Changes in the upper and lower tails of the temperature distribution are key characteristics of how global warming will impact climate (Hartmann et al., 2013). These expected changes in temperature are in line with a series of recent high-profile extremes witnessed across Europe (2003, 2010; Coumou and Rahmstorf, 2012; Schär et al., 2004; Barriopedro et al., 2011), western North America (van Mantgem et al., 2009), the Amazon (2005, 2010; Philips et al., 2009; Lewis et al., 2011) and Australia (2012/2013; van Gorsel et al., 2018). Changes in temperature extremes are not only limited to the warm tail; the cold tail has also seen a notable change, with observed decreases in cold extremes particularly across North America (Wolter et al., 2015). Given the wide-ranging impacts of temperature on vegetation function (Berry and Björkman, 1980; Gunderson et al., 2009; Valladares et al., 2014; van Gorsel et al., 2016; Kumarathunge et al., 2019), health (McMichael and Lindgren, 2011), socio-economics (McEvoy et al., 2012; Colombo et al., 1999; Zander et al., 2015) and land–atmosphere feedbacks (Fischer et al., 2007; Teuling et al., 2010; Miralles et al., 2012; Kala et al., 2016; Donat et al., 2017), projecting the impact of changes in temperature extremes is critical.

Our understanding of how temperature extremes will change is based on simulations using coupled climate models, e.g. the Coupled Model Intercomparison Project (CMIP5) (Eyring et al., 2016). To build confidence in these projections, models should be consistent with our understanding of changing temperature extremes, the impact on the vegetation and the associated feedback on the climate. However, current models are known to have key weaknesses in simulating both temperature extremes (Sillmann et al., 2013; Sippel et al., 2017) and the response of the vegetation to these extremes. For example, most climate models represent broad geographic regions with a single photosynthetic temperature response function, which varies only with plant functional type (Smith and Dukes, 2013; Lombardozzi et al., 2015; Mercado et al., 2018). This assumption would seemingly contradict empirical evidence, showing that the temperature response of photosynthesis varies as a function of climate (Berry and Björkman, 1980; Gunderson et al., 2009). Furthermore, studies show that plants adjust their temperature response of photosynthesis and respiration to changes in ambient temperature (Way and Sage, 2008; Lombardozzi et al., 2015). Although model improvements in the representation of physiological responses to temperatures need to be informed by data from leaf-level and manipulation experiments, data from eddy covariance are also of value. For example, Keenan et al. (2019) recently quantified an apparent inhibition of daytime ecosystem respiration, showing that the diurnal pattern differed from expectations using the global FLUXNET network.

Improving how well models simulate temperature extremes and how vegetation responds to these extremes requires empirical data. The global network of eddy covariance towers (commonly known as FLUXNET), which includes over 900 sites and over 7000 site years, provides measurements of the exchange of carbon, energy and water between the land and the atmosphere. Therefore, eddy covariance measurements provide our best ecosystem-scale estimate of the vegetation's response to heat extremes (Ciais et al., 2005; Teuling et al., 2010; Wolf et al., 2013; von Buttlar et al., 2018; Flach et al., 2018; De Kauwe et al., 2019) although some limitations inevitably remain (e.g. lack of energy closure; see Wilson et al., 2002). Although the length of the temporal records varies across sites, some sites extend back several decades, allowing estimates of the impact of natural variability and climate trends on carbon, energy and water fluxes to be examined.

From each FLUXNET site, measurements of the exchange of latent heat flux (Qle), sensible heat flux (Qh) and net ecosystem exchange (NEE) are available at 30 to 60 min resolution, alongside meteorological variables (including air temperature, net radiation, precipitation and relative humidity). By providing simultaneous and co-located measurements of both the meteorological forcing of the surface, and the associated turbulent energy fluxes, FLUXNET provides a critical resource for understanding ecosystem responses to temperature extremes and for the development, evaluation and benchmarking of land surface models. Importantly, the scale of recorded flux measurements (roughly a square kilometre) is directly relevant for evaluating land surface schemes used in CMIP-type climate models (e.g. Krinner et al., 2005; Abramowitz et al., 2008; Blyth et al., 2011). As a result, land surface modellers routinely use these data to parameterise and evaluate models for extreme conditions. For example, van Gorsel et al. (2016) synthesised eddy covariance data from seven Australian sites alongside a land surface model, to investigate the impact of heat extremes on the exchange of carbon and water fluxes during the record-breaking heat wave in 2012–2013. They found that water-limited woodlands and energy-limited forest ecosystems responded differently to the heat wave, with the forests showing greater resilience to short-term heat than the woodlands. Ukkola et al. (2016) used FLUXNET data to show systematic errors in how well models captured land–atmosphere feedbacks during periods of water stress as a landscape transitioned into drought. In general, as the land surface dries, the surface energy balance tends to partition available energy increasingly towards Qh and less towards Qle, which has important implications for atmospheric temperature, moisture and atmospheric boundary layer depth (Seneviratne et al., 2010). This understanding of land–atmosphere processes was used by Miralles et al. (2014) to link soil desiccation to the amplification of extreme heat waves via land surface feedbacks.

