Granger and Gray (1989) developed the G–D model from the complementary relationship of the combination model of Bouchet (1963) and Penman (1948). The G–D model extends the potential evaporation model to non-saturated surfaces using the relative evaporation term, G. This is defined as a ratio of actual to potential evaporation which depends on the relative drying power of the air, D. The underlying theory is based on the reduction of water availability as a surface dries, but the “potential” evaporation increases due to a subsequent rise in surface temperature.

The development of G eliminated the need for observations of surface temperature and vapour pressure. As a result, estimates of actual evaporation were obtainable for non-saturated surfaces with atmospheric data alone (Granger, 1989). The general form of the equation is 1E=ΔGQ*-Qg+γGEAΔG+γ. The available energy term is driven by net radiation, Q* (W m-2), calculated as a sum of the net shortwave and longwave radiation components and the ground heat flux, Qg (W m-2), and the slope of the saturation vapour pressure curve, Δ.

The aerodynamic term includes the psychrometric constant, γ, and “drying power of the air”, EA, calculated using a Dalton-type formula. 2EA=f(u)ea*-ea, where f(u) is a vapour transfer function, and the atmospheric vapour pressure deficit at 2 m height is derived from ea* (saturated) and ea (actual). Pomeroy et al. (1997) empirically derived f(u) as a function of wind speed and aerodynamic roughness length and Z0 (m) from extensive field data collected for prairie, boreal forest, and northern cold-region environments in western Canada. 3f(u)=8.19+22Z0+1.16+8Z0u, where u is the mean daily wind speed (m s-1).

The G–D complementary feedback method is driven by the non-linear relationship between G and the relative drying power of the air, D. 4G=10.793+0.2e4.902D+0.006D, where D is a function of the humidity deficit and available energy is given by 5D=EAEA+Q*-Qgλ, and λ is the latent heat of vaporisation (kJ kg-1).

The feedback model approach is generally applicable for regions indicated above and can be used when detailed soil moisture information is lacking, except under conditions of severe moisture stress. Armstrong et al. (2008) evaluated point-scale evaporation estimates obtained with the G–D model over mixed grasses at the same study area under conditions of non-limiting soil moisture using field data collected during 2006. Estimates obtained with the G–D model compared well with eddy covariance (EC) measurements and was less data intensive than the Penman–Monteith model, which also performed well in that study.

Armstrong et al. (2010), however, found that the use of appropriate soil moisture constraints was required to obtain reliable estimates with the G–D model when applied under drought conditions to a mixed-prairie site at Lethbridge, Alberta, Canada.

Inputs needed to parameterise Eqs. (1)–(5) included mean daily net radiation, air temperature, humidity, wind speed, and surface roughness length. Outgoing radiation components were derived from high-resolution digital (Canon PowerShot A70 camera) and thermal (FLIR ThermaCAM P20 imaging radiometer) images taken on 5 August 2007 during a Cessna flight at midday. Meteorological and surface observations (radiation components, temperature, humidity, and wind speed) were recorded at two station locations. Figure 1 shows the relative locations, which included one reference site and one validation site. An independent station operated by the National Water Research Institute (Environment Canada, Saskatoon) was also available which provided a second validation site for estimates of daily net radiation. Air temperature, humidity, wind speed, incoming radiation, and surface temperature were recorded as 15 min averages. Incoming and outgoing shortwave and longwave radiation were measured with a CNR1 net radiometer (Kipp & Zonen, Delft, the Netherlands). Air temperature and humidity were measured at 2 m height using a shielded Vaisala HMP45C (Campbell Scientific, Inc., Logan, Utah, USA). A shielded Exergen infrared (IR) temperature sensor (IRTC; Exergen, Watertown, Massachusetts, USA) was used to measure surface temperature.

Canopy spectral reflectance was independently sampled on 21 August 2007 for validating albedo estimates derived from the visible images taken on 5 August. Canopy reflectance was collected according to the methods of Disney et al. (2004) and Zhang and Guo (2007). Samples were taken at 4.5 m intervals along a site transect at 1 m height at the nadir (25∘ field of view) with an ASD FR Pro spectroradiometer (Analytical Spectral Devices, Inc., Boulder, Colorado, USA); the spectral range was 350–2500 nm with a 1 nm resolution. Samples were taken between 12:00 LT (local time) and solar noon. Reflectance was recalibrated every 10 min using a white Spectralon reflectance panel (Labsphere Inc., North Sutton, New Hampshire, USA).

