informal settlements or slums (Van Leeuwen et al 2017, Lwasa et al 2018, Richmond et al 2018. Within the city of Kampala, both morphological and socio-economical characteristics largely differ, distinguishing wealthy districts characterized by asphalted roads, modern houses and large gardens, from informal settlements composed of densely built shacks made of corrugated metal sheets that are only accessible via small alleys (Vermeiren et al 2012, Hemerijckx et al 2020. Recently, Brousse et al (2019) classified these intra-urban variations into Local Climate Zones (LCZ, Stewart and Oke 2012). This classification includes 7 vegetated and 10 built classes, each class exemplifying uniform surface cover, structure, material and human activity that span hundreds of meters to several kilometers in horizontal scale (Stewart 2011). Importantly, LCZs are designed to reflect the thermal environments as a consequence of their intra-urban variations. LCZ are thus expected to also reflect heat stress variations in the city (Kabano et al 2021, Van de Walle et al 2021), similar to the findings in Nairobi (Kenya), concluding that informal settlements are particularly prone to heat stress (Scott et al 2017).
However, observational studies investigating this heterogeneity have been depreciated, because of the characteristic meteorological data scarcity in the region (Roth 2007). Six weather stations were set up in Kampala only recently, thanks to the Trans-African Hydro-Meteorological Observatory (TAHMO, van de Giesen et al 2014) project, collecting meteorological data from the synoptic station at Makerere university and five instrument shelters placed in open school gardens, in accordance with the official World Meteorological Organization standards (WMO 1986). Despite this great observational effort, no stations are placed in more densely built environments where most of Kampala's population lives.
To better represent the variations of heat stress throughout the city of Kampala, including densely populated areas, this study put in place an observational network of 45 low-cost iButton sensors. These sensors recorded near-surface air temperature and relative humidity for three 42 d periods between August 2018 and April 2019 (Van de Walle et al 2021). From these measurements, the Humidex Index is computed, providing a good estimate for feel-like temperature (Masterton and Richardson 1979). High relative humidity decreases a person's evaporation ability and thereby the effectiveness of the body's natural cooling system (Malchaire et al 2000, Hass et al 2016. Particularly in hot and humid cities like Kampala, high values of the Humidex Index might cause dangerous health conditions. We therefore focus on extreme heat recorded at the different stations, and explain the observed patterns based on relevant satellite-derived earth observations. For example, vegetation is known to generally play a twofold role, decreasing temperature but enhancing humidity by transpiration (Hass et al 2016). Results are discussed from two different perspectives: insights in spatial heterogeneity of heat in Kampala and occurrences of heat above great discomfort thresholds among different urban environments.

iButton observations
The iButton sensor, a product of Maxim Integrated, is a low-cost sensor containing a temperature and humidity logging system (Hubbart et al 2005). With a logging frequency and data accuracy programmed at 15 min and 11 bit respectively, each sensor can store 42 consecutive days of data. Afterwards, a manual download is required. To protect the sensors from radiation and splash water, they are shielded by a folded thin light reflective film (figure S1 available online , Nakasero (Ns), Industrial area (Ia), Nkeere (Nk) and Buziga (Bz). A set of three sensors is installed per location (at the exact centre of each top view image), reducing the uncertainty due to different installation conditions. Exact locations of the sensors are listed in table S1. Each top view image is 500 × 500 m 2 , retrieved from Google Earth imagery. The land surface temperature estimation for the cloud-free day of 27 February 2021 is derived from Landsat 8 and MODIS remote sensing products via Parastatidis et al (2017), and shows large intra-urban variations of surface temperatures. at stacks.iop.org/ERL/17/024004/mmedia), designed and produced by the Maryland Institute College of Art and Johns Hopkins University. The sensorcontaining shields are zip-tied preferably to a wooden material, at about 2 m height and at a shaded location (table S1). In this study we reused the material of the observational campaign by Scott et al (2017) held in Nairobi, Kenya.
The resulting network of 45 sensors installed throughout the city of Kampala (figure 1) aimed at properly representing the city's surface heterogeneity. Besides a good spatial coverage of both openly and densely built environments, the choice of sensors location also considered the security of the sensors from vandalism. Preferred locations thus included schools or houses of local acquaintances. At each location, a set of three sensors were installed close together, reducing the uncertainty of the triple-sensor-mean due to different installation conditions such as attachment material, shade fraction or ground cover. For each location, the 15 min resolution triple-sensor-mean information is reduced to the minimum and maximum values per day. In addition, the daily mean is computed as the average over 24 h. Three downloading rounds provided data for 3 periods of 42 d each between August 2018 and April 2019 (figure S2). For further analyses, daily minimum and daily maximum values are computed as the averages over all 3 × 42 d, still referred to as daily minimum and daily maximum values. These periods do not cover a full year, yet we monitored humidity and temperature for both dry and wet seasons. The intra-seasonal variations in temperature and humidity are generally rather limited in the region (figure S3). On a longer timescale, warm spells can occur, with the start of 2019 as an example. The second of our observational periods covers that period ( figure S3). Due to a slightly different downloading moment, measurement periods may slightly differ for locations located far from each other. In addition, some data is missing for the second and third periods due to technical issues, especially low battery. At the end of the third period, 32 out of 45 sensors remained active.
In addition, one sensor set was put indoor at the informal settlement of Acholi Quarters. The building is a concrete block, no windows, an iron door and iron roof sheets. Given the fact that it is only one site, results are qualitatively described in the discussion section.