While eddy covariance data have been widely used to examine the impact of temperature extremes, the measurement of temperature and the measurement of Qle, Qh and NEE are independent in terms of the instrumentation used. However, the measured temperature is provided in published data, along with measurements for the site of net radiation, wind speed, humidity etc. alongside measurements of Qle, Qh and NEE. A land surface modeller requires all these data to drive a land surface model for evaluation or process-based studies. We are therefore interested in the relationship between measurements of temperature, and in particular extreme temperatures, and concurrent measurements of Qle, Qh and NEE. Our aim is to characterise, for example, whether direct observations of Qle, Qh and NEE are biased towards the temperature mean and lacking at the tails of the temperature distribution, or whether they are biased to one tail of the distribution. If biases exist, is this true for all FLUXNET sites, or are there specific regions or climates where the tails of the temperature distribution are rich with measurements of Qle, Qh and NEE? We use measurements of temperature and Qle, Qh and NEE from FLUXNET sites because they provide co-located measurements of meteorological variables and land surface fluxes. We seek to identify those sites with data useful to explore land surface processes under extreme temperature conditions, and potentially those sites with the meteorological forcing measured concurrently with the fluxes required to drive land surface models. We therefore do not blend the measured fluxes with meteorological observations taken elsewhere to ensure the land surface fluxes are fully representative of the concurrent meteorological conditions.

Our goal is to identify those FLUXNET sites with data useful to explore land surface processes under extreme temperature conditions. We therefore first investigate which parts of the temperature distribution have simultaneous measurements of Qle, Qh and NEE for a given site. We then aggregate the answers to this question to ask which sites contain the most measured Qle, Qh and NEE relative to measured temperatures. This question is posed separately for the flux measurements over the whole temperature distribution and for the upper and lower tails of the distribution. We therefore seek to identify which FLUXNET site data are most suitable for analysing processes under extreme temperature conditions with the goal of identifying those sites most useful for land surface model development and evaluation of the surface energy, water and carbon budgets during extreme temperatures.

Figure 1a shows the normalised frequency distribution of temperature, aggregated over all sites. Values range from about -40 to 40 ∘C and are approximately normally distributed. However, the upper tail ends more abruptly than the lower tail. Figure 1a also shows the normalised frequency of Qle, Qh and NEE for different values of temperature. The shapes of the distributions for Qle and Qh are similar and measurements exist across the entire range of sampled temperatures. Not surprisingly, the normalised frequency of measurements for both Qle and Qh are lower than for measured temperature. Notably, the frequency of NEE is much lower than for Qle and Qh. Figure 1b shows the ratio of the number of measurements of Qle, Qh and NEE relative to the number of measurements of temperature. In all cases, the ratios increase as a function of increasing temperature, indicating that fluxes are better sampled for warmer than colder temperatures. At the lowest temperatures, ratios for Qle and Qh range from ∼0 to ∼0.3 but these increase as temperatures increase to maximum ratios of ∼0.8 at around 20 ∘C and remain at ∼0.8 through to 30 ∘C. For NEE ratios increase to ∼0.6 at around 35 ∘C for NEE. A minor dip in Qle and Qh ratios occurs at 0 ∘C associated with the phase change of water, which most likely affects the operation of instrumentation. At the upper extreme of the temperature distribution, ratios decline between 30 and 45 ∘C from ∼0.8 to ∼0.6 for Qle and Qh and from ∼0.6 to ∼0.5 for NEE. However, in each case a secondary peak of high ratios occurs for the very highest temperatures. This peak is associated with temperatures >44 ∘C, which are rare and associated with measurements at Au-Cpr (there are only 68 individual measurements in excess of 44 ∘C at this site), AU-GWW (23 individual measurements), AU-Stp (24 individual measurements) and SN-Dhr (33 individual measurements). Of these, the Australian sites tend to have high measurement ratios and this peak at very high temperatures almost entirely reflects observations from Australian sites. Figure 1b highlights where there are less than 1000 measurements in an individual bin and as expected they occur at the upper and lower tails of the distribution.