Eddy covariance observations were sampled at approximately 2 m height with a CSAT3 three-dimensional sonic anemometer (Campbell Scientific, Inc., Logan, Utah, USA) and a KH20 ultraviolet krypton hygrometer (Campbell Scientific, Inc., Logan, Utah, USA). The raw EC data were reported as 15 min averages and post-processing was done with MATLAB (Mathworks, Natick, Massachusetts, USA) which included flux corrections using a standard planar-fit axis rotation algorithm (Wilczak et al., 2001). Data filtering indicated there were no missing or bad data values on the study day, and station positioning was not considered to be an issue. EC data collected at the validation site were used for a comparison against a mean estimate obtained from the G–D model upwind of the EC station.

Given the higher frequency of flux reporting (15 min averages), axis rotation, and data filtering applied, and a potential mismatch in measurement footprints, no further corrections were used to force an energy balance closure on the 15 min averages.

At midday, reference observations of the shortwave, K↓ (835 W m-2), and longwave, L↓ (320 W m-2), irradiance and albedo (0.153 over grass) were obtained from the CNR1. Both K↓ and L↓ were assumed to be uniform over the field given the images used for calculations were cloud free. The reference values were used in the calculation of midday normalised ratiometric indexes for albedo and net radiation. A mean daily reference value of 155 W m-2 was obtained for Q* which was the input for distributing the final estimates of the mean daily net radiation using the ratiometric index (discussed in Sect. 3.4).

The daily ground heat flux measured at two locations was relatively small, 8 and 2 W m-2, due to the continuous and tall grass canopy cover. For this study the ground heat flux was ignored, in part due to the small measured values, but this was also to limit errors introduced from using a numerical solution to estimate ground heat flux, which would require more-detailed cover information at each image pixel. To standardise the parameterisation of the G–D model for calculating the mean daily estimates of evaporation, the reference values of mean daily air temperature (19.6 ∘C), humidity deficit (1.1 kPa), and wind speed (3 m s-1) were also taken to be uniform. Differences in mean daily air temperature between the measurement locations were approximately 1 ∘C.

Potential impacts of assuming uniform humidity and wind speed for parameterising the aerodynamic term was also considered. This was done by examining G–D evaporation estimates derived from 2006 field data collected over different surfaces at the same study area (Armstrong et al., 2008). During 2006 a fixed station obtained continuous measurements over a mixed-grass upland, and a portable EC station was moved periodically to obtain concurrent measurements over different land covers. Meteorological observations from the fixed station were initially used to estimate evaporation with the G–D model. Another estimate was obtained after substituting observations of humidity and wind speed taken from the portable EC station. This resulted in only minor variations in the respective evaporation estimates (root mean square deviation (RMSD) of 0.02 mm) derived from the concurrent measurements of humidity and wind speed taken from the different sites.

Implications of neglecting the ground heat flux and treatment of the meteorological parameters (above) with respect to modelling uncertainty are given in the discussion.

Exploratory analysis of key surface variables and evaporation estimates was conducted using the R software environment (Grunsky, 2002). Data analysis consisted of boxplot summaries giving the 25th and 75th quartiles (i.e. the interquartile range) defined by the box, and an inner solid line indicating the median value was also given. Data points were plotted with jitter to reflect the density of outliers greater than 1.5 times larger or smaller than the interquartile range with the upper and lower limits defined by box whiskers.

At midday on 5 August 2007 a reference albedo of αref=0.153 was obtained from observed irradiance and the reflectance of shortwave radiation over the mix of green grasses. The albedo map resulting from Eq. (11) and locations of reference and validation points is shown in Fig. 3. Vegetation was similar at both EC station locations, and the scaled albedo estimate (0.164) agreed well with the measured value (0.167) at the validation site.

The validation of the mean and range of albedo estimates obtained for major land covers in Fig. 3 is summarised in Table 1. Estimates of albedo from the image compared well with values expected for grasses, agricultural crops, deciduous trees, and bare soils reported in Brutsaert (1982). The root mean square error (RMSE) of albedo estimates from the pixels and measured albedo values was approximately 3.5 %, which is within an expected error of 2 % to 5 % for research purposes (Liang, 2000).

