Barometric pressure (in hPa) is measured in the fibreglass-reinforced polyester logger enclosure (Fig. , number 9). The logger enclosure is generally located 1.5 m above the ice surface. The barometer manufacturer reports a measurement accuracy of ±2 hPa within the -40 to +60 ∘C temperature range (Table ; see also the Appendix for more information).

Air temperature (in ∘C) is measured inside a fan-aspirated radiation shield (Fig. , number 6). The sensor is located approximately 2.6 m above the ice surface (i.e. as high as possible underneath the sensor boom). The measurement height varies when a winter snow cover is present. The temperature sensor is a PT100 probe that changes its electrical resistance with temperature and has an accuracy of ±0.1 ∘C (Table ; see also the Appendix for more information). A secondary air temperature reading (in ∘C) is made in the aspirated shield from the HygroClip temperature/humidity sensor described in the following, which also has a manufacturer-stated accuracy of ±0.1 ∘C, but we consider the HygroClip temperature to be less accurate than the PT100, given the need for more frequent sensor recalibration.

Relative humidity (RH; in %) is measured alongside the PT100 in the aspirated radiation shield using a HC2A-S3 (or HC2) HygroClip (Fig. , number 6). The sensor measures relative humidity with ±0.8 % accuracy. Relative humidity is measured relative to water. For temperatures below freezing, relative humidity is recalculated relative to ice in post-processing (see Sect. ). To distinguish between the two relative humidities in the PROMICE data products, the prior humidity (unadjusted below freezing) is called “relative humidity with respect to water”, whereas the latter is simply referred to as “relative humidity”. The conversion of relative humidity relative to ice is after . Every 1–2 years, the HygroClip is replaced by a sensor recalibrated in a closed chamber at room temperature with constant relative humidities of 10 %, 35 %, and 80 %.

Wind speed and direction (in ms-1 and degrees respectively) measurement height is approximately 3.1 m above the ice surface and, like the other measurements, has a reduced measurement height if a winter snow layer is present (Fig. , number 4). An AC sine wave voltage signal is produced by the rotation of the four-bladed propeller, and the pulse count converts to wind speed using a multiplier. According to the manufacturer, the sensor can measure wind speeds between 0 and 100 ms-1, with an accuracy of ±0.3 ms-1 or 1 % if the measured value is higher than 30 ms-1.

Wind direction is measured through changes in the vane angle by a precision potentiometer housed in a sealed chamber on the instrument. The output voltage is directly proportional to vane angle wind direction and is measured between 0 and 360∘ with an accuracy of ±3∘. Every 3 years the sensor is replaced and tested for drift and functionality with an “anemometer drive” rotating the propeller at a known rate. The instrument's orientation is logged and reset to “geographic north” during each maintenance visit to keep wind direction data accurate within ±15∘ (although much larger station rotations have been encountered).

Horizontally levelled up- and down-facing Kipp & Zonen CNR1 or CNR4 record solar radiation (in Wm-2) respectively. Measurement height is at the sensor boom level of 2.7 m over the ice surface (Fig. , number 1). Short-wave radiation is measured by the pyranometers within plastic meniscus domes, allowing minimal water droplet adhesion. The manufacturer reports that sensor uncertainty is 10 %. In practice, this sensor uncertainty has been found to be ca. 5 % for daily totals in Antarctica . The radiometers are recalibrated at Kipp & Zonen every 3 years. The radiometer is one of the few variables stored in the data logger in voltage (V) units, because every radiometer has a different set of calibration coefficients, whereas all logger programs running on PROMICE AWSs are identical, for practical reasons. In post-processing, sensor readings SRraw are converted into a physical measurement SRm as follows: 1SRm=SRrawCSR, where CSR (in V(W m-2)-1) is a sensor calibration coefficient, and SRm is either the converted downward or upward short-wave irradiance. Short-wave radiation measurements are corrected for sensor tilt following in post-processing, which means that the PROMICE AWS dataset contains both uncorrected and corrected values.

