1 Introduction

The economy of a country depends on accurate and high resolution short-, medium- and long-range weather forecasts (Kumar et al. 2013a), used in agriculture, water resources, aviation, and disaster mitigation, on a daily basis (Strand 2000). Continuous verification of numerical weather prediction (NWP) models with observations from ground-based instruments is obligatory in evaluating the quality of model forecasts, and to provide feedbacks on improving the performance of numerical models. We have performed a number of studies (Kumar et al. 2011, 2013a, 2014a, b) on the verification of NWP model-predicted meteorological parameters (e.g. surface temperature, moisture, wind speed, and rainfall) with conventional observations. Quite a few earlier studies estimate errors in the predicted surface fluxes (e.g. Oncley and Dudhia 1995; Zamora et al. 2003; Cheng and Steenburgh 2005; Prabha and Hoogenboom 2010; Kumar et al. 2013a). Marciotto (2013) discussed the relation between the radiation budget and the ground heat flux, the energy stored in the soil, and the residual of the energy balance in urban and suburban areas of Oklahoma City, USA, together with the variability of the energy fluxes across the urban area. The evaluation over land surfaces of the flux parametrization in atmospheric models has been carried out by Beljaars and Holstag (1991), inter alia, while Mahura et al. (2009) investigated the effects of urban schemes in the HIRLAM numerical weather forecasting system. The surface energy fluxes obtained from the Weather Research and Forecasting (WRF) model simulations were evaluated by Jaksa (2012) for the Snake basin, while recent studies (Baker et al. 2013; Kleczek et al. 2014) have evaluated surface and upper-air parameters from WRF fine-scale model simulations. However, no dedicated study of the diurnal evaluation of NWP high-resolution fluxes have been made over diverse agro-climatic settings such as in India using micrometeorological measurements for different vegetative systems.

A network of 22 INSAT (Indian National SATellite System)-linked micrometeorological towers has been established across the country on the initiatives of the Indian Space Research Organization (ISRO) (Bhattacharya et al. 2009). Observations from micrometeorological towers provide a unique opportunity to validate model-predicted meteorological parameters as well as surface fluxes from the same platform. These tower observations may help improve the quality of the NWP model forecasts, and the description of land-surface characteristics in NWP models. The objective therefore is to validate several important surface parameters available from WRF model simulations against tower observations.

2 Micrometeorological Tower Data

A network of 10-m tall micrometeorological towers was installed under a national project titled “Energy and Mass Exchange in Vegetative Systems” under the aegis of the ISRO Geosphere-Biosphere Programme. These towers have instruments that measure the surface temperature, humidity, and wind speed at 30-min intervals at three vertical levels, soil temperature and soil moisture at three depths, two-depth soil heat fluxes, and shortwave and longwave radiation fluxes on a 30-min basis over different land-cover types across 15 broad agro-climatic zones (Bhattacharya et al. 2009, 2013b). The short vegetation surfaces include crops, grasses, natural shrubs, wetland vegetation, and young forest where the mean maximum canopy height varies from 1 to 2.5 m. Twenty-two micrometeorological towers were installed in India over 15 agro-climatic regions within the footprint of INSAT, each one with an optimal fetch of 500–1000 m, maintaining a fetch ratio of 1:50–1:100. These towers are spread over alluvial, black, red, grey-brown, and desert soils distributed across humid to arid climates. Soil temperature and moisture measurements are available at 0.05-, 0.1- and 0.2-m depths. The tower observations from 22 locations were collected through the data relay transponder on board the ISRO’s INSAT 3A satellite, with Yagi antenna attached to each tower transmitting observations of 26 data fields through four signal packets. Micrometeorological tower observations are archived in the Meteorological and Oceanographical Satellite Data Archrival Centre (www.mosdac.gov.in) maintained by the Space Applications Centre of ISRO; the spatial distribution of ISRO micrometeorological towers is shown in Fig. 1. The prevailing cover types at tower locations and land-use/land-cover (LULC) types defined in the WRF model collocated to tower locations are summarized in Table 1. The match is quite good where the tower fetch footprint (500–1000 m) has similar surrounding extended land-cover types within the WRF model spatial resolution (5000 m). But mismatches arise at those tower locations where existing land-cover types in the tower fetch footprint are surrounded by different cover types within the WRF model grid.

Fig. 1
figure 1

Distribution of micrometeorological towers over the Indian region

Full size image
Table 1 Wide range of climate, soil, cover types represented by micrometeorological tower and corresponding land-use/land-cover used in WRF model
Full size table

The sensible and latent heat fluxes are generally estimated from the tower measurements either using the Bowen ratio energy balance approach assuming energy balance closure, or using the flux-profile method with aerodynamic assumptions. Both approaches have their inherent uncertainties thereby leading to errors in energy balance closure to \(\pm \)30 % (Bhattacharya et al. 2013a; Singh et al. 2014). Direct and improved measurements of the sensible and latent heat fluxes depend on the eddy-covariance approach, but in India, no systematic network of eddy-covariance sites exists.

