## Introduction

Based on the first law of thermodynamics, the energy balance, Eq. 1, states that the net radiation (R n) available to a patch of land surface is consumed in the exchange of latent (λE) and sensible (H) heat with the atmosphere and the change of heat storage within the system (S):

$${\mathit{\mathsf{R}}}_{\mathsf{n}}=\lambda \mathit{\mathsf{E}}+\mathit{\mathsf{H}}+\mathit{\mathsf{S}}$$
(1)

R n depends on the net difference between down- (↓) and up-welling (↑) short- (S) and long-wave (L) radiation, i.e.

$${\mathit{\mathsf{R}}}_{\mathsf{n}}=\downarrow \mathit{\mathsf{S}}-\uparrow \mathit{\mathsf{S}}+\downarrow \mathit{\mathsf{L}}-\uparrow \mathit{\mathsf{L}}$$
(2)

A key component of R n is the ratio of up-welling to down-welling shortwave radiation termed albedo (α):

$$\alpha =\uparrow \mathit{\mathsf{S}}/\downarrow \mathit{\mathsf{S}}$$
(3)

Different types of land surfaces differ in their R n which, through Eq. 1, determines how much energy is available for λE, H and S, which in turn critically affects the near-surface climate (e.g. Stegehuis et al. 2013; Seneviratne et al. 2006). For example, it was shown by Bonan (2008) and Bala et al. (2007) that grasslands and croplands, as opposed to forests, have a cooling effect at higher latitudes because the albedo of grasslands and croplands is typically higher, in particular when covered by snow, compared with forests, which absorb more solar energy. In contrast, in tropical regions, the difference in albedo between forests and grasslands is compensated by the cooling through the large amount of water transpired by (tropical) forests. In order to understand how past (e.g. Brovkin et al. 2006) and potential future (e.g. Bala et al. 2007; Brovkin et al. 2009) changes in land use affect the Earth’s climate, it is crucial to understand how changes in land surface properties affect R n and the partitioning into λE, H and S. For example, it was shown by Chapin et al. (2005) that warming-induced shorter periods of snow cover in the Arctic and associated trends of shrub/tree expansion are likely to cause local warming similar in magnitude to the warming expected from a doubling of atmospheric carbon dioxide concentrations.

The surface energy balance and its components can be quantified by a hierarchy of methods across spatial scales: At the largest scale, merging several satellite data streams with models allows estimating all four components of the energy balance (e.g. Diak et al. 2004; Kalma et al. 2008; Glenn et al. 2007) on a global scale. At the scale of catchments, evapotranspiration may be deduced on an annual basis by difference between precipitation and discharge (e.g. Peel et al. 2010). At the ecosystem-scale, i.e. typically a few hectares characterised by similar vegetation and soil, micrometeorological methods, such as the eddy covariance technique (Baldocchi et al. 1988; Aubinet et al. 2000), allow the direct quantification of both H and λE, with R n and S typically being estimated on/from the tower which supports the turbulence equipment (fast-response sonic anemometer and hygrometer). Within the FLUXNET network, the four terms of Eq. 1 are presently measured continuously at >400 sites globally (Baldocchi et al. 2001; Williams et al. 2012). Finally, at the plot, single plant and leaf scale, sap flux (Wilson et al. 2001), various types of chambers and lysimeters (Wohlfahrt et al. 2010a) can be used to quantify (evapo)transpiration.

In this comprehensive hierarchy of methods, it is the lower end of the microscale (Orlanski 1975), that is, spatial variability at the scale of square meters, which is presently poorly represented (e.g. Ahrends et al. 2012). Landscape variability at this spatial scale is much smaller than the typical pixel size of remote sensing-based approaches and also considerably smaller than the typical footprint of micrometeorological measurements. The only approaches suited for this spatial scale, lysimeters and ecosystem chambers, on the other hand, are generally impractical for surveying a large number of distributed samples within the footprint of eddy covariance flux measurements or in a landscape context.

We thus argue that, in micrometeorological, catchment hydrological and landscape ecological studies, there is the need for the development of approaches for spatially distributed energy balance measurements which can be applied at the lower end of the microscale and yet are portable enough to allow making a large number of spatially distributed measurements within short periods of time. To this end, we propose a mobile device which allows quantifying the small-scale (a few square meters) spatial heterogeneity of the energy balance over short-statured (<1 m) canopies. In the following, we first present the design of the mobile device, followed by four case studies which are meant to illustrate its potential and conclude with a discussion of its strengths and weaknesses, as well as an outlook on potential future developments.