Figure 2 shows the geographic distribution of measurement ratios for Qle (Fig. 6 provides the actual ratio values associated with each site and temperature range). The ratio over the whole temperature distribution shows most sites (63 %) exceed 0.7 and some sites (5 %) exceed 0.9 (Fig. 2a). These ratios drop considerably if the lower tail (Fig. 2b) is examined. Since the lower tail is calculated for each site independently this result is not surprising for mid- and high-latitude sites where snow, freezing and frosts would affect measurements. However, this result is more surprising in southern Europe and south-eastern Australia where the lower tail is warm relative to some sites with higher ratios that are colder (e.g. Japan, northern China, Scandinavia). In contrast, for the upper tail, Fig. 1c shows many (67) sites with ratios exceeding 0.9 (see also Fig. 6). While we focus on the US, Europe and Australia, we note sites in Japan, China, South America and Russia with ratios exceeding 0.9. We also note few sites with measurement ratios >0.8 over some regions with very high temperatures, including Africa and the Middle East, and no sites in India, Pakistan and Greece, for example. Figure 3 shows a broadly similar result for Qh although overall the ratios are higher (on average 0.79) than for Qle (on average 0.73). This is most apparent for the upper tail (Fig. 3c), where many of the sites with ratios of 0.8–0.9 for Qle are above 0.9 for Qh.

Figure 4 shows the geographic distribution of measurement ratios for NEE (see also Fig. 8). There is a sharp contrast with the maps of Qle (Fig. 2) and Qh (Fig. 3) and the overall average is 0.42 compared to 0.79 for Qh and 0.73 for Qle. In terms of the overall metric (Fig. 4a), no sites exist with a ratio exceeding 0.9, and only one exceeds 0.8 but 18 exceed 0.7. Two sites located in the eastern US (US-Orv, US-Wi0) exceed 0.7 for the lower tail (Fig. 4b). Multiple sites (11) over North America exceed 0.9 for NEE at the upper tail of temperatures (Fig. 4c) together with isolated sites over Europe (IT-Tor, ES-Ln2), China (CN-HaM, CN-Cha, CN-Dan) and Australia (AU-Ade).

To examine these results further, Fig. 5 shows the measurement ratios as a function of mean annual precipitation and mean annual temperature. Note the amounts of rainfall shown in Fig. 5 are accumulated only over times when temperature data are selected and therefore cannot be compared with observations taken at meteorological stations. Figure 5 shows little relationship between temperature or rainfall and the measurement ratios. For example, some cool dry sites have high measurement ratios whereas others have low ratios. Similarly, some hot wet sites have high and some have low ratios for both Qle and Qh and for the upper tail of NEE. Few sites have high ratios for the overall temperature distribution or for the lower tail of NEE. In other words, the temperature or rainfall at specific FLUXNET sites does not explain why some sites have a high frequency of flux measurements while other sites rarely observe Qle, Qh and NEE. For Qle, Qh and NEE, Fig. 5 also shows the lack of high ratios for the lower tail relative to the upper tail and the low ratios for NEE compared to Qle and Qh. At the upper tail, many sites (e.g. AU-Cpr, DE-Akm and US-NR1) exceed measurement ratios of >0.9 for Qle and Qh. Overall, Fig. 5 shows 5–10 sites with high measurement ratios at temperatures above ∼25 ∘C for the upper tail and for Qle, Qh and (to a lesser degree) for NEE; these are predominantly FLUXNET sites located over Australia.

