Use of remotely sensed land surface temperature as a proxy for air temperatures at high elevations: Findings from a 5000 m elevational transect across Kilimanjaro

High elevations are thought to be warming more rapidly than lower elevations, but there is a lack of air temperature observations in high mountains. This study compares instantaneous values of land surface temperature (10:30/22:30 and 01:30/13:30 local solar time) as measured by Moderate Resolution Imaging Spectroradiometer MOD11A2/MYD11A2 at 1 km resolution from the Terra and Aqua platforms, respectively, with equivalent screen‐level air temperatures (in the same pixel). We use a transect of 22 in situ weather stations across Kilimanjaro ranging in elevation from 990 to 5803 m, one of the biggest elevational ranges in the world. There are substantial differences between LST and Tair, sometimes up to 20°C. During the day/night land surface temperature tends to be higher/lower than Tair. LST‐Tair differences (ΔT) show large variance, particularly during the daytime, and tend to increase with elevation, particularly on the NE slope which faces the morning Sun. Differences are larger in the dry seasons (JF and JJAS) and reduce in cloudy seasons. Healthier vegetation (as measured by normalized difference vegetation index) and increased humidity lead to reduced daytime surface heating above air temperature and lower ΔT, but these relationships weaken with elevation. At high elevations transient snow cover cools LST more than Tair. The predictability of ΔT therefore reduces. It will therefore be challenging to use satellite data at high elevations as a proxy for in situ air temperatures in climate change assessments, especially for daytime Tmax. ΔT is smaller and more consistent at night, so it will be easier to use LST to monitor changes in Tmin.


Introduction
There is concern that mountain areas in comparison to adjacent lowlands may show enhanced warming in response to greenhouse forcing due to factors such as snow-albedo feedback, changes in atmospheric moisture, and the laws of physics [see Rangwala and Miller, 2012;Pepin et al., 2015]. However, surface temperature observations are skewed toward lower elevations [Lawrimore et al., 2011], and the current observational network is inadequate to determine whether high elevations are warming more rapidly than lower elevations. Numerous studies have investigated observational evidence for elevation-dependent warming, but while some show an increase in warming rates in high mountains [Diaz and Bradley, 1997;Ohmura, 2012;Yan and Liu, 2014], others show a decrease [Lu et al., 2010;Vuille and Bradley, 2000;Ceppi et al., 2010;You et al., 2010] or a more complex picture .
Remotely sensed temperatures have the advantage of extensive spatial coverage, often at fairly high resolution [Merchant et al., 2013], but so far, relatively few studies have used this potential data source in the context of temperature trends at high elevations [Qin et al., 2009]. Although this is partly because of the short length of the satellite records, this is becoming less relevant with time. However, satellites measure land surface or "skin" temperature (hereafter referred to as LST) [Jin and Dickinson, 2010] as opposed to air temperature at screen level (~2 m above ground; hereafter referred to as T air ). The latter is commonly used for climate change assessments [Hartmann et al., 2013] and is a critical measurement in ecology, hydrology, and climate science. LST is more strongly controlled by the surface radiation balance [Benali et al., 2012], which is in turn controlled by the interplay between local land surface characteristics and solar geometry, while T air tends to be more regional in scope, although rapid variations can occur over short distances due to factors such as cold-air drainage [Daly et al., 2010;Lundquist et al., 2008]. Thus, if satellite records are to be applied to examine elevation-dependent warming, studies which compare LST and T air at extremely high elevations (>4000 m) are required. So far, comparisons have been limited to lowland [Coll et al., 2009;Vancutsem et al., 2010] or moderate elevation environments [Wenbin et al., 2013;Shamir and Georgakakos, 2014]. This paper therefore compares Moderate Resolution Imaging Spectroradiometer (MODIS) LST from the MOD11A2/MYD11A2 products [Wan, 2006] with T air along a transect of 22 weather stations across Kilimanjaro, including sites at elevations up to 5800 m. The elevational range of nearly 5000 m is one of the largest in the world. The main aim is to understand the contrast between LST and T air and uncover the factors that control this. Previous studies are discussed in section 2. The study area is outlined and methods are explained in section 3. The main behavior of the difference between LST and T air is examined in section 4, along with the factors that control its variation, before the consequences of our findings are discussed.