Measurement quality
The manufacturer's evaluation of the iButtons reports a thermochron and hygrochron accuracy of 0.5 • C and 5% respectively. This is confirmed by calculating the mean deviation per sensor triple, explained and summarized in table S2. In addition, the triple-sensor-mean is evaluated against the Makerere automatic weather station data, part of the TAHMO network (van de Giesen et al 2014). With an overall temperature bias of 0.10 • C, 0.11 • C and 0.53 • C in the three periods, and relative humidity bias of −2.50%, −2.66% and −1.43%, the iButton sensors tend to slightly overestimate the air temperature and underestimate humidity. This however varies throughout the day as observed nighttime temperatures by the iButton sensors are higher than the ones measured by the automatic weather station, while observed daytime temperatures are lower ( figure S4). This results in an underestimation of the diurnal temperature range. Nighttime relative humidity is underestimated by ∼4%, but daytime observations compare well to the Makerere automatic weather station.

Humidex index
To better estimate the human-experienced heat, the Humidex Index (hereafter referred to as 'Humidex' , H) is computed every 15 min from observations of both temperature (T in • C) and relative humidity (RH in %), following Masterton and Richardson (1979): ) . (1) The resulting quantity increases non-linearly with both air temperature and relative humidity, and can be understood as feel-like temperature in degrees Celsius. Humidex values above 40 • C lead to 'great discomfort' , values exceeding 45 • C are 'dangerous' (Masterton and Richardson 1979). Humidex information is reduced to the minimum and maximum values per day, as well as the daily mean which is the average over 24 h.

Explanatory variables
We aim to explain spatial Humidex patterns by comparing them against potential explanatory factors, including distance to the lake, vegetation presence, built-up fraction, surface elevation, population density and surface albedo ( figure S5). The choice of each of these six factors is argued below. First, located next to Lake Victoria, Kampala is affected by a daytime lake breeze developing on a daily basis (Thiery et al 2015, 2016  , is translated to a 30 m resolution grid. Sixth, the MODIS instrument also provides directional hemispherical (blacksky) near-infrared albedo at 0.7-5.0 µm wavelength at 500 m horizontal resolution, possibly distinguishing different roofing types within the city of Kampala (Brest 1987). If these variables can explain the observed Humidex variations, a simple statistical model could extrapolate operational weather station data to the entire city, providing heat stress information about hardly accessible locations such as informal settlements. Assuming linear behaviour, a multiple linear regression technique is applied, expressing the Humidex (H) in terms of the explanatory variables x i : where the Ordinary Least Squares method estimates the best fitting coefficientsβ i based on the observations. Initially, all N explanatory variables are included, but a t-test decides on the elimination of the least significant variable. This backward elimination continues iteratively until all explanatory variables are significant at 95% confidence level. Ultimately, the remaining multiple linear regression model no longer suffers from multicollinearity, ensuring its coefficients to be optimally stable (Halinski and Feldt 1970).