The 5 m resolution albedo map highlighted surface information within the landscape that would be more generalised using coarser data. For example, the albedo map depicted distinct boundaries separating regions dominated by brome (BG) and mixed grasses (MG), cultivated or crop area (C), sparsely vegetated, fallow area (F), and wetland fringe vegetation (W). Also, wetland extents and fringe vegetation were observable in areas surrounded by other vegetation types (see Fig. 1 for reference). The detailed variations are expected to impact evaporation estimates through relative increases at pixels with relatively lower values of albedo, which results in an increase in the available energy, whereas relatively higher values will reduce the available energy and estimates of evaporation.

Figure 4 shows the resulting map of emitted longwave radiation, L↑, derived from the image of observed surface temperatures, with estimates ranging between 380 to 480 W m-2. At the two EC station locations a comparison was made among the midday surface longwave measurements from the P20 camera, a narrow-beam Exergen infrared thermocouple (IRTC) radiometer, and surface-emitted longwave values obtained from the CNR1.

The P20 measurement compared well with the IRTC values with differences less than -12 W m-2. Compared with the CNR1 values, the differences were slightly larger at -30 W m-2 but was still within 8 % error. Generally, the differences were small considering relative changes in midday surface net radiation, and the magnitude of incoming radiation components is considerably larger. The variable footprints of the different measurements and absorption properties of water vapour, which can reduce the signal, are likely sources of error. Also, dust and heating of the CNR1 downward-facing pyrgeometer may introduce another source of error.

For estimating evaporation, the smaller emitted longwave values resulting from lower surface temperatures will increase the available energy relative to larger values which imply reduced water availability and evaporative cooling.

Figure 5 shows the resulting classified surface roughness length, Z0, map. A visual comparison with the study area map (Fig. 1) and albedo map (Fig. 3) suggests that the classification map provides a physically realistic representation of variations in roughness over a large portion of the study area. Notable changes in the reflected surface properties and roughness length classification among the key land cover types can be observed. This is particularly true where distinct boundaries separate the transition between the dominant cover types such as where taller shrubs and trees surround ponds, compared to more broad regions of the fallow, cropped areas and grasses.

Normalised indexes for albedo, surface temperature, and roughness were considered to examine the relationships of variations in these variables compared to the evaporation ratio. Figure 6 shows the expected physical behaviour of these variables within the G–D model. Only the actual range of values computed for the ratiometric index was considered so that the physical variations could be shown more clearly. In this case, the impacts of relative changes in the apparent inverse linear relationships between ER and αR and L↑R and slight non-linear relationship between ER and Z0R on evaporation estimates can be computed. The results can be used to illustrate the impacts of required changes in key driving variables to produce notable changes in the evaporation ratio.

For example, a relative increase in the surface temperature via L↑R of 0.18 (or 18 %) was shown to reduce ER by 10 %. By comparison, an increase in reflected shortwave radiation via αR of 0.30 (or 30 %) reduced ER by only 5 %. In the case of surface roughness, an increase of 250 % was needed to reduce ER by 10 %. Consequently, a relative reduction in evaporation rates may be expected where albedo, surface temperature, and surface roughness tend to be larger. In the latter case, step changes in surface roughness would increase the relative drying power, D, but the relative evaporation, G, results in a relative decrease due to the inverse non-linear relationship. The increased sensitivity of ER to L↑R also indicates that detailed spatial variations in surface longwave radiation is an important factor for estimating evaporation.

Net radiation is known to vary dynamically on a sub-daily basis. Equation (10) shows how a radiation ratio, QR*, can be used as a temporal transfer function to scale estimates of mean daily net radiation over an image. In order to scale the normalised net radiation ratios from a temporal “point” at midday to a mean daily value, it was necessary to examine whether a stable proportionality existed between measured values at midday and mean daily net radiation. Verification of a proportionality under clear skies would eliminate the need for an empirical scaling function.

Historical records were examined for a period from 1 May to 1 September at three Canadian Prairie locations at similar latitudes (49–52.2∘). The analysis included two field seasons at the SDNWA study site (2006–2007). Archived data were also obtained at two short grass prairie locations, an AmeriFlux network site at Lethbridge, Alberta, Canada (1999–2004), and Kernen Farm located at Saskatoon, Saskatchewan, Canada (1999–2000). Figure 7 shows a moderately strong relationship at each location even with the y intercept set to 0, and r2 is equal to 0.54–0.62.