Previous Studies Comparing LST and T air in Different Regions
MODIS LST is produced as a series of MOD11 (Terra) and MYD11 (Aqua) products, validated with in situ measurements in more than 50 clear-sky cases in the temperature range of 263-331 K (for details see Wan et al. [2002Wan et al. [ , 2004, Wan [2006], Wan [2008], Wang et al. [2008], and Coll et al. [2009]). A generalized split-window algorithm [Wan and Dozier, 1996] is used to retrieve LST at 1 km resolution. LST has been compared with T air in a range of environments including the Arctic [Urban et al., 2013;Østby et al., 2014;Westermann et al., 2012], midlatitude continental areas [Shen and Leptoukh, 2011;Hachem et al., 2012;Benali et al., 2012;Wenbin et al., 2013], and the tropics [Maeda et al., 2011;Vancutsem et al., 2010]. Vegetation plays a key role on LST especially in the tropics, but snow cover and ice cover are important at higher latitudes and elevations.
Comparing LST and T air at Voi in Kenya (savannah vegetation), Maeda et al. [2011] showed a high correspondence at night (22:30 local solar time) but significant differences during the day (10:30 local solar time). Larger differences were observed around the equinoxes when solar input was strongest. Mostovoy et al. [2005] also showed seasonal differences in the LST/T air temperature relationship over Mississippi. Larger differences in summer were related to changes in vegetation. Vancutsem et al. [2010] also show that differences between daytime LST and T air vary according to the ecosystem and solar radiation input at the surface. Benali et al. [2012] developed statistical models to predict LST from T air in Portugal using variables including day length, elevation, and continentality. In high latitudes the presence of extensive snow cover reduces variability in the difference between LST and T air [Hachem et al., 2012], but LST is consistently colder. Cloud cover could introduce systematic bias in LST [Westermann et al., 2012], and it is difficult to separate cloud cover and snow cover, especially when patchy, due to similarities in thermal emission. Shamir and Georgakakos [2014] examined contrasts between LST and T air at moderate elevations in eastern Turkey in an area with seasonal snow cover. The relationship between the two was strongly dependent on snow cover. Daytime LST excesses of 15-20°C were recorded in summer when snow was absent, but daytime LST sometimes became colder than T air when snow was present.
Most studies conclude that LST is much more spatially variable than T air , particularly during periods of strong insolation (daytime) and when there is patchy cloud and/or snow. Such situations are common in the tropics (convective cloud) and at high elevations. The extreme radiation climate at high altitudes [Barry, 2008] is thus expected to increase the variability between LST and T air further. Both ephemeral snow cover and vegetation contrasts (e.g., across treelines) are also expected to influence differences at high elevations. However, there have been no detailed comparisons of LST and T air at elevations above 4000 m because of the paucity of reliable air temperature data.
Much of the work using satellite temperature data in Africa and in the Kilimanjaro area has concentrated on the effect of vegetation on surface temperatures [Maeda et al., 2011;Maeda and Hurskainen, 2014]. This is unsurprising given the central influence of land use change, in particular deforestation and land cover change, on tropical environmental change [Soini, 2005]. Rapid change has been observed on Kilimanjaro [Shugart et al., 2001;Lambrechts et al., 2002;Hemp, 2005;Fairman et al., 2011]. Vegetation density and state can be measured by the normalized difference vegetation index (NDVI) [Townshend and Justice, 1986] which can be obtained from the MODIS sensor reflectance measurements. The NDVI is calculated as follows: where ρ NIR is the near-infrared reflectance and ρ Red is the red reflectance. MODIS products can offer NDVI imagery with different preprocessing levels in terms of atmospheric correction, cloud screening, and bidirectional reflectance distribution function effect removal. Maeda and Hurskainen [2014] showed that NDVI (as measured by MOD13Q1) had a strong influence on LST patterns across Kilimanjaro during the daytime in addition to elevation, drier areas with lower NDVI having increased LST, but the strength of this effect varied seasonally. During the night this influence disappeared and elevation became the dominant control.