Local climate zones
In addition to this quantitative approach, locations are classified based on the building structure, land cover and human activity according to the LCZ classification scheme ( 5), characterized by open arrangement of 3-9 story buildings and abundant plant cover. The Industrial Area (Ia) station is located in a large low-rise (LCZ 8) class, characterized by extensive paved surfaces between large, low buildings, often with an industrial or commercial function. The Nakasero (Ns) station is located in compact mid-rise (LCZ 2) class, defined by buildings of 10-25 m separated by narrow streets and inner courtyards, and with few or no trees. In addition, we classify the stations in Namungooma (Ng), Bwaise (Bw) and Najjanankumbi (Nj) as compact low-rise (LCZ 3) and Nateete (Nt), Nkeere (Nk) and Acholi Quarters (Aq) as lightweight low-rise (LCZ 7), often called informal settlements or slums. Both classes consist of small buildings tightly packed along narrow streets with no or little vegetation. Typical for the latter class are the lightweight building materials (thatch, wood or corrugated metal) and often formless arrangement of the buildings (Stewart 2011).

Results
Measurements show clear differences in Humidex values at the different sensor locations (figures 2(g)-(i)). For example, the average Humidex varies between 30.6 • C and 32.1 • C for the urban stations, except for the Makerere (Mk) station, located near the city centre, with a substantially lower average Humidex (29.0 • C). Also the city's outskirts, represented by Buloba (Bl), Bukerere (Bk) and Kawanda (Kw) stations, are cooler (29.3 • C-29.9 • C, figure 2(h)). At night (figure 2(g)), both the city outskirts and Makerere are cool, in contrast with high Humidex values observed nearby central stations. Particularly the industrial area (Ia), Nakasero (Ns), Namungoona (Ng) and Nkeere (Nk) experience warm nights, with differences up to 2.3 • C compared to Makerere. The intra-urban heterogeneity is most pronounced when comparing daily maximum Humidex values (figure 2(i)), ranging from 34.1 • C at Makerere to 41.6 • C at Najjanankumbi.
The observed Humidex heterogeneity is a result of intra-urban temperature and relative humidity variations. The latter are considered at times of daily minimum and maximum Humidex (figures 2(a), (d) and (c), (f) respectively). Around sunrise, when Humidex reaches its minimum, the air temperature almost entirely determines the Humidex variations between the sensor sites, with the urban air being clearly hotter and drier than at the outskirts (figures 2(a), (d) and (g)). Specific humidity between those sites is similar (not shown). Around noon, when Humidex reaches its maximum, the situation is different. Then, the highest air temperatures are observed at Nateete (Nt), Nkeere (Nk) and Acholi Quarters (Aq, figure 2(c)), all classified as lightweight low-rise LCZ 7. These locations also have the lowest relative humidity (figure 2(f)). While the temperature at Najjanankumbi (Nj, compact low-rise LCZ 3) is very similar to the temperatures at Nateete or Nkeere, it has a clearly higher Humidex (figure 2(i)). This can be explained by Najjanankumbi's high relative humidity compared to Natalee and Nkeere. We therefore need both temperature and relative humidity to properly explain spatial variations in heat. Importantly, neither Najjanankumbi's air temperature or relative humidity is exceptional compared to other stations such as Nkeere or Makerere: the temperature distribution is similar to the one at Nkeere (figure S6(c)), and relative humidity values exceeding 70% occur as frequent at Makerere (figure S6(f)). Instead, it is the combination of both compound drivers that creates high maximum Humidex values in Najjanankumbi (figure S6(i), Zscheischler et al 2018Zscheischler et al , 2020.
Relating the observed Humidex heterogeneity to the LCZ classification, the open low-rise (LCZ 6) environments, together with the open mid-rise Makerere University campus (LCZ 5), report generally lowest temperature, highest relative humidity and lowest Humidex values (figure 2). These cool environments contrast with the warm compact and lightweight low-rise classes (LCZ 3 and LCZ 7), as well as the compact mid-rise central business centre (LCZ 2). Kampala's industrial area (Ia, LCZ 8) is rather warm at night, but shows relatively mild temperatures during daytime. Especially poorly vegetated and compactly built neighbourhoods in Kampala are thus more prone to heat stress than the outskirts or Makerere station.
Not only for daily maximum Humidex, also for daily minimum or average Humidex, Nkeere (Nk), Acholi Quarters (Aq) and Namungoona (Ng) show highest exceedance frequencies of all Humidex thresholds (figure S7). When also accounting for the duration of exceedance by considering the hours above a certain threshold ( figure S8, top), or for the heat intensity by computing the mean heat-degreehours ( figure S8, bottom), the densely built environments often experience extreme heat (1%-10% of the observed time), especially in Najjanankumbi (Nj).
The intra-urban heterogeneous Humidex values correlates with six proposed explanatory variables (figure S9). Strongest correlations are found between Humidex and NDVI as well as ISA fraction. Moderate negative correlations of −0.5 to −0.65 are found between Humidex and near-infrared albedo, while Humidex and population density are positively correlated, with values between 0.5 and 0.76. Smaller correlations are found between Humidex and the proximity to the lake or elevation. Importantly, the explanatory variables are not independent from each other, with high ISA fractions prohibiting abundant vegetation (correlation of −0.93) while also implying lower near-infrared albedo due to the strongly modified land cover (correlation of −0.82, figure S10). Also population density is not independent from ISA fraction, NDVI or near-infrared albedo, with correlations of 0.71, −0.78 and −0.72, respectively. The elevation shows a moderate correlation of 0.51 with NDVI, probably related to Kampala's vegetated hills.
Due to this collinearity between the explanatory variables, the stepwise backward elimination procedure only retains NDVI as explanatory variable in the linear regression model for minimum, mean and maximum Humidex (figure 4). With R 2 = 0.79, the NDVI has a high explanatory power for minimum (early morning) Humidex, meaning that 79% of the the intra-urban Humidex variability can be explained. This explanatory power is similar for average (75%), but lower for maximum (midday) Humidex (52%).   The explanatory power of NDVI allows us to apply the regression model by extrapolating the point observations of minimum, mean and maximum Humidex to the greater Kampala region ( figure 5). This map provides information on the spatial heterogeneity of heat stress in Kampala.