This result confirms that a general proportionality exists and suggests that daily net radiation, Qd*, could be scaled to all pixels across an image from a single measured reference value, Qdref*, as a function of the normalised ratio index for QR* at midday, stated here as 12Qd*=Qdref*QR*. Qdref* was assigned a daily measured reference value of 155 W m-2, which was obtained from the CNR1 at the reference station. QR* was derived from Eq. (10) using the measured reference values of incoming shortwave and longwave radiation components, and albedo and emitted longwave maps are used as inputs. The resulting mean daily net radiation map and a comparison of estimated and measured Qd* at two validation locations are shown in Fig. 8. The main validation site was equipped with a CNR1 and a second independent site maintained by Environment Canada was equipped with an NR Lite radiometer (Kipp & Zonen, Delft, the Netherlands).

The distributed estimates of daily net radiation ranged from approximately 120 to 190 W m-2. The Qd* estimates at the validation sites were 6 W m-2 higher (4 % error) compared to the CNR1 measurement and 11 W m-2 lower (8 % error) compared to the NR Lite observation. These results indicate that accurate estimates of net radiation can be scaled across detailed midday images from a single measured value of mean daily net radiation. More importantly, detailed accurate estimates of net radiation would be valuable for improving point-scale evaporation estimates.

The mean daily net radiation map derived from Eq. (12) was supplied as input to the G–D model to estimate the mean daily actual evaporation at every land surface pixel. The reference humidity and wind speed values (discussed earlier) and the surface roughness length map were used for calculating the aerodynamic terms in Eqs. (2) and (3). An estimate of mean daily evaporation was then calculated at every pixel location based on Eq. (1).

The resulting map of distributed actual evaporation estimates is shown in Fig. 9. Due to the level of detail captured in the net radiation map of Fig. 8, the map shows a physically realistic pattern of evaporation for the range of land cover types. For example, areas where vegetation was less dense (e.g. fallowed/cropped) and soil surface conditions were drier showed lower rates of mean daily evaporation, and higher rates were associated with the more-densely grassed areas. The highest evaporation estimates were obtained where water availability is expected to be higher, e.g. among wetland fringes and depressions where albedo and surface temperatures were lower, which can be attributed to increased water availability and evaporative cooling.

The prevailing wind direction for the day (Udir) was from a north-northwest direction. A mean evaporation estimate of 2.7 mm was obtained along a linear transect for pixels containing a brome grass surface immediately upwind of the EC station (Fig. 9). The estimate of 2.7 mm was 0.5 mm or 23 % higher than the EC-measured mean evaporative flux of 2.2 mm d-1. Based on the cumulative flux model of Schuepp et al. (1990), it is expected that 80 % of the cumulative flux would come from within an upwind distance of 100 m from the EC station along a similar path to the linear transect. The estimate of 2.7 mm d-1 might be lower if ground heat flux was found to be a factor for reducing the available energy. However this would need to be reliably accounted for at every 5 m pixel.

On the study day the mean daily energy fluxes in W m-2 (including latent (LE) and sensible heat (HE)), and the ratio of the energy balance closure at the validation site was (LE + HE)/(Q*-Qg)=(63+55)/(144-2). This results in an energy balance closure of approximately 83 %. The resulting Bowen ratio of 0.87 was reasonable for the drying conditions and later timing in the growing season for the grasses in this semi-arid region. In this case it is expected there will be uncertainty in both estimated and measured EC fluxes due to modelling and measurement errors and different footprint scales. For instance, it is possible that LE and HE could be under-measured or the ground heat flux could be more variable upwind of the EC station than estimated.

Measured evaporation rates were not available for trees (dominated by aspen) in 2007, but the G–D estimates were compared against archived values from Boreal Ecosystem–Atmosphere Study (BOREAS) data for an old-aspen site from August 1996. The G–D mean daily values in the order of 3 mm were reasonable compared to evaporation from similar trees reported by Hogg et al. (2001).

In general, the results are instructive because the G–D method has been demonstrated to provide a reasonable estimate of evaporation from the remote-sensing images taken over a complex landscape. Armstrong et al. (2008) have also shown similar accuracy for daily and multi-day periods from meteorological data alone during the 2006 field season at the same study area. More importantly, the ratiometric-index method of scaling may be valid for distributing surface radiation components across remote-sensing imagery obtained from satellites, planes, or near-surface aerial platforms such as UAVs.