## Material and methods

In addition to the four components of R n (up- and down-welling short- and longwave radiation; W m–2), the data logger outputs the net radiometer body and infrared surface temperature (°C), air and soil temperature (°C), soil moisture (% volumetric soil moisture based on general calibration for mineral soil and raw mV output), wind speed (m s–1) and direction (°) as well as a digital sonic data quality flag.

The protocol at each measurement point is the following: First, soil temperature and moisture sensors are put into place. Then the operator gets into position by pointing the net radiometer horizontally (or slope-parallel; see below) towards South and waits 2 min before taking three (pseudo-)replicate measurements at the same spot. The 2-min delay accounts for the time response of the various sensors. The soil moisture sensor and the sonic anemometer have no quoted time response, while the air temperature/humidity sensor and the net radiometer have a quoted time response of <20 s (90 % response). The soil temperature sensor has a quoted response time of <80 s in air at a wind speed of 1 m s−1 (63 % response), no indications are given for response times in soil, which are likely to be longer. Following data acquisition, time and place of the measurement and environmental conditions (e.g. cloud cover) are noted in a field book, and any additional measurements are made on the plot (see case study 3 below for an example).

## Results and discussion

In the following, we illustrate the potential of the EcoBot by reference to four selected case studies:

### Case study 1: within eddy covariance footprint heterogeneity of Rn

In eddy covariance energy flux studies, R n and S are typically measured either on the tower which supports the turbulence equipment or a nearby additional tower, and in the vast majority of cases, measurements are made at a single location only. As the footprint of eddy covariance flux measurements is typically orders of magnitude larger than the footprint of R n and S (Schmid 1997), any analyses of H and λE that make use of single-point R n and S rely on the implicit assumption of their values in the flux footprint being equal those measured on the tower. Case study 1, shown in Figs. 2 and 3, was selected as an example illustrating a situation where the tower-based measurements of R n differ from R n in the footprint due to spatial heterogeneity in vegetation cover caused by land use. Briefly, eddy covariance H and λE, R n and soil heat flux (G; assuming other heat storage to be negligible) measurements were made from a 2-m tower above grassland ca. 20 km to the East of Innsbruck (Austria). The site was situated in the middle of the flat Inn Valley in an area dominated by intensively used grasslands interspersed with various crops (mostly vegetables; Fig. 2). In order to explore the within footprint heterogeneity of R n, mobile measurements with the EcoBot were conducted on a sunny day (10 May 2012) at the seven dominant land use types within the eddy covariance footprint. To this end, one representative plot was selected within each of the seven land use types and revisited every 30 min and between 8 and 16 UTC and three pseudo-replicate EcoBot measurements made. Figure 3 shows that mobile down-welling shortwave and longwave radiation agreed with the flux tower to within their temporal variability (data from the flux tower were saved as averages and standard deviations over 30 min). In contrast, up-welling shortwave (and thus albedo) and longwave (and thus infrared surface temperature) radiation differed by up to 85 and 35 W m−2, respectively, between the seven major land-use types and the stationary measurements (Fig. 3). Due to compensating effects between larger/smaller up-welling radiation fluxes, R n at individual plots differed from the stationary measurements at the flux tower by up to ±60 W m−2. Depending on the aerial extent of the various land surface types and their contribution to the flux footprint (Fig. 2), these differences may need to be accounted for when relating R n to latent and sensible energy fluxes or when attempting to close the energy balance (Foken 2008). Doing so will require a two-dimensional footprint model (e.g. Detto et al. 2006; Kljun et al. 2002), which allows weighting R n of the various land use types with their flux contribution.

### Case study 2: estimating slope-parallel Rn

Measurements of R n are typically made horizontally, assuming a horizontal underlying surface. However, in case of measurements above sloping terrain, slope-parallel measurements are required to be able to relate R n to latent and sensible heat fluxes (Whiteman et al. 1989). Algorithms for correcting horizontal R n measurements for slope and aspect of the underlying non-horizontal surface exist, but, however, usually account only for differences between the angle of incident direct solar radiation and the surface (e.g. Matzinger et al. 2003) and, similar to case study 1, do not account for heterogeneity in slope and aspect within the flux footprint (but see Hammerle et al. 2007).