Discussion
With average intra-urban differences of 1.2 • C, and an afternoon difference of 6.4 • C on average, the Humidex Index is heterogeneously distributed over the city of Kampala. These large intra-urban differences are also reflected in the number of days exposed to extreme heat. At some locations, daily maximum heat stress exceeds the great discomfort level, defined by Humidex >40 • C, for more than 50% of the 3 × 42 observation days. In comparison, for the same period, this level was never reached at the Makerere station and only a few times in the city outskirts. Moreover, we identified the regions in Kampala that are most prone to heat, pointed to non-vegetated, densely built environments using linear regression and extrapolated the result to the greater Kampala region. The resulting map could complement remote sensing products for land surface temperature, adding information about 2 m temperature and relative humidity combined in the Humidex Index. Such map can guide anticipatory action plans that help reduce the impact of heatwaves, by providing concrete information about heat-prone areas. With special attention to those heat-prone areas, a heat action plan commits to public awareness about heat risks (Singh et al 2019), which has been demonstrated to be successful in reducing heat-related mortality in Ahmedabad, India (Knowlton et al 2014, Hess et al 2018, Nastar 2020. In addition, a heat action plan also accounts for the vulnerability of urban dwellers. In general, densely built and informal settlements house a large part of the population belonging to the lower socioeconomic status with income and livelihood insecurity, making them particularly vulnerable (Vermeiren et al 2012, Lwasa et al 2018, Hemerijckx et al 2020, Twinomuhangi et al 2021. An important factor increasing their vulnerability is the housing infrastructure, not offering any protection to heat. As a test case, we collected heat observations inside a building in the informal settlement of Acholi Quarters during the three periods of the observational campaign. Instead of offering protection against heat, the house seems to act like a heat trap for evening and nighttime heat, times when people are living inside. Only heat before and at noon is slightly lower then outside observations (figure S11). A recent study in Ghana investigated indoor temperatures and concluded large effects of building materials (Wilby et al 2021).
A second implication concerns urban adaptation planning, and follows from our finding that Kampala's intra-urban varying heat is strongly correlated with NDVI, explaining up to 77% of the intraurban variability in daily mean Humidex. Despite their elimination in the regression analysis due to clear correlations with NDVI, other explanatory variables, particularly ISA fraction, might also be important. From a Local Climate Zone perspective (Stewart and Oke 2012), the warmest stations were found in compact LCZ classes, characterized by densely built environments with little or no vegetation. In particular, they include compact mid-rise (LCZ 2), compact low-rise (LCZ 3) and lightweight (LCZ 7) classes. Cooler environments in Kampala include open low-rise (LCZ 6) and open mid-rise (LCZ 5) classes. Overall, sparsely built and highly vegetated areas thus experience substantially lower heat, with an observed mean difference of 6.4 • C in the afternoon. This result implies that greening the city could mitigate urban heat (Bowler et al 2010, Demuzere et al 2014, Gunawardena et al 2017. Yet, a more detailed investigation is needed to the overall effects of different types of vegetation on human well-being, including the effects on local climate, air quality and aesthetics (Salmond et al 2016). In fact, the Kampala City Council Authority (KCCA) announced plans to plant 0.5 million trees in Kampala as part of its climate action strategy, which is developed with stakeholders. Concretely, the strategy has embarked on taking stock of the trees in the city coupled with mapping of natural assets in the city to form the basis for implementation of the climate strategy. This climate action strategy is challenging, especially because most available land is in Kampala's residential areas where urban tree canopies are already evident, while we showed that the need for cooling is most urgent in densely built environments and informal settlements. Yet, Lwasa et al (2014) claimed the potential for expansion of tree canopy cover in the city's densely built and most vulnerable areas as well. For this, initiatives from local actors and inhabitants should be strongly supported, which can only be achieved by properly informing the inhabitants (Hintz et al 2018), followed by incentive-based mechanisms that allow the planting, nurturing and taking care of the trees in dense settlements by the developers. As an example, some local politicians and influencers challenged Ugandans via (social) media to commemorate every marriage, death, birth and graduation by planting a personal tree.