The negative relationship between NDVI and daytime LST has been used as the basis of the temperature/vegetation index method [Nemani and Running, 1989], which applies this correlation to correct LST to provide an improved estimate of T air . LST becomes more similar to T air when vegetation is healthy (high value of NDVI) because the canopy has an effective radiative temperature close to the air temperature. However, this method has shown limited success at high elevations on the Tibetan Plateau [Wenbin et al., 2013], and at extremely high elevations vegetation is absent.
The aim of this study is to understand the different climatologies of LST and T air across Kilimanjaro, a location with an extreme elevation range and strong environmental gradients in moisture, vegetation, and snow cover. It is expected that LST will be more spatially and temporally variable than T air , particularly during the day, and that decoupling between LST and T air will increase at the highest elevations.

Study Area
Kilimanjaro is an isolated mountain in northern Tanzania with a summit elevation of 5895 m. The base lies below 1000 m, meaning that this is one of the largest and highest free-standing mountains in the world. The mountain covers numerous ecological zones ( Figure 1) [Hemp, 2006a]. The lower slopes (submontane belt: 1000-1800 m) are dominated by agriculture and support a high population, mainly through coffee and banana cultivation and small-scale subsistence farming ( Figure 1a) [Hemp, 2006b]. Above 1800 m most of the natural cloud forest has been preserved, especially on the southern slopes (Figure 1b), although the northern slopes are much drier. The cloud forest extends up to around 3000 m, above which is the giantheather zone (~3100-3900 m; Figure 1c) dominated by Erica arborea. Above 3900 m the dominant vegetation becomes alpine heathland/moorland including Helichrysum cushion vegetation ( Figure 1d). Above 4500 m there is very little vegetation and the environment is arid and dominated by the extreme radiation regime of high elevations ( Figure 1e). There are small summit ice fields within the summit crater ( Figure 1f).
The climate of Kilimanjaro has been extensively studied [Coutts, 1969;Hastenrath, 1991;Hastenrath and Greischar, 1997], most recently in connection with the ablation of the summit ice fields [Mölg et al., 2009]. The summit of the mountain lies above a trade wind inversion, meaning that the majority of the cloud cover and precipitation occurs on the lower slopes [Hemp, 2001]. The precipitation peaks at about 2700-3000 mm at 2200 m elevation on the southern slopes [Røhr and Killingtveit, 2003;Hemp, 2006c;Schüler et al., 2014]. There are two main rainy seasons [Camberlin and Phillipon, 2002] known as the "long rains" in April/May and "short rains" in November/December. However, cloud patterns are slightly different between the two. In the short rains, mornings are often clear, but showers and thunderstorms develop in the unstable atmosphere by the afternoon [Mölg et al., 2009]. During the long rains there is more widespread persistent precipitation and cloud cover. Outside this period, above 4000 m, skies are typically clear, particularly in the June-July-August-September (JJAS) dry season, which means an extreme radiation climate with intense sunlight by day and clear cold conditions at night. This in turn sets up a thermal circulation on the mountain slopes [Duane et al., 2008], which sometimes brings cloud and moisture up the mountain during the afternoon  only to subside by dusk. Because the mountain is equatorial, sunrise and sunset are at approximately 06:30 and 18:30 East African standard time (EAST) throughout the year.

Data
In September 2004 a transect of 10 air temperature/relative humidity stations was installed on the southwest (SW) slope (Machame route) ranging from 1800 m to 5800 m in elevation [Duane et al., 2008]. This was supplemented by lower stations down to 900 m and another transect of 10 stations on the NE slope (Rongai route) at equivalent elevations to the SW slope in 2012 (see Figure 2). Station details are listed in Table 1 (U23-001) are installed at 1.5 m above ground level, attached to vegetation on the lower slopes or on tripods above treeline (>4000 m). Sensors are screened with white tubing and orientated (almost) horizontally with the long end facing north-south to ensure that no direct sunlight would enter the tube. A slight angle (<5°) was preferred to avoid air pooling within the tube and increase ventilation. One of the two summit sensors is sited on the northern ice field as part of the weather station operated by the University of Massachusetts used for mass balance research. The other is sited at an equivalent elevation~200 m away on the ice-free summit of Albion (see Figure 2). At Albion a downward facing sensor installed from 2010-2015 allowed an identification of presence/absence of snow cover, based on reflected shortwave radiation.