We finally aggregate our analyses for the overall ratio, the lower tail and the upper tail separately for Qle, Qh and NEE (Figs. 6–8), and we identify each FLUXNET site in terms of the measurement ratio. Figures 6–8 are then combined in Fig. 9 to highlight those sites with high measurement ratios for all of Qle, Qh and NEE and for just Qle and Qh for the overall metric (Fig. 9a), the lower tail (Fig. 9b) and the upper tail (Fig. 9c). Taking the overall statistic first (Fig. 9a, additional details are listed in Table S1), no sites are found with Qle, Qh and NEE ratios exceeding 0.9. Only two sites, both in the US (US-Whs, US-WiO), have measurement ratios above 0.8. If NEE is omitted, 19 sites are selected where both Qle and Qh ratios exceed 0.9 (Fig. 9a, listed in Table S1). These include eight sites over the US; four sites over Australia; two over China; and single sites from Denmark, Germany, France, Italy and Portugal. Even if the threshold is reduced to only Qle and Qh ratios exceeding 0.8, there are still no sites over South America, Africa, and, perhaps critically for high temperatures, over Central America and the majority of the Mediterranean region. The freely available FLUXNET data sets provide no data over India, Pakistan or the Gulf States.

If we are interested in the lower tail of temperatures and we seek sites with measurement ratios exceeding 0.8 for each of Qle, Qh and NEE, we have two choices (US-Orv, US-Wi0). If only Qle and Qh are needed, the choice widens to 18 sites with 7 sites in Australia; 4 in the US; and 1 each in China, Canada and France (Fig. 9b, Table S2). Here, we note that very cold regions are poorly sampled with no sites in Alaska, Russia, the Himalayas, Greenland or Antarctica.

At the upper tail, 16 sites have ratios exceeding 0.9 for each of Qle, Qh and NEE and are in Canada (7), the US (6), China (3), Spain (1), Australia (1) and Italy (1) (Fig. 7c, Table S3). If only Qle and Qh are required above 0.9 there are many sites (32) and above 0.8 there are 3 sites in South America, 1 in Botswana, several in the southern US and southern Europe, and 1 in Israel. No sites remain in India, Pakistan, the Gulf States, Central America and the majority of the Mediterranean region.

We also examined whether the measurement ratio varied by time of day for each site (Fig. S2). These examples are provided to illustrate individual site behaviour and to emphasise that major variations at each site are present. At Au-ASM, a weak diurnal cycle is visible in the measurement ratio with very similar and consistently high ratios of Qle, Qh and NEE being slightly lower. At a second Australian site, AU-Tum measurement ratios increase from dawn throughout the day, and then drop off just before dusk. At CA-NS4 behaviour is similar to AU-ASM until late in the day when the measurement ratios drop sharply. At DE-Hai there is little variation though the day and Qh is much higher than Qle, and only NEE shows any diurnal variation. DE-Meh shows Qle and Qh are consistent throughout the day and are almost identical. DK-NuF shows Qle and Qh falling from dawn to around 10:00 LT, then stabilising at a low value (∼0.3–0.4) and then increasing strongly from 14:00 LT to ratios >0.7 while NEE increases weakly from ∼0.2 gradually though the day. It-Tor shows little diurnal variation in Qle and Qh, but there is a strong diurnal variation in NEE. Finally, US-Whs shows high measurement ratios for Qh and Qle, but falling slightly throughout the day with NEE increasing strongly from dawn to 11:00 LT and then slowly declining throughout the day. If we assume that the hottest part of the day is around 13:00 LT, those sites that provide useful observations of Qh and Qle coincident with these temperatures clearly require site-by-site evaluation. Thus, if sites are being composited, the knowledge that different sites sample different parts of the diurnal cycle, and sample Qle, Qh and NEE differently across the diurnal cycle, needs to be taken into account.

The FLUXNET eddy covariance flux measurements are among the most valuable observations available for understanding processes, and for developing, evaluating and benchmarking land surface models. Under future climate change, warming driven by radiative forcing is likely to be amplified by changes in the partitioning of available energy between latent and sensible heat at the surface (e.g. Seneviratne et al., 2010; Miralles et al., 2014; Donat et al., 2018; Ukkola et al., 2018). This change in the partitioning, linked with soil desiccation or changes in stomatal conductance under higher CO2, provides an amplification of the large-scale meteorology and can lead to more extreme conditions via the coupled land-boundary layer system (Seneviratne et al., 2010; Miralles et al., 2014). As the continental surface warms, some regions will experience temperatures beyond the historical record. Building land models for CMIP-type climate models that properly capture mechanisms and processes occurring in a region experiencing higher temperatures is helped if observations from other regions already experiencing those temperatures are available (so called climate analogues, or space-for-time substitutions). In this context, observations from FLUXNET are particularly valuable if they sample existing hot locations, and if they actually measure fluxes at those locations at the upper tail of temperature.