Five important challenges have to be addressed by future research. First, besides Humidex, many other indices can describe heat as a combination of multiple factors influencing heat stress. To explain human discomfort, many factors play a role, including air temperature and humidity (Epstein and Moran 2006, Barnett et al 2010, Fischer et al 2012, Lange et al 2020, but also wind, radiation, physiological factors, physical activity and clothing (Quayle and Doehring 1981, Roth 2007, Potchter et al 2018. This study includes temperature and relative humidity only, with differences between similar heat measures shown to be small (Barnett et al 2010). Second, to explain the observed Humidex values, vegetation is an important factor with a twofold role. On the one hand, enhanced transpiration by vegetation implies air temperature cooling, tending to reduce the Humidex. On the other hand, enhanced transpiration increases the humidity which contributes to a higher Humidex (Hass et al 2016). By showing a strong negative correlation between Humidex and vegetation fraction (NDVI), we conclude that the first role dominates. Concretely, adding vegetation might increase the relative humidity, the reduction in temperature is superior. Yet this study did not quantitatively investigate the interplay between temperature and humidity in detail. A study in Indiana, US, highlighted the importance of different vegetation types and structures to unravel this interplay (Souch and Souch 1993). Individual tree species or open grass areas have the largest cooling effect, street trees the smallest. Also in Kampala, different vegetation types could be obtained or derived from detailed landcover maps. Even if such land-cover maps would help the interpretation of the results by further linking heat to different vegetation types, our study already integrates the level of greenness and vegetation by using NDVI and ISA variables. Third, our observational iButton network was only active for a short 3 × 42 d period, during a warm spell. Results might therefore slightly overestimate the multi-year situation. In addition, our network was not complete to cover all neighbourhoods. Moreover, paper-made shields might not be sufficient to perfectly protect the sensor from direct sunlight. However, the role of irradiation of the sensor shields is not investigated since the time of the day at which the sensor is sunlit/shaded was not inventorized. Deviations between the sensor triple observations are small, suggesting that the effect of irradiation is also small. Despite the good performance of our sensors, we still recommend using a pair or triplet of iButtons per site, regardless of the reduced number of observational sites. Besides our iButton network, weather station observations are increasingly conducted in the region via the TAHMO project (van de Giesen et al 2014), yet the meteorological network is still sparse. Besides TAHMO, the availability of crowd-sourced weather data has grown worldwide in recent years, offering possibilities for high-resolution observational urban heat studies (Muller et al 2015, Chapman et al 2017, Meier et al 2017, Venter et al 2020, Fenner et al 2021. Unfortunately, no studies have collected and analysed such data in tropical Africa. Fourth, there is no common definition of heatwaves (Hintz et al 2018). This study investigated different definitions including total days/nights or hours with extreme heat and heat degree hours. Similar to Europe, the US and Australia, a region-specific heatwave definition should be derived from mortality and morbidity studies (Kovats and Hajat 2008, Guo et al 2014, Mora et al 2017, Xu and Tong 2017, Li et al 2020. Yet such studies are currently completely lacking for Africa (Harrington and Otto 2020). In addition, given an appropriate metric, health impact research should also question whether certain thresholds are meaningful for different regions of interest based on more metabolic information on people living in these regions. In this study we provided the general (dis)comfort levels accompanying the Humidex Index. Yet, experienced heat is expected to depend on the region of the world (Potchter et al 2018).
In some areas, heat is experienced as 'great discomfort' below/above the proposed 40 • C threshold. Surveys in Dar es Salaam (Tanzania) suggested that the comfort range was well above the one defined in temperate climates (Ndetto and Matzarakis 2017). While concrete information about applicable levels/thresholds for Kampala is lacking, this study provisionally explored different thresholds for the Humidex Index, ranging from 30 • C to 50 • C. Fifth, though the explanatory power of the regression model is high (75%), four factors could lead to improvements of the model. First, it would definitely benefit from more observations, perhaps supplemented by remote sensing data. Second, this study explored only six explanatory variables, but more could be added, for example providing information about building morphology or materials. Though this information is implicitly already included in LCZ classes, high resolution material maps will possibly appear in the near future. Third, future research should investigate the robustness of the model when adding new observations. Fourth, we found strong collinearity between different explanatory factors, challenging the causality between vegetation fraction and Humidex Index in our linear regression model.