Data were recorded every hour (all T air times are EAST:UT +03:00) and records run from September 2004 to September 2015 (11 years) on the SW slope and September 2012 to September 2015 (3 years) on the NE slope. There are some data gaps due to a combination of phased logger installation, vandalism, and logger malfunction (see Table 1), but overall most sites have considerable data of good quality. Data have been evaluated in previous publications [Duane et al., 2008;Pepin et al., 2010] through comparison with National Centers for Environmental Prediction/National Center for Atmospheric Research reanalysis temperatures at equivalent elevations and at the highest site calibrated against the automatic weather station run by the University of Massachusetts. The suitability of this type of data logger for climate monitoring has been assessed by Whiteman et al. [2000].
LST is obtained from the MODIS MOD11A2 and MYD11A2 products (Terra and Aqua platforms), which corresponds with 10:30/22:30 local solar time (daytime/nighttime; Terra) and 13:30/01:30 (daytime/nighttime) local Journal of Geophysical Research: Atmospheres 10.1002/2016JD025497 solar time (Aqua) (e.g., https://lpdaac.usgs.gov/dataset_discovery/modis/modis_products_table/mod11a2). The data are stored on a 1 km sinusoidal grid. We extracted LST for the relevant pixel for each station ( Figure 2). All coordinates were based on universal time meridian and the World Geodetic System 84 ellipsoid. The LST data are produced as an 8 day composite. The composite itself is only derived from clear days within the 8 day period, so in some cloudy periods the composite will be based on only a few days. Occasionally, no value is available because all days were cloudy within the 8 day period. Although a 1 day product exists  (MOD11A1), the high frequency of cloud cover on the slopes of the mountain means that many (sometimes well over 50%) of individual days are missing. The use of the 8 day product avoided this problem. The 8 day product also easily overlapped with the 16 day NDVI product (see below).
Quality control was undertaken using the quality assurance layers provided by the MOD11A2/MYD11A2 products. For each individual pixel the QC layer contains information on measurement quality, emissivity error, and average LST error. We excluded pixels in which LST was not produced due to atmospheric interference or not processed due to poor quality. The clear-sky coverage layer provides an assessment of clear-sky conditions within the 8 day period. Over land MODIS LST is only calculated for pixels at clear-sky conditions at 95% confidence for regions below 2000 m above sea level (asl) and at 66% confidence for regions above 2000 m asl [Wan, 2008].
NDVI imagery was obtained from the MOD13Q1 product. This offers 16 day composites at a spatial resolution of 250 m, in which each pixel represents the best quality measurement within a 16 day period [Justice et al., 2002]. A data quality control was undertaken considering the MOD13Q1 pixel reliability layer, in which pixels tagged as cloudy were removed.

Analytical Methods
To compare LST and T air across the range of elevations and ecological zones, air temperatures for 11:00/ 14:00/23:00/02:00 EAST were compared to 10:30/13:30/22:30/01:30 LST measurements, respectively. Solar noon occurs at 12:30 EAST, and the satellites pass over on average 90 min before/after this. We aggregated T air into the same 8 day periods as the LST measurements. This is likely to be a better approximation when there are many clear days within the composite than when many days are cloudy.
Because some stations on the crater (Mach8/Mach9 and Mach10/Albion) are very close together they fall within the same MODIS pixel. Yet there can be a substantial elevation difference between them. The variability of elevation within each pixel (Figure 2) can be as much as 550-718 m on the crater wall. To account for the potentially large variation within a 1 km pixel, we corrected LST using a standard lapse rate of À6.5°C/km to the elevation of the station (from the mean pixel elevation). The maximum elevation difference was 335 m (Mach9) which translated to a correction of À2.18°C, but most corrections were less than 1°C (see Table 1). We assess error due to intrapixel elevation variability in the discussion.