Long-wave radiation (in W m-2) is also measured by the CNR1/CNR4 radiometer mounted at approximately 2.7 m over the ice surface (Fig. , number 1). The radiometer contains a pair of up- and down-facing pyrgeometers, with a spectral range of 4.5 to 42 µm. In the same manner as for short-wave radiation, long-wave radiation is stored in voltage units (LRraw) in the data logger and transformed to physical units (LRm) in post-processing as follows: 2LRm=LRrawCLR+5.67×108⋅(Trad+T0)4, where CLR (in V (W m-2)-1) is the sensor calibration coefficient, Trad is the sensor temperature measured in the radiometer casing (in ∘C), and T0=273.15 ∘C.

The height of the sensor boom (in metres) is measured by a sonic ranger attached to the boom itself approximately 0.1 m below the boom (Fig. , number 5a), while the height of the stake assembly is measured about 0.1 m below an aluminium boom connecting stakes drilled into ice (Fig. , number 5b). The sensor outputs a distance (Hraw) that requires an air temperature correction in post-processing. The temperature adjustment is performed as follows: 3Hm=Hraw⋅Tair+T0T0. After temperature correction, the measurement uncertainty of the SR50A sonic ranger reported by the manufacturer (Campbell Scientific) is ±1 cm or ±0.4 % of the measured distance. The uncertainty of sonic ranger readings in PROMICE was investigated utilizing data from a wintertime accumulation-free period of more than 2 months at the location SCO_U. The associated standard deviations for the two sensors were found to be 1.7 and 0.6 cm after spike removal, amounting to 0.7 % and 0.6 % of the measured distance respectively . In addition to the sensor uncertainties, occasional problems with the stake assembly occurred, primarily in terms of stability during storms when melted out several metres. Also, an unknown amount of melt-in of the stake assembly can occur, but we speculate that this only happens (1) when surface melt since installation has been considerable, increasing the height and, thus, the pressure applied by the stake assembly, and (2) when the stake bottoms are not plugged with caps, as was only the case until 2010.

The PROMICE AWSs are also equipped with a pressure transducer assembly (PTA) that measures surface height change due to ice ablation (Fig. , number 7). The assembly was first constructed and implemented in Greenland in 2001 by but was further developed within PROMICE . The PTA consists of a 50 / 50 antifreeze / water mixture-filled hose with a pressure transducer attached at the bottom. Drilling the hose typically more than 10 m into the ice, the pressure signal registered by the transducer will be that of the vertical liquid column over the sensor, where the upper level is a bladder fixed on the tripod in a shielded box. This allows inflow/outflow of antifreeze due to compression while keeping a steady level at roughly 1.5 m above the ice surface depending on the AWS. Figure  illustrates the free-standing AWS tripod that floats on the ice surface and moves down with the ablating surface, whereas the hose itself melts out of the ice, which, in turn, will reduce the hydrostatic pressure from the vertical liquid column over the pressure transducer at the bottom of the hose. The measured reduction in pressure at the bottom of the hose translates directly into ice ablation. As for the radiometer, every pressure transducer has a different calibration coefficient, which is why measurements are stored in the data logger in voltage units and transformed to a physical measurement in post-processing. Measurement height (Hm), or in fact depth relative to the PTA bladder, is calculated as follows: 4Hm=CPTA⋅ρwρaf⋅Hraw, where CPTA is the calibration coefficient. The constants ρw and ρaf are the densities of water and the 50 / 50 antifreeze / water solution respectively.

Subsurface temperatures (in ∘C) are measured by a 10 m thermistor (temperature-dependent resistor) string (Fig. , number 11). The string measures at 1, 2, 3, 4, 5, 6, 7, and 10 m depth, although depths vary due to the surface ablation and accumulation. The string is constructed at GEUS (see the Appendix for more information).