The following sections briefly discuss the statistical distributions of the driving variables obtained from the images used for estimating evaporation and their impacts on the resulting evaporation estimates. The benefit here is the statistical distributions of evaporation and key driving factors are considered to be physically meaningful.

Figure 10 shows the frequency distributions of evaporation estimates and relative contributions for the energy balance and aerodynamic components. G–D model evaporation estimates appear normally distributed with a mean of 2.8 mm d-1 and a relatively low coefficient of variation (cv of 0.07). Distributions for the energy and aerodynamic terms were notably different due to interactions among the driving variables and different roughness classes. For the energy component, the distribution was continuous with a mean of 1.1 mm and a cv of 0.11. The distribution mean for the aerodynamic term was 1.7 mm and was bimodal due to the differences in magnitude of the discrete roughness classes used.

Figure 11 shows boxplot summaries of distributions for albedo (α), surface temperature (Ts), net radiation (Q*), relative evaporation (G), mean daily evaporation (E), and G plotted against net radiation. In the case of α, values below approximately 0.10 and lower “outliers” can be attributed to uncertainty associated with the wetland vegetation and likely where surface water was not completely masked out near the edges of ponds. In general, the distributions for α and Ts which drive the net radiation appear to show more notable variability compared to Q* and the resulting estimates of E. For example, the distributions of α and Ts showed opposite skewness, -0.83 and 0.36, respectively. The variability for α (cv of 0.19) was larger than Ts (cv of 0.14), which is also much larger than apparent variability of both Q* (cv of 0.65) and E (cv of 0.66). Mean values of the distributions for α, Ts, Q*, and E were close to the median values shown within the boxes depicting the interquartile ranges.

The resulting boxplot for G and the plot of G against net radiation reflects the interaction across the discrete roughness length classes. The statistical distribution of G was not continuous due to the larger step change in roughness length from 10 to 40 cm. Plots of G against net radiation resulted in three distinct linear relationships which can be attributed to using a uniform mean daily wind speed and humidity deficit for calculating the drying power of the air, EA. In this case, the potential variability of G associated with the relative drying power, D, is limited to variations in roughness length, but this may vary more with changes in wind speed and the humidity deficit.

Generally, despite the larger variability of key factors driving the energy component the complex interactions within the G–D model appear to reduce the overall variability of evaporation estimates.

There is a potential for areal estimates of evaporation to vary depending on how upscaled estimates are calculated from the underlying driving factors or the smaller point-scale estimates. For example, a mean areal estimate of evaporation calculated from all image pixels was 2.8 mm d-1, which accounts for all of the variability available from 5 m pixels. An areal estimate may also be calculated as a weighted mean of evaporation estimates obtained for each roughness length class. This is similar to the mosaic approach used within land surface schemes based on fractional land cover areas. The mean daily evaporation rates for the 5, 10, and 40 cm roughness length classes increased with each step change in roughness class as 2.6, 2.8, and 3.0 mm d-1, respectively. The distribution of land area associated with each roughness class was 48 % (10 cm roughness), 30 % (5 cm roughness), and 22 % (40 cm roughness). The weighted areal evaporation (Eareal) can be obtained as 13Eareal=(0.30⋅2.6)+(0.48⋅2.8)+(0.22⋅3.0)=2.78mmd-1. In this case there was only a small difference in areal estimates obtained based on the distribution mean or a weighted mean based on fractional areas of each roughness class. This may be partly due to the relatively small variation in evaporation rates across the roughness length classes and the level of detailed variability captured by the 5 m pixel resolution. Eareal was recalculated using Eq. (13) with different combinations of the fractional areas, which only produced a minor difference of ±0.1 mm. In other words, in order for there to be a larger difference between the areal estimates, greater variability may either be required in the evaporation estimates distributed over the field or among the mean rates for each roughness class.

The mean areal estimate, Eareal of 2.8 mm d-1 for all pixels, was also 0.3 mm higher than a G–D model estimate of 2.5 mm d-1 obtained from station meteorological observations of net radiation, air temperature, humidity, and wind speed. Variations among the estimates and measurements vary by approximately 22 % to 27 %, which is not surprising given the differences in calculation techniques and potential mismatches in associated footprint scales.