We calculated LST-T air (hereafter referred to as ΔT). If ΔT is positive it means that the surface is warmer than the air and vice versa. We examined the variation of ΔT with elevation and slope aspect and factors such as season, cloud cover, NDVI, and snow cover. NDVI was provided at 250 m, and the nearest pixel to the station was used (this is a subset of the larger 1 km LST pixel). The 16 day NDVI composite was used to cover two consecutive 8 day periods.  (Figure 3) shows a fairly regular decrease with elevation, especially at night (Figure 3b). There is a rapid decrease in daytime temperature (Figure 3a) where the rainforest begins (~1800 m). Temperatures are then fairly stable up until treeline (~3000 m). A weak lapse rate develops around 2500-3000 m in the austral summer and an inversion forms above treeline (3000-3500 m) in the winter (JJA). During the night this complexity disappears and there is a more or less consistent decrease in temperature with elevation. The nighttime freezing level lies around 4300 m and varies little throughout the year. T air is in general slightly lower at night during the austral winter (JJAS), more especially at lower elevations.

Raw Data
The sharp change at the lower forest limit during the day is also picked up by the satellite LST (Figure 4a), with the cloud forest causing considerable cooling in comparison with the warm agricultural zone immediately below. However, there is also considerable surface heating in the alpine moorland zone (4000-4500 m) during the daytime, particularly in the dry seasons (JF and particularly JJAS). LST reaches over 25°C during July-September. At this elevation there is a lack of cloud in these months, and the dry surface heats up Journal of Geophysical Research: Atmospheres 10.1002/2016JD025497 effectively due to minimal latent heat flux. The topography also flattens into a plateau (Shira) between 3800 and 4500 m, which would enhance surface heating due to both the mass elevation effect [De Quervain, 1904] and overhead Sun angle. A strong inversion therefore develops above treeline during the afternoon. This seasonal contrast in LST remains at the highest elevations (>5000 m), with temperatures during the wet seasons (November-December and March-April-May) being reduced in comparison with the dry seasons. At night LST depends more on elevation alone and is much more regular. Seasonal variations are subdued. Figure 5 shows a similar plot of ΔT. The most striking feature is the strong variability of daytime ΔT, which reaches over 20°C during the dry seasons at around 4000 m and above. Daytime differences below 3000 m are much smaller, although still predominantly positive. At night, differences tend to be weakly negative,   Figure 6 plots the correlation (r) between individual 8 day LST and T air observations at each time of day as a function of elevation. The right-hand graph is reversed so the two horizontal axes represent a transect across the mountain from SW to NE. The potential number of 8 day periods on the SW slope is around 500, although n is usually around 300 because of cloud contamination and missing air temperature data. The northeast slope shows more noisy results because there are less data. The most striking pattern is the decrease in correlation at high elevations during the day (11:00 and 14:00 EAST) on both slopes. At night the decrease in correlation disappears and there is a less systematic deterioration of r with elevation.