3 Methodology and Model Description

3.1 Model Description

The Weather Research and Forecasting (Skamarock et al. 2008) model version 3.5 used herein is a limited area, non-hydrostatic, primitive equation model with multiple options for various physical parametrization schemes. This version employs Arakawa C-grid staggering for the horizontal grid and a fully compressible system of equations; a terrain-following hydrostatic pressure coordinates with vertical grid stretching is implemented vertically. The time-split integration uses a third-order Runge-Kutta scheme (Wicker and Skamarock 2002) with a smaller timestep for acoustic and gravity wave modes. The WRF model physical options used consist of the single moment 6-class simple ice scheme for microphysics (Lin et al. 1983); the Kain–Fritsch scheme (Kain 2004) for the cumulus convection parametrization, and the Yonsei University planetary boundary-layer scheme ( Hong and Dudhia 2003). The rapid radiative transfer model (RRTM; Mlawer et al. 1997) and the Dudhia scheme (Dudhia 1989) are used for longwave and shortwave radiation, respectively. The Noah land-surface model (Chen and Dudhia 2001) is selected in this version of the WRF model, and contains four soil moisture layers, namely 0–0.1, 0.1–0.4, 0.4–1.0, and 1.0–2.0 m below the surface. The thicknesses of the lowest atmospheric layers are 75.8, 103.7, 132.4, 166.8, 212.3 and 260.2 m. Table 2 summarizes the various schemes used in the WRF model version.

Table 2 Summary of the selected WRF model schemes
Full size table

The WRF model and its three-dimensional variational (3D-Var) data assimilation system (Barker et al. 2004) are configured and implemented to provide high spatial resolution (5000 m) short-range weather forecasts over India. WRF 3D-Var aims at producing an optimal estimate of the true atmospheric conditions by minimizing the cost function defined by

$$\begin{aligned} J(x)=(x-x^\mathrm{b})^\mathrm{T}B^{-1}(x-x^\mathrm{b})+(y-H(x))^\mathrm{T}(E+F)^{-1}(y-H(x)), \end{aligned}$$
(1)

where x is the analysis state composed of atmospheric and surface variables and \(x^\mathrm{b}\) is the first guess comprising values from the previous forecast. Conventional and satellite observations are represented by y, B and E represent the background and observation error covariance matrices respectively, and the observation operator H transforms the analysis to observation space \(y=H(x)\). The inaccuracies introduced on account of the observation operator are contained in the representative error covariance matrix F. A detailed description of the 3D-Var system can be found in Barker et al. (2004).

3.2 Design of Experiment

The WRF model is integrated in a triple nested domain configuration (Fig. 2) with two-way nesting and a horizontal resolution of 45, 15 and 5 km over India with \(260\times 235\), \(352\times 373\) and \(676\times 721\) grid points in the x (longitude; east to west) and y (latitude; north to south) directions for domains 1, 2 and 3 respectively. The model has 36 vertical levels with the top of the model located at 10 hPa. The WRF model forecast is generated for the next 72 h initialized from 0000 UTC (0530 IST local standard time), with data available daily from satellites as well as conventional observations assimilated at the same time (here 0000 UTC) using 3D-Var data assimilation to improve the initial conditions of the WRF model (Kumar et al. 2011, 2014a; Singh et al. 2011). ISRO automatic weather stations, Oceansat-2 scatterometer (Kumar et al. 2013b) and advanced scatterometer winds, satellite retrieved temperature and moisture profiles, satellite radiances, refractivity, and conventional and satellite observations available from EUMETCast (European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT)’s Multicast Distribution System) and National Centers for Environmental Prediction (NCEP) are used to improve the model initial conditions. The NCEP Global forecasting system (analysis and 6-h forecasts) is used to generate the initial and lateral boundary conditions of the WRF model. Hence, 3-hourly WRF model forecasts of air and soil temperatures, relative humidity, wind speed and atmospheric pressure at 5000-m spatial resolution throughout the year of 2011 (365 sample days) over the Indian region are compared against micrometeorological tower observations. The WRF model radiation and ground heat fluxes are compared with tower fluxes for a limited period November–December 2011 due to the availability of well-calibrated measured fluxes from the towers. The average measurement fetch for atmospheric sensors such as air temperature, relative humidity, wind speed, atmospheric pressure, incoming shortwave and longwave radiation at the micrometeorological towers is about 500 m. However, it is still less (about 50 m) for soil sensors such as soil temperature and ground heat flux. The tower data closest to the relevant WRF model grid point (here 5000 m) are directly compared without any interpolation.