An upscaled areal mean estimate of evaporation can also be obtained from mean values of the key factors driving the energy and aerodynamic terms. The general form of Eq. (1) can be restated to derive the individual components more directly as 14E=ΔGQ*ΔG+γ+γGEAΔG+γ. For the entire area and also each roughness length class, mean values of the driving factors were derived and evaporation estimates were recalculated using Eq. (14). The relative evaporative contributions attributed to the energy and aerodynamic terms are provided in Table 2. For the different roughness length classes, the range of evaporation estimates attributed to E_energy was only 0.2 mm d-1 and nearly 0.7 mm d-1 for E_aero, and the difference in total evaporation, E_total was 0.5 mm d-1. A bias toward larger evaporation estimates might be expected given the increase in energy availability and enhanced turbulence with an increase in roughness. However, the potential bias was offset by the interaction of G and Q* and also G and EA.

Table 2 also compares evaporation estimates calculated from only mean values of G and Q*, and EA to the “expected” rates for each roughness length class. Expected rates were calculated from the mean values for all pixels assigned within each roughness length class. Evaporation rates derived from the mean input values alone were only between 0.14 and 0.2 mm less than the expected mean rates for each roughness length class. Upscaling the driving factors to the entire area also had no impact on the resulting estimate as the difference was just 0.1 mm.

Whether evaporation estimates might be influenced by a spatial covariance between driving factors was also examined. The Pearson correlation coefficient, r, was used to evaluate correlations among the driving factors distributed over the field area. By definition, Pearson's correlation is the ratio of the covariance (the numerator) between two variables normalised by the product of their standard deviations as follows: 15r=∑Xi-X‾Yi-Y‾n∑Xi-X‾2n∑Yi-Y‾2n, where Xi and Yi are the respective values of the variables, the overbar denotes the mean value, and n is the number of pairs. A strong correlation between two variables might suggest the existence of covariance that could influence upscaled estimates of evaporation. Given that the roughness classes used represent discrete data and there is a lack of more detailed meteorological data to parameterise the aerodynamic term, further evaluation related to climate factors would not be meaningful. Therefore only an examination of the factors driving the energy term are considered here.

In this case, a potential covariance between G (dimensionless) and Q* (expressed in mm d-1) can be considered directly. By rearranging Eq. (15) the covariance can be obtained by multiplying the correlation coefficient and the product of the standard deviations of Q* (0.34 mm d-1) and G (0.021 mm d-1). The correlation between Q* and G over the field area produced a coefficient, r=-0.67.

When multiplied in series (r=-0.67)⋅0.34⋅0.02 this results in a covariance of approximately -0.0049 mm d-1. This suggests interactions between Q* and G for the methods applied here would have no further statistical influence on upscaled evaporation estimates. Unfortunately, no comment can be made regarding covariance between G and the turbulent flux component in the G–D model. Such analysis would require more detailed observations of air temperature, humidity, and wind speed and possibly a more sensitive combination model. Nevertheless, the impact of combined interactions within the G–D model effectively reduced the overall variability of point-scale evaporation estimates and also upscaled estimates derived from different computation methods for parameterising the model.

Implications of general modelling assumptions and the uncertainty of the methods applied are briefly discussed here. The daily ground heat flux was considered to be negligible for the study day as the mean daily flux was relatively small at the two stations where there was good canopy coverage. Where local cover properties are similar (e.g. brome grass and mixed-grass areas) the ground heat flux may be of a similar magnitude, but it would be greater where surface cover is less dense or where water availability is limited, due to differences in the surface properties.

Further, the relative evaporation, G, which is based on the relative drying power, D, integrates available energy which in this case directly includes the spatial variability of surface temperature. As such, evaporation might be overestimated in areas with more ground exposure and higher surface temperatures, where ground heat flux may be appreciably larger (e.g. fallowed/crop area). This interaction could result in a further reduction in the energy available for the evaporation estimate.

C3-type vegetation was at the study area, but the model behaviour may differ for C4 plants and may require new vapour transfer equations for surface types and plants other than which the G–D model was developed on. In the current study representative roughness lengths were selected based on reported values for surfaces of a similar type and height. A larger uncertainty in evaporation estimates may be expected for the roughness length class associated with the shrub/tree areas which could result in variable changes in wind speed and enhanced turbulence associated with larger roughness lengths.