Patterns of LST Versus T air
The mean LST-air temperature difference (ΔT) is plotted in Figure 7 in a similar format. Again, there are large contrasts between day and night. During the day LST tends to be warmer than T air (positive ΔT), although differences are fairly small on the lower forested slopes (particularly in the cloud forest around 2000 m). ΔT increases rapidly above treeline (3100 m), peaking on the higher slopes around 5000 m. On the northeast slope mean differences at 14:00 EAST exceed 20°C at Rong8 (4966 m). Even on the southwest slope differences are significant (between 5 and 15°C). At night the contrast between the slopes disappears, and the LST is uniformly colder than the air but not by a large amount. The elevational gradient in the difference also disappears at night. These distinctive patterns show that LST and T air are certainly not the same measurement.  (Figures 8a and 8b), but there is a less consistent drop at night and perhaps on the NE slope in general. Differences between clear and cloudy periods are subtle. Figures 9 and 10 show the seasonal patterns of ΔT at a selection of low/high-elevation sites, respectively. During the day (11:00 and 14:00 EAST), differences are much larger during the dry seasons when solar heating is strong and the surfaces drier. At Moshi (on the plains) the peaks in February/March and September/October correspond with both the Sun being directly overhead at noon and dry seasons. In the rainforest zone (Mach2 and Rong2) seasonal changes are more subdued. Seasonal changes are also smaller at night in general at all sites. Perhaps against intuition, the biggest differences (most negative ΔT) are often  during the wet seasons (May or November) at the low-elevation sites, the reason for which is unclear because surface radiation loss should be limited by moist surfaces. At higher elevations ( Figure 10) the influence of wet and dry seasons becomes more important. At Mach7, Mach10, and Albion, the short rains (November/ December) mark a time when convective cloud build up and precipitation is much more likely at crater level. This appears to reduce daytime ΔT. There is a large contrast between the two dry periods, with the least cloudy JJAS period mostly responsible for the large daytime ΔT values in excess of 20°C. Similar to low-elevation sites, nighttime ΔT shows much less seasonal variation in general.

Seasonal Patterns in ΔT
In summary the differences between LST and T air show strong patterns controlled by the frequency of cloud cover and its interaction with the changing radiation balance over the course of the day and year. In future sections we examine how seasonal and short-term (intraweekly) changes in moisture and relative humidity, vegetation, and snow cover can help explain the variation in ΔT. Figure 11 plots ΔT against the number of cloud-free days within the composite for two SW slope sites during the day (11:00 EAST; Figures 11a and 11b) and night (23:00 EAST; Figures 11c and 11d). The 14:00/02:00 EAST patterns are broadly similar (not shown). The x axis represents the number of clear days within the composite (more cloudy conditions on the left). Mach2 is in the rainforest, and Mach8 in the alpine desert. Accepting a null hypothesis that cloud has no effect on ΔT, then ΔT should become less variable and closer to zero as the number of clear days increases, purely because LST is based on more of the same days as T air (which is based on all 8 days). However, although the variance in ΔT decreases (probably due to increasingly data similarity) the size of the daytime difference tends to increase. Thus, under composites with more frequent clear-sky conditions (stronger radiation input) the surface tends to heat up more in comparison with the air, particularly at the higher site (Mach8). At night the variance in ΔT also decreases for clearer composites. However, the mean difference does not always become more negative (which we would expect based on rapid surface cooling under clear skies). At some sites such as Mach2 in the rainforest the difference becomes smaller (closer to zero), but at others (Mach8) it becomes larger (more negative). Thus, differences between sites impact on nighttime cloud effects on ΔT.

Effect of Cloud Cover and Relative Humidity
The data loggers also recorded relative humidity as well as temperature. Apart from one or two exceptions, there are strong negative correlations between daytime ΔT and relative humidity (%) at many locations ( Table 2). The left-hand columns represent the correlation based on raw 8 day values and the right-hand columns the correlation remaining when the common seasonal cycles in both ΔT and relative humidity have been removed. Although correlations tend to be stronger using the raw values rather than anomalies, the difference is small near the summit of the mountain, meaning that the seasonal cycle is not responsible for the negative correlations between ΔT and relative humidity (RH) above 4500 m. Strong interweekly variability in moisture caused by synoptic variations must therefore account for much of the change in ΔT on the summit crater and at some sites on the NE slope (e.g., Rong4, Rong5, and Rong7). More humid conditions (which will mean a much higher likelihood of cloud cover at the elevation of the station) lead to subdued differences. At Mach7 and Rong7 (4555 m) for example nearly 50% of the variance in the difference can be accounted for by afternoon RH, and a significant relationship remains (p < 0.01) when annual cycles have been removed.