In case study 2, the EcoBot was used to investigate differences between horizontally and slope-parallel measured R n and to quantify the reliability of approaches to correct for slope and aspect. The study site was again ca. 20 km to the East of Innsbruck (Austria), but this time on a grassland site situated high up on a North facing slope with an average inclination of 30°. EcoBot measurements were made on a clear day (14 May 2013) for three times during the day (morning, noontime, afternoon) at seven positions around the flux tower characterised by different slopes and expositions. Two measurements, each with three pseudo-replicates, were made at each plot—the first one horizontally and a second one with the net radiometer approximately inclined according to local slope and aspect based on a manual assessment of the operator. The fraction of diffuse radiation was quantified continuously with quantum sensor (BF3H, Delta-T, UK) on the flux tower. As shown in Fig. 4, it is obvious that horizontal measurements overestimated R n measured slope-parallel on this steep North-facing slope by a factor of almost 2. Correcting for local slope and aspect following Hammerle et al. (2007) reduced the discrepancy to the slope-parallel measurements (on average down to 3 %; Fig. 4); however, from the spread of data (differences up to 100 W m−2), it is clear that the correction did not completely remove the bias at all locations and times. Apparently, there are other local factors, such as the fraction of cold sky/warm vegetation seen by the pyrgeometers, which vary within the footprint and are not well captured by the common approach of correcting only for the angle between the incident direct short-wave radiation and the underlying surface. In addition, the manual assessment of local slope and aspect is likely to introduce uncertainty, which might be reduced by adding an electronic tilt sensor to the EcoBot capabilities in the future.

### Case study 3: drivers of landscape-scale variability in soil temperature and albedo

Case study 3 is meant to illustrate the potential of EcoBot and concurrent auxiliary measurements to study landscape-scale variability in R n and its components and drivers. Briefly, the study was conducted between June and October 2011 and June 2013 in the Stubai Valley (Western Austria), in the Matscher/Mazia Valley and in the Tauferer-Ahrntal Valley (both in Northern Italy), at 51 different grassland and shrub ecosystems. The study sites covered an altitudinal range from 850 to 2,500 m asl and included abandoned areas and differently managed hay meadows and pastures. At each site, two to five replicate EcoBot measurements were taken as described above. At the same sites, the above-ground plant area index (PAI) was estimated directly based on harvesting and plant area determination and/or indirectly based on canopy light transmission measurements using a line quantum sensor as described in Wohlfahrt et al. (2001). The total above-ground biomass was quantified by harvesting the vegetation in a 0.3 × 0.3 m area. Species composition and dominance were estimated in a 2 × 2 m area based on Braun-Blanquet (1964) and the vegetation association according to Tasser et al. (2010).

Soil temperature affects numerous soil processes (e.g. weathering, mineralisation of organic material) and through the biogeochemical cycling of carbon, nutrients and water, vegetation composition and status. Spatial differences in soil temperature on a landscape-scale reflect these differences in soil and vegetation, in addition to topographical and environmental factors. Figure 5 illustrates the potential of the EcoBot to explore and explain landscape-scale spatial patterns in soil temperature using a forward stepwise linear regression. We hypothesised that a combination of site, vegetation and land-use variables would best predict spatial soil temperature patterns. The following site variables were used: altitude (as proxy for the altitudinal climate gradient), slope angle and aspect, all parameters measured by the EcoBot (see Fig. 1), day length, time of day, total vegetation cover, cover of grasses, forbs, dwarf shrubs and open soil, mean canopy height, PAI and phytomass (total, green, woody and dead plant matter). With these independent variables, a total of 83.7 % of the spatial variability in soil temperature could be explained (Durbin Watson: 1.4), with 11 variables contributing significantly (Fig. 5b). Spatial patterns of soil temperature were positively correlated with air temperature, which explained the largest fraction of the total variability (Fig. 5a), the time of day and the degree of South exposition. Negative correlations were observed with variables expressing the amount and cover of above-ground plant area (total vegetation and dwarf shrub cover, PAI and green biomass), soil moisture, altitude, day length and the degree of North exposition. While this simple empirical model is likely to have little utility outside the conditions under which the data have been acquired, it may be used in the study areas, together with land cover/use maps and information on the litter quality of single different land cover types (e.g. Gamper et al. 2007), to predict soil mineralisation rates for carbon budgeting studies (e.g. Smith et al. 2005).