The inclinometer is installed on the sensor boom (Fig. , number 2) and is aligned with the radiometer to allow for tilt correction of short-wave radiation measurements. The inclinometer measures the tilt (in degrees) across (left–right) and along (up–down) the sensor boom, which translates into tilt-to-east and tilt-to-north when the sensor boom is perfectly oriented north–south. The tilt sensor readings in voltage units (Tiltraw) are converted into tilt in degrees as follows: 5Tiltm=21.1⋅|Tiltraw|-10.4⋅|Tiltraw|2+3.6⋅|Tiltraw|3-0.49⋅|Tiltraw|4, where all constants were determined at GEUS (Table ). Ice ablation causes the AWS tripod to melt downward; this changing (slippery) surface often results in AWS tilt changes of more than several degrees.

We use a single-frequency GPS receiver to measure the position (in ∘ N/∘ W) and the elevation (metres above sea level) of each station to quantify ice flow velocity (Fig. , number 9). The GPS antenna, as well as the receiver contained in the Iridium 9602-LP modem, is placed inside the data logger enclosure. The receiver type is described as follows: NEO-6Q, 1575.42 MHz (L1), 16-channel, and C/A code. The accuracy is reported to be within 2.5 m. In the PROMICE AWS set-up, the GPS receiver is powered up for 5 min preceding each Iridium transmission (hourly in summer and daily in winter), during which it attempts to acquire location data every 20 s. The return (out of a maximum of 15) that reports the lowest horizontal dilution of precision is written to memory. To date, NUK_U, NUK_L, MIT, and QAS_L have been repositioned during maintenance visits over distances larger than several tens of metres. The main reason for this is to reduce the influence of location change on the AWS variables measured, but stations have also been relocated to move them away from a region with opening crevasses. Table  shows the horizontal and vertical displacement due to glacier flow and AWS relocation during maintenance visits.

Table  provides filtering information used in the processing chain. We remove unrealistic spikes from the data by using upper and lower thresholds for each measurement. Measurements outside these (generous) threshold limits, which could occur for a number of known and unknown reasons, are considered erroneous and set to -999. Known reasons will be discussed in Sect.  (living data section). Derived variables are also set to -999 when one or more of the listed “core” AWS measurements that serve as input fall outside the threshold limits.

The specific humidity q (in kg kg-1) is calculated from relative humidity with respect to water/ice above/below freezing (RH) using the following equation: 6q=RH100⋅qsat, with 7qsat=ϵ⋅esice/waterp-(1-ϵ)⋅esice/water, where ϵ=0.622 is the ratio between the specific gas constants for dry air and water vapour, p is air pressure (in Pa), and esice/water is saturation water vapour pressure (in Pa) over ice (below freezing) or water (above freezing) calculated after .

The surface temperature Ts (in ∘C) is derived using the measured downward and upward long-wave irradiance (LRin and LRout respectively): 8Ts=LRout-(1-ϵ)⋅LRinϵ⋅5.67×10-80.25-273.15, where ice sheet surface emissivity ϵ=0.97.

The sensible and latent heat fluxes (SHF and LHF respectively; in W m-2) are estimated using vertical gradients in wind speed, potential temperature, and specific humidity between the measured boom height and the surface described by and . According to the Monin–Obukhov similarity theory, SHF and LHF can be approximated as follows: 9SHF=ρCpκ2uln⁡zuz0-ψuT-Tsln⁡zTz0,T-ψT,10LHF=ρLs/vκ2uln⁡zuz0-ψuq-qsln⁡zqz0,q-ψq. Here, ρ is the density of air, and Cp=1005 JK-1kg-1 is the specific heat capacity at constant pressure. Ls=2.83×10-6 Jkg-1 and Lv=2.50×10-6 Jkg-1 are the latent heat values of sublimation and evaporation respectively, and κ=0.4 is the von Kármán constant. When estimating turbulent heat fluxes, we need the measurement heights (zu, zT, zq; Table ) of wind speed (u), temperature (T), and specific humidity (q) as well as the surface roughness lengths for momentum z0, for heat z0,T, and for moisture z0,q. We use z0=0.001 m, and z0,T=z0,q is calculated using the formulation from for rough surfaces. We use the stability correction functions ψu,T,q from Eq. (12) in for stable atmospheric conditions, and we follow for unstable conditions. The surface temperature (Ts) is calculated from long-wave radiation (see Eq. ), and the surface specific humidity is assumed to be at saturation (qs=qsat).