Our results highlight multiple positives for those wishing to probe vegetation responses to temperate extremes and/or evaluate land surface models. Figure 9 shows many sites with high measurement ratios for Qle and Qh at the upper and lower tail, indicating a rich source of available observations. Conversely, if we seek observations of Qle, Qh and NEE, these data are more limited, with only two sites with a measurement ratio >0.8, none >0.9 at the lower tail and 16 sites at the upper tail (see Tables S2 and S3). Of course, the >0.9 measurement ratio is arbitrary and more sites become available at lower ratios; however, it is somewhat confronting that at >0.8, 87 % of the sites in Table S2 are located in Europe, North America and Australia and for the upper tail, 88 % of the sites in Table S3 are located in these three regions. The sites outside Europe, North America and Australia are not distributed globally: Fig. 9 shows virtually no sites with high (>0.8) measurement ratios in the tropics, Africa or South America for Qle and Qh, and no sites at all in India or the Gulf States. These typically hot regions may be surrogates for how continental surfaces behave under future climate scenarios in the mid-latitudes and it is unfortunate that FLUXNET lacks observations in these regions.

In the absence of measurements from hot regions, the availability of observations from Australia becomes particularly important because these sites cover a wide rainfall gradient, ranging from water- through to energy-limited sites. We note two possible reasons for the lack of freely available data in many regions. First, there may be a lack of sites, or sites that exist may have low measurement ratios. Second, the high number of sites identified in our analysis with high measurement ratios located in Europe, North America and Australia largely reflects the high number of sites in the FLUXNET data. Similarly, the low number of sites in Africa, South America, India and the Gulf States reflects the rarity of FLUXNET sites in these regions. There are, however, four sites in Africa, three in South America and one in Israel in FLUXNET, but these are excluded due to the shortness of the data record, and the low temperature measurement ratios. This is not intended as a criticism; it is a consequence of history (where groups grew with the capacity to maintain measurements and the common desire to run measurement sites near home institutions).

One result from our analysis is that, overall, measurement ratios for Qh are higher than for Qle and both of these are much higher than NEE. This is true for the overall distribution of temperatures, and for the lower and upper tails of the distribution. This result can be quickly visualised by comparing Figs. 6, 7 and 8. In part, this is associated with the actual temperatures at the sites influencing the measurement ratios once aggregated. Figure 5 shows that the measurement ratios are generally lower at the lower tail than the higher tail for Qle and Qh. Furthermore, for the lower tail, the ratios are generally lower at colder temperatures than warmer temperatures. We propose multiple reasons explaining these findings. Qh, Qle and NEE are all products of turbulent transport. While there have been significant improvements in instrumentation over the last 20 years, measurements of these fluxes over long periods and across a range of weather conditions remains challenging.

Measurement ratios of <1 for Qh, Qle and NEE are expected due to data loss caused by instrument failure, precipitation, ambient conditions that violate the assumptions of the eddy covariance method (particularly low- or non-stationary turbulence) and other artefacts (Foken et al., 2010; Burba, 2013). The lower ratios for Qle in comparison to Qh are likely to be associated with measurement methods. The majority of sites use a sonic anemometer and an open-path gas analyser to measure Qh, Qle and NEE. Both devices use measurement techniques over a physical path (sound waves for the sonic and infrared for the open-path gas analyser). Anything that partially obscures the measurement path (condensation, mist, drizzle, snow, ice, etc.) can interfere with the measurements. The sonic anemometers are robust to all but very intense rain but the open-path gas analysers are more sensitive to anything that blocks the optical path (Foken et al., 2010). The Qh measurements only involve the sonic anemometer while Qle and NEE use measurements from the sonic (for vertical velocity component) and from the open-path gas analyser (for water and CO2 concentration). Measurements for Qle and NEE are therefore inherently more complex than for Qh, which explains the lower measurement ratio for Qle relative to Qh.