Effect of Surface Characteristics (Vegetation and Snow Cover)
The relationship between NDVI MOD13Q1 and daytime ΔT is shown in Figure 12 for two bottom and two high-level sites on each transect. In the lowlands (Moshi and Rong1) there are strong negative relationships, with increased NDVI (greener vegetation) reducing ΔT as in previous studies. At some high-level sites (but not Mach10 on the ice field) the relationship reverses. However, NDVI is very low at the highest elevations (there is no vegetation) and is more likely to be influenced by snow cover. Table 2 (right-hand column) lists the correlation between daytime ΔT and NDVI MOD13Q1 for all locations. Correlations tend to be more significant on the SW slope, but this is probably because there is a larger sample size. At night there was an absence of significant correlations with MOD13Q1 (not shown).
At Albion where snow cover is sporadic, we assessed the effect of snow cover on the relationships between NDVI and ΔT and on the correlation between LST and T air . Snow cover was estimated to be present if reflected shortwave radiation at 11:00 EAST exceeded 237 W/m 2 (around 43% of days). Although there was no significant difference in the LST/T air correlation between snowy (4 or more days) and non-snowy (less than 4 days) composites at Albion, LST was reduced much more than T air by the presence of snow cover (Figure 13a). There is also a significant difference (two-tailed t test: p = 0.0012) between NDVI values recorded under conditions of snow cover versus snow-free composites (Figure 13b), with more negative values representative of snowy conditions. When mean NDVI at a station falls below 0.1 the daytime correlation between LST and T air also falls dramatically (figure not shown).

Discussion
The rapid deterioration (and even reversal) in correlation between LST and T air at high elevations is important to the debate on elevation-dependent warming. Although satellite data have great potential to complement air temperature data at high elevations [Qin et al., 2009], unfortunately, LST is increasingly poorly related to T air as elevation increases. One likely cause is that the surface radiation balance becomes increasingly variable both temporally (as solar geometry changes) and spatially (as factors such as aspect, topographic screening, and land surface contrasts become increasingly important in high mountains) [Barry, 2008]. The low air density at high elevations means that despite rapid surface response to energy balance fluctuation, this is not transferred readily into T air at screen level, which is more decoupled from the surface response (LST) than it would be at sea level. At extremely high elevations a negative relationship between LST and T air could arise because increased moisture warms the air through increased longwave emission [Rangwala et al., 2009], but cools the surface through increased likelihood of snow cover, and increased latent heat flux.
Intense sunlight at high elevations can warm the surface far above T air , which explains the daytime increase in ΔT with elevation. Although the rate at which direct radiation input increases with elevation depends on solar angle and atmospheric transmissivity [Kastrov, 1956; cited by Barry, 2008], solar radiation intensity at the summit of Kilimanjaro (~6000 m) is typically 5-12% greater than at sea level [Barry, 2008] and frequently exceeds 1000 W/m 2 [Mölg et al., 2008] on the northern ice field. The NE aspect of the mountain faces the Sun in the morning, explaining why differences between LST and T air are slightly smaller on the SW aspect at both 11:00 and 14:00 EAST. Although the SW slope should heat up in the afternoon, convective cloud limits afternoon heating Mölg et al., 2009]. Although daytime values of ΔT can exceed 20°C, similar results have been obtained in other studies under less extreme radiative conditions [Shamir and Georgakakos, 2014], and we have no reason to suspect that these values are erroneous.
Because surface heating is dependent on direct radiation input, any patchy cloud cover, or any topographic screening will change LST but have less influence on T air . We examined the influence of changing pixel at the highest-elevation sites (Mach10 and Albion) to the neighboring 1 km square (which could be topographically contrasting), and in some cases differences in LST were substantial (>10°C). It is informative to compare this with error due to altitude variation within a pixel (up to ±350 m). In the worst case (the western crater wall) this translates to an error of ±2.28°C, assuming a standard lapse rate of À6.5°C/km [Maeda, 2014]. However, elevation error is still insignificant compared with mean values of ΔT, which can reach over 20°C ( Figure 10) on the higher slopes, and compared with interpixel variability of LST due to differential aspect/radiative load (>10°C). Thus, small differences in positioning of satellite measurements over time (i.e., because of changing cloud cover) would create significant changes in LST. Thus, there is a lot of small-scale noise in the LST record, which would need to be corrected for if satellite data were to be used to represent high-elevation temperature trends over large areas.