The potential of the EcoBot for exploring spatial differences in albedo driven by land use is shown in Fig. 6 for the same dataset. Abandoned areas with a high cover of dwarf shrubs or taller woody shrub species such as Pinus mugo or Alnus virridis reflected much less shortwave radiation compared with differently managed grasslands or Alpine grasslands above the tree line (Fig. 6). These findings are critical for accounting for biophysical feedbacks (Bonan 2008) from ongoing changes in land use (e.g. Tappeiner et al. 2008a, b) and climate (e.g. Pauli et al. 2007; Gobiet et al. 2014) in the Alps.

### Case study 4: using the EcoBot for inferring evapotranspiration

Another application, which primarily motivated the inclusion of air temperature and wind speed measurements, is to use the EcoBot for estimating sensible and latent heat fluxes. In a bulk approach, the sensible heat flux may be derived from the gradient (K) between the aerodynamic surface temperature (T aero) and air temperature (T air), divided by the aerodynamic resistance to heat transfer (r aero; s m−2) multiplied by the product of the density and specific heat of air (ρc p ; J m−3 K−1):

$$\mathit{\mathsf{H}}=\rho {\mathit{\mathsf{c}}}_{\mathit{\mathsf{p}}}\left({\mathit{\mathsf{T}}}_{\mathsf{aero}}-{\mathit{\mathsf{T}}}_{\mathsf{air}}\right)/{\mathit{\mathsf{r}}}_{\mathsf{aero}}$$
(4)

For near-neutral conditions, the aerodynamic resistance may be estimated on the basis of the logarithmic wind law using measured wind speed and estimates of the zero-plane displacement height (m) and roughness length (m), which may be derived from canopy height (Campbell and Norman 1998). If the aerodynamic surface temperature is replaced with the measured infrared surface temperature, Eq. 4 may be solved for H exclusively on the basis of EcoBot measurements. Replacing the heat storage (S) in Eq. 1 with the soil heat flux (G) and assuming G to represent some fraction of R n (Sauer and Horton 2005) and/or by empirically relating it to measured soil temperature and water content, H derived from Eq. 4 together with estimated G and measured R n allows inferring λE as the residual of the energy balance, i.e.

$$\lambda \mathit{\mathsf{E}}={\mathit{\mathsf{R}}}_{\mathsf{n}}-\mathit{\mathsf{H}}-\mathit{\mathsf{G}}$$
(5)

This approach was applied to the data presented in case study 1, and the results are shown in Fig. 7 for the four components of the energy balance equation. Note that, in contrast to the data shown in Fig. 3, here we present EcoBot measurements only from the grassland plot (#6 in Figs. 2 and 3) which is identical to where the eddy covariance flux tower is situated and which makes up a major fraction of the flux footprint. In order to enable comparison with the EcoBot calculations, which close the energy balance by definition (Eq. 5), eddy covariance sensible and latent heat fluxes were adjusted for the lack of energy balance closure (20 % residual energy on average) using three different approaches (Wohlfahrt et al. 2009): The first approach forces closure by assigning the residual energy to H and λE according to the Bowen-ratio and is used as the reference (solid lines in Fig. 7) below. The second approach assigns the entire residual energy to either H or λE, and the third approach uses H and λE as measured, i.e. applies no closure operation. The second and third approaches represent the range of possible closure operations and are highlighted in Fig. 7 by grey shading. The soil heat flux was estimated as 18 % of R n measured by the EcoBot, based on soil heat flux measurements at the eddy covariance flux tower (see also Hammerle et al. 2008). It can be seen that, despite a clear underestimation of H before noon, overall λE inferred from the Ecobot measurements nicely corresponded with the one measured by eddy covariance (λE EcoBot = 1.05 λE EC , r 2 = 0.79, RMSE = 40.3 W m−2) and was mostly within the range of the uncertainty of the eddy covariance λE measurements due to the lack of energy balance closure.