The lower ratios at lower temperatures are likely to be associated with the occurrence of condensation (dew), which is more common at cooler temperatures – hence the observed dependence of the ratio on measured air temperature. However, the assumptions underpinning the measurement of surface fluxes using the eddy covariance method are violated in low-turbulence conditions, which occurs mostly at night (excluded in our analysis) and low temperatures (e.g. at dawn where radiative cooling leads to a stable surface layer). For fluxes that are significantly different from 0 at night (e.g. NEE due to ecosystem respiration) this leads to an overwhelming bias in the measurements unless low-turbulence conditions, where the assumptions of the eddy covariance method fail, are excluded from the analysis. Therefore, friction velocity (u*) is used as a proxy for turbulence, by finding the site-specific value for u* above which NEE is independent of u* and removing all observations in which u* is below this threshold (Aubinet et al., 2012). This often results in less than 20 % of NEE data being available for estimating ecosystem respiration. The application of this turbulence filter causes the ratio for NEE to be much lower than the ratio for Qh and Qle. The occurrence of these conditions is more likely in lower-temperature conditions, contributing to the slope in Fig. 1b. We avoid the consequences of these procedures in quality-controlling and gap-filling data by only using those data that are directly observed.

Our analysis has a specific weakness, which requires consideration when interpreting our results. There may be a temptation to interpret the ratios we report as a metric linked with measurement quality. To discourage such a temptation we draw attention to two hypothetical FLUXNET sites, one with ratios around 0.9 and another around 0.3. In the former, the efforts around measurement quality are superficial and data are included unless a specific problem identified. At the latter, the efforts around measurement quality are rigorous and any doubts whatsoever about the data lead to it being discarded. For the latter case, one would suggest that the resulting data reported to the FLUXNET2015 or La Thuile archives are likely to be of the highest quality and most reliable to use in process-level examination of models or understanding of the surface energy and carbon balance. The more complete data in the former example could in fact be misleading. In short, our analysis does not report on data quality, it only relates to coincident data availability and identifies those sites where measurements are available with high frequency and with a QC =0.

Our methodology contained several assumptions, for example we excluded sites with less than 8 months of data. We tested the sensitivity to this assumption, examining whether the sites identified with high measurement ratios changed if we required 12 months of data. If we set a minimum length of record as 12 months, US-Wi0 (one of two sites with Qle, Qh and NEE >0.8), US-SP1, US-Orv and ES-Ln2 are excluded in Table S1. The only sites with Qle, Qh and NEE >0.8 are excluded from the lower tail (US-Orv and US-Wi0), along with DK-Fou, US-SP1 and NL-Lan. At the upper tail multiple sites (AU-Rob, PT-Mi1, NL-Lan, Es-Ln2, US-Wi0, US-SP1 and US-Bar) are excluded. Therefore, requiring a 12-month data set has a significant impact on some of the otherwise most useful sites. Given the purpose of our analysis is to examine the tails of the distributions at each site, we suggest that imposing longer measurement periods than absolutely required may prove counterproductive. In addition, we examined two other attributes of the FLUXNET data – whether our measurement ratio changes between the first half of the data and the second half (i.e. to examine whether the measurement ratio improved over time) and whether any relationship exists between the total number of QC =0 observations and the measurement ratio. The first analysis found no evidence that higher measurement ratios were apparent in the first or second half of the data, something that might have been expected if the ability to sustain measurements improved over time. The second analysis also found no evidence of a relationship between the measurement ratio and the length of data (Fig. S1).

One obvious criticism of our measurement ratio metric is the temptation to interpret the results as a way to select FLUXNET sites for model development and evaluation without further thought. Clearly, a high measurement ratio is only one aspect of a valuable data set. A modeller might, for example, prefer a large number of actual measurements with a low overall measurement ratio rather than a site with few measurements but a high overall measurement ratio. We have noted above that we find no correlation between data length and measurement ratio but some sites (see Tables S1–S3) have both high measurement ratios and large amounts of data and others have high measurement ratios and low amounts of data. For example, the two sites with the highest measurement ratios overall (US-Whs and US-Wi0) sharply contrast in the amount of data (63 619 and 4621 temperature measurements respectively). In this case, US-Whs covers 2922 d of measurement and 93 % of the time temperature data are reported (Table S1), whereas US-Wh0 only measures for 365 d and only 62 % of the time temperature data are reported. In contrast, sites such as CA-NS1 and CA-NS3 display very similar measurement ratios for Qle, Qh and NEE: both cover 1826 d but CA-NS1 includes 30 269 temperature measurements while CA-NS3 includes only 22 689 temperature measurements. Clearly, many characteristics of a data set make it valuable for model development or model evaluation and our analysis should be viewed only as one of these characteristics. One way forward to resolve how to choose FLUXNET data for extremes is to combine an analysis of meteorological sites with FLUXNET sites. Using sites maintained by meteorological agencies to identify extreme events (e.g. heat waves) and then interrogate the FLUXNET sites near to the meteorological site for the availability of measurements of Qle, Qh and NEE could enable a modeller to choose suitable sites for land surface model development and evaluation. While one possible way forward, inconsistencies between observations from meteorological agencies relative to FLUXNET (location, geographical distribution, height of measurements, standardisation of measurements over short grass) highlight the challenges in using meteorological observations that are physically separate from the FLUXNET observations.