The strong negative relationship between daytime ΔT and relative humidity means that moister air leads to reduced surface heating. This pattern is particularly strong at midlevel stations on the SW slope (i.e., Mach6, Mach7, and Mach8), where there is large day to day variance in the amount of cloud cover and moisture brought up the slope through the thermal circulation mechanism [Duane et al., 2008]. The strong relationship of ΔT with atmospheric moisture opens up the possibility, at least in theory, of using humidity measurements to predict ΔT and subsequently adjust satellite temperatures to predict air temperatures, if atmospheric soundings were available. However, previous work on Kilimanjaro  has demonstrated large differences between in situ moisture (at screen level from surface stations) and free air measurements at the same elevation (from reanalyses and radiosonde ascents), and such differences would complicate the process. Healthy vegetation can also dampen daytime surface warming, through increasing the latent heat flux. ΔT was smaller and more consistent at the lower vegetated sites than higher up the mountain. However, even at lower sites the surface still warmed during the day in comparison with T air . In the cloud forest the satellite is measuring the canopy surface temperature (above the temperature sensor) as the effective radiative surface. NDVI showed a strong negative correlation with ΔT at the low-level sites, with greener vegetation (higher index) leading to relative surface cooling as supported by previous studies [Nemani and Running, 1989;Wenbin et al., 2013]. However, at the highest elevations, where vegetation is reduced or absent, the biophysical interpretation of NDVI variation is difficult, and the relationship reverses. The main source of variability in surface properties at the higher stations (excluding the ice field itself) is transient snow cover. At Albion snow cover occurred around 43% of the time, and although it did not appear to have a strong influence on the LST/T air correlation, it did reduce LST in preference to T air and thus reduce ΔT. NDVI also reduces under conditions of snow cover, and this probably reverses or weakens the relationship between ΔT and NDVI at some higher stations. Unfortunately, reliable snow cover data needed for a more extensive analysis are not available at a fine enough resolution, both Mach10 (permanent ice) and Albion (transient snow cover) falling in the same pixel for example, but snow effects should be a focus for further research.
At night it would be expected that the surface would cool more at higher elevations (since radiation escapes more effectively into space), but although LST is consistently lower than T air , there is no strong elevational signal in ΔT. More consistent differences means that using satellite data as a proxy for air temperature could therefore be much more reliable, and since many studies have shown nighttime warming rates to be both faster globally than mean temperatures [Karl et al., 1993] and also to show more of an elevational-dependent signal [Rangwala and Miller, 2012], it would be wise to first use satellite temperatures to examine trends in T min at high elevations. Despite elevation-dependent warming providing a strong context for our research, we have not analyzed temporal trends in LST, T air , or ΔT in this study because the length of the record (11 years) is not long enough to draw robust conclusions, but our findings show that LST and T air require careful comparison before LST is used as a proxy for T air in mountainous terrain.

Conclusions
Our study has shown that there are large instantaneous differences between LST as measured by the MODIS MOD11A2/MYD11A2 products at 1 km resolution and equivalent T air measured from in situ weather stations. These differences are variable, particularly during the daytime, and the predictability of the differences tends to reduce with elevation. Both relative humidity and vegetation state can be used to some extent to predict ΔT, but at extremely high elevations this method is less successful. Our study is only based on one transect on one tropical mountain, but it is likely to represent the challenges of using LST to represent T air in other highelevation areas worldwide, with similar changes in vegetation, moisture, and snow cover with elevation. Treelines and snowlines are of universal importance on mountains worldwide, so the large-scale controls on ΔT in this study are broadly similar to those in all mountains. High-elevation radiation balance is controlled by the fundamental laws of physics [Ohmura, 2012], so the results shown here would be expected to occur in other mountain ranges. There is no permanent snowline on Kilimanjaro (snow is transient, even on the summit crater), and further research is required on the role of persistent deep snow cover on the climatologies of LST and T air [e.g., Hachem et al., 2012]. The general conclusion that LST becomes more different from air temperature as elevation increases is an important one, because careful validation is therefore required before such LST records can reliably be used to assess elevational dependency of T air warming rates in mountainous areas.