Conclusion
From a network of low-cost temperature and humidity sensors, we compute the Humidex Index, quantifying the heat stress throughout the city. Daily minimum, mean, maximum as well as extreme heat stress are heterogeneously distributed over the city, with poorly vegetated and densely built-up environments being the most heat-prone areas. Their inhabitants, generally vulnerable people due to their socioeconomic status, are often exposed to great discomfort or even dangerous heat. Future research should bridge the gap to indoor heat and its dependence on house structure and building materials. Two recommendations follow from this study: to mitigate heat stress, urban greening should be considered in urban planning strategies, and urban heat action plans should account for the large intra-urban heat stress variations.

Data availability statement
The data that support the findings of this study are openly available at the following URL/DOI: http://doi.org/10.5281/zenodo.5105570.

Acknowledgments
This work was financially supported by Internal Special Research Fund (BOF-479C1 Project C14/17/053) and the REACT project. M Demuzere is supported by the ENLIGHT project, funded by the German Research Foundation (DFG) under Grant No. 437467569. The authors thank Daniel S Ddumba (Makerere University) and Sadam Y Kiwanuka (KCCA) for their support during the field work, and local people for taking care of the sensors. OB was financed by the Remote Sensing for Epidemiology in Sub-Saharan African CiTies (REACT http://react. ulb.be/) project, funded by the STEREO-III program of the Belgian Science Policy (BELSPO, SR/00/337) and the Wellcome Trust HEROIC project (Grant No. 216035/Z/19/Z).

Code availability
Scripts to process the raw temperature and relative humidity observation are provided via https://github.com/jonasvdw/kampalasensors.