Our analysis poses interesting questions about the FLUXNET data that deserve further exploration. Why do sites with a similar climate vary so greatly in terms of their frequency of reporting of Qle, Qh and NEE in comparison to temperature? Why are some sites able to do this routinely while others cannot, and can expertise be shared to resolve this? What are the implications of aggregating FLUXNET data given the large variations in which parts of the temperature distribution are sampled? Why are there major variations in the measurement ratios between sites over the diurnal cycle and what does this mean in terms of using site data from FLUXNET? Clearly, the FLUXNET data do provide our best ecosystem-scale estimate of the vegetation's response to heat extremes (Ciais et al., 2005; Teuling et al., 2010; Wolf et al., 2013; von Buttlar et al., 2018; Flach et al., 2018; De Kauwe et al., 2019) but given the need to build land models representing extreme conditions these data cannot be used without further evaluation of the specific site data. We do not know if there are opportunities for the global community to prioritise new sites in regions that currently lack data, or directly support those measurements in regions with low measurement ratios. However, we suggest investment in either new sites or in existing sites in countries that experience temperatures that are higher than those experienced across North America and Europe to enable land models to be developed in anticipation of further warming. Virtually all sites (∼90 %) with high measurement metrics for Qle, Qh and NEE, or just Qle and Qh, whether examining the whole distribution or just the lower tail or just the upper tail, are located in North America, western Europe and Australia. There are no sites in India, South America, Africa or the Middle East and few sites in China. In terms of vulnerability, the freely available FLUXNET data therefore cover regions representing 12 %–14 % of the global population. Indeed, the poorest country with measurements (based on gross domestic product, Portugal) suggests all countries ranked from Portugal (ranked 47th) to the poorest country (ranked 211th) lack any measurements. Another perspective is if countries are ranked on average temperature, none of the warmest 98 countries contain a site and Australia is the hottest country with sites with high measurement ratios. Conversely, North America, western Europe and Australia have multiple sites with observations of Qle and Qh and some with NEE with high measurement ratios for both the lower and upper tail of the temperature distribution. For these three regions, therefore, FLUXNET data provide a rich source of data for understanding how fluxes of energy, water and carbon behave under extreme temperature conditions. Overall, we have noted more frequent observations of Qh than Qle and both these fluxes are much more common than NEE. An implication of this is that some regions, particularly very hot regions that will be the first to experience novel climates, require observations. We also highlight a wide discrepancy between the measurement ratios across FLUXNET sites that is not related to the actual temperature or rainfall at the site.

We have examined the FLUXNET data by evaluating the availability of Qle, Qh and NEE observations at time steps where temperature is measured (with a quality control flag QC =0). We have analysed this spatially to identify those sites with a high availability of flux measurements, relative to temperature measurements, across the whole temperature distribution, and at the upper and lower tails of the distribution.

Virtually all sites (∼90 %) with high measurement metrics for Qle, Qh and NEE, or just Qle and Qh, whether examining the whole distribution or just the lower tail or just the upper tail, are located in North America, western Europe and Australia. There are no sites in India, South America, Africa or the Middle East and few sites in China. This discrepancy between the measurement ratios across FLUXNET sites is not related to the actual temperature or rainfall at the site. Clearly, some sites seem able to retrieve Qle, Qh and NEE reliably at extreme temperatures while others cannot. This may provide an opportunity for the FLUXNET community to share best-practice strategies to identify ways to ensure measurements at the tails of the temperature distribution.

Finally, we restate a key caveat to our paper to avoid any misunderstanding. Our analysis does not highlight the “best data”. A site might have high ratios because of poor QC control, or low metrics because of strict controls. However, our paper does highlight sites with frequent observations of Qle, Qh and NEE coincident with temperature observations where all have a QC =0. A modeller might of course reject some of these sites for reasons of data record length, vegetation type, soil type or a multitude of other reasons. However, we suggest that our analysis provides one way for modellers to identify sites from the FLUXNET archive that warrant closer scrutiny for development and evaluation of land surface models under extreme temperature conditions.