Several sources of uncertainty apply to the calculation of SHF and LHF. The aerodynamic surface roughness length z0 is known to vary with surface type and through time . Using the constant value of z0=0.001 m could be an overestimation of surface roughness in the presence of snow and could subsequently lead to an overestimation of both turbulent fluxes. As most PROMICE stations are located in the ablation area, the snowpack is melted during spring and the surface becomes snow-free for most of the ablation season. The calculation of surface temperature also relies on certain assumptions (see the section above). Several studies have evaluated the performance of the Monin–Obukhov similarity theory in Greenland. Using one- and two-level methods vs. eddy covariance and evaporation lysimeters, found an underestimation of downward LHF during extreme stability cases. used a similar method for calculating SHF and reported a root-mean-square difference (RMSD) of 8.7 Wm-2, with an average bias of -7.0 Wm-2, when compared with their two-level eddy-covariance estimation of SHF. emphasized that SHF records from one-level approaches often cover longer time periods. investigated the use of an unrealistically high z0 to get agreement between surface energy balance (SEB) closure and observed ablation rates during extreme sensible and latent heat-driven melt events.

Tilt correction of solar radiation is performed following . Downward short-wave radiation (SRin) consists of a diffuse and direct beam part. It is only the direct beam part of SRin that requires tilt correction. For a horizontal radiation sensor, the direct beam, which equals SRin, is reduced by its diffuse fraction (fdif). For the tilted radiation sensor, SRin is calculated from the measured value, SRin,m, and a correction factor, C, as follows: 11SRin,cor=SRin,mC1-fdif+Cfdif, with 12C=cos⁡(SZA)⋅(sin⁡(d)sin⁡(lat)cos⁡(ϕsensor)-sin⁡(d)cos⁡(lat)sin⁡(θsensor)cos⁡(ϕsensor)+cos⁡(d)cos⁡(lat)cos⁡(θsensor)cos⁡(w)+cos⁡(d)sin⁡(lat)sin⁡(θsensor)cos⁡(ϕsensor)cos⁡(w)+cos⁡(d)sin⁡((θsensor)sin⁡(ϕsensor)sin⁡(w))-1, where SZA is the solar zenith angle, d is the sun declination (the angle of the sun above the plane formed by the Earth's Equator), w is the hour angle (the angle between the sun's current position in the sky and its position at solar noon), lat is the site's respective latitude in radians, and θsensor and ϕsensor are the radiometer's tilt angle and direction respectively. The calculation procedures for d, w, and SZA are detailed in . Table  illustrates the average bias or correction made for the incoming solar radiation based on Eq. (). The standard deviation indicates that the average correction is minor (below 15 Wm-2) for most AWSs, whereas a few AWSs have corrections values spread out over a wider range.

We estimate fdif spanning from 0.2 for clear skies to 1 for overcast conditions, while assuming a linear dependency on the cloud cover fraction . We approximate the cloud cover fraction from the dependence of the near-surface air temperature (Tair) on LRin . For this purpose, we calculate a theoretical downward long-wave radiation flux corresponding to clear-sky conditions using the equation from : 13LRclear=5.31×10-14⋅(Tair+T0)6. We calculate a theoretical downward long-wave radiation flux corresponding to overcast conditions assuming the black-body radiation as follows: 14LRovercast=5.67×10-8⋅(Tair+T0)4. The cloud cover (limited to the [0:1] range) is then calculated as 15cloudcov=LRin-LRclearLRovercast-LRclear=fdif-0.20.8.

Surface broadband solar reflectivity in the 0.3 to 2.5 µm wavelength range, also known as albedo (unitless), is calculated from 10 min tilt-corrected downward and upward solar irradiance data. Hourly averaged albedo values are calculated for cases when the sun hits the radiometer top at angles exceeding 20∘ (i.e. when measurements are most reliable for this sensor type). Daily albedo averages are computed from available hourly data. AWS obstruction of sunlight, casting a shadow within the radiometer's field of view, may lower the albedo on average by 0.03 , but this depends on the surface type and height. Also of relevance to measured albedo is the contrast of the surface relative to the AWS battery box, legs, mast, and enclosure, as well as whether a melt pond forms beneath the AWS. examined spatial variograms in unoccupied-aerial-vehicle-derived albedo vs. satellite and PROMICE albedo and found increasing differences for some PROMICE sites toward the late melt season when the AWS point measurements lack representativity of the increasingly inhomogeneous surface cover. A study by found a 5 % uncertainty on pyranometer measurements, although the manufacturer, Kipp & Zonen, estimates a more conservative value of 10 % uncertainty. We conservatively assume 10 % uncertainty in the calculated albedo.

The pressure transducer assembly (PTA; Fig. , sensor 7) set-up is influenced by variations in air pressure. The air pressure contributions to the measured PTA signal HM are eliminated using the following equation: 16HL=HM+PC-PAgρl, where PA (in hPa) is air pressure, PC (in hPa) is the known pressure given by the manufacturer to which the sensor was calibrated, g=9.82 m s-2 is the gravitational acceleration, and ρl=1090 kg m-3 is the antifreeze mixture density at 0 ∘C. Changes in HL are equal to ice ablation. compared PTA time series to hose measurements manually performed in the field and recorded distances from sonic rangers to quantify instrument inaccuracies, which were found to be accurate to within 0.04 m.

The time reported in our data products specifies the hour/day/month during which the measurements are taken, as opposed to other products that list the exact timestamp of the end of the averaging period. Hourly averages are calculated from 10 min values if at least one value is available (10 min data are seldom missing). We then calculate daily averages from hourly averages if at least 20 values (∼80 %) are available for a dataset variables with a clear diurnal variability. Less transient variables require at least one measurement to calculate an average. Lastly, we calculate the monthly averages from daily averages if at least 24 values (∼80 %) are available.

To illustrate the PROMICE AWS data coverage, we determined the “success rate” in terms of available daily averages for all measured variables that are required for estimating the surface energy budget: air pressure, air temperature, humidity, wind speed, and downward and upward short-wave and long-wave radiation. Success rate is defined as the ratio of the counts of successful variable estimate and the number of days since AWS installation. The performance for the critical variables for each station and their measurement periods are illustrated in Fig. . A total of 18 of the 26 stations have at least an 85 % success rate for all critical surface energy budget variables, while 6 have experienced significant periods with power failure, station toppling, snow accumulation exceeding the instrument height, or even crevasse formation underneath the station.

Time series spanning the years 2012 through 2014 of weekly median wind speeds and maximum 10 min wind speed within that week for TAS_L and UPE_U are displayed in Fig. . Median wind speeds are lower at TAS_L than at UPE_U, whereas the opposite is true for maximum wind speeds because TAS_L is located in a region well-known for its piteraq storms.

The daily average air temperature for the two stations near Upernavik is shown in Fig. . The temperature is higher at the lower station, UPE_L, than at the upper station, UPE_U, due to an elevation difference of more than 700 m. The tendency of the temperature to have a higher variability during winter months than during summer is also evident from these time series.

Figure  presents the surface energy fluxes at UPE_U in 2012. The plots show how UPE_U experiences a shorter period with solar radiation, due to the more northerly location, when comparing it to the TAS stations further south. Furthermore, Fig.  shows how the outgoing long-wave radiation becomes stable during the main melt season when the surface temperature is at the melting point. The sum of all the fluxes determines if there is a surplus of energy at the surface, which can be used for snow or ice melt.