Fig. 2
figure 2

The nested domain used in WRF model simulations

Full size image

4 Results and Discussion

In this study, mean error (BIAS) and root-mean-square difference (RMSD) are considered as the standard statistical measures to evaluate the quality of WRF model forecasts during the year 2011. WRF model analyses and 3-hourly forecasts of surface (2-m air) temperature and relative humidity, 10-m wind speed and direction, soil temperature, surface pressure, downward shortwave (\(S^{\downarrow }\)) and longwave (\(L^{\downarrow }\)) radiative fluxes at the surface, and ground heat flux \((G_\mathrm{S})\) are compared with micrometeorological tower observations. The above three fluxes are direct measured quantities at the micrometeorological towers.

4.1 Temperature

BIAS and RMSD values for the WRF-model simulated surface temperature analyses and forecasts are compared with tower observations as shown in Fig. 3. The left (primary) y-axis shows the RMSD and the right (secondary) y-axis shows the BIAS statistics. Since these tower observations are not used in the data assimilation process, the verification can be considered as an independent validation. A gross quality control check is performed for validation, such that if the difference in the model forecast values and tower observations valid at the same time is more than 20 K, the corresponding points are discarded. This quality control is applied to avoid unrealistic cases, which may be due to signal interruption while transmitting data from the INSAT-3A data relay transponder to the ground receiving station. Approximately, 9 % of temperature data are discarded after this check. The RMSD and BIAS values of 2.6 and 0.5 K, respectively, are found in the WRF model analysis, which shows that the analysis itself has an error in the range of 2–3 K. The RMSD values remain below 4.2 K (Fig. 3) throughout the 72-h model run, but RMSD increases from \(\approx \)2.6 to \(\approx \)4.2 K as the forecast length increases. The RMSD values in temperature forecasts show a semi-diurnal cycle that has a local minimum in the 12-h forecasts (e.g. 12-h, 24-h etc.) compared to neighbouring forecast lengths. A positive BIAS value occurred at all forecast times, showing that the WRF model overestimated temperature when compared with tower observations. Additionally, monthly error statistics of surface temperature forecasts (Table 3) demonstrate that a slightly smaller RMSD value (\(<\)3.0 K) is observed during the Indian summer monsoon (June–September) period while larger RMSD value (\(>\)3.3 K) is found in the pre-monsoon periods in 24-h forecasts. During July and September 2011, minimum RMSD values (2 K) occur in the 24-h surface temperature forecast. Slightly larger RMSD and BIAS values are observed in the 48-h surface temperature forecast compared to the 24-h temperature forecast. The elevation differences in the WRF model grids and corresponding micrometeorological stations could also contribute to errors in the surface air temperature. Generally, elevation differences between micrometeorological stations and the corresponding WRF grid point lie within 4–35 m, which corresponds to approximately 0.02 to 0.2 K error in temperature. For the Dehradun station, having a 202-m elevation difference, this error is approximately 1.3 K.

Fig. 3
figure 3

Temporal distribution of RMSD and BIAS values in 3-h surface temperature forecasts during the year 2011 (total 365 sample days)

Full size image
Table 3 Monthly statistics of 24-h WRF model forecast when compared with micrometeorological tower observations
Full size table

4.2 Relative Humidity

The temporal distribution of BIAS and RMSD values of surface relative humidity (RH) analysis and forecasts are shown in Fig. 4. As with surface temperature, a gross quality control check is used for RH verification. If the difference between model-simulated RH and tower-observed RH values is more than 50 %, the corresponding points are discarded during error computation. The unrealistic values of data fields could be due to the perturbation of satellite transmission of any of the four signal packets that carry a total of 26 data fields of tower measurements to the data relay transponder. Approximately 30 % of data are discarded after this quality control. The model initial condition has RMSD and BIAS values of 18 and \(-\)9 %, respectively; RMSD varies from \(\approx \)16 % (in 9-h forecast) to \(\approx \)21 % (in 45-h forecast) as the forecast length is increased but remains below 22 % throughout the 72-h forecast. The RH RMSD values show a diurnal cycle with a minimum around 0900 UTC and an increase in the next 9 h (to 1800 UTC). The maximum variation occurred between 1200 and 1500 UTC. Similarly, BIAS values follow a diurnal cycle with the maximum difference between 1200 UTC and 1500 UTC and minimum BIAS values at 0900 UTC. A low BIAS value is found for all forecast lengths, which shows that the WRF model underestimates the tower-observed relative humidity. Higher RH values are predicted during July, August and September 2011 (figure not shown), which matches well with tower observations. Higher RMSD values (\(>\)20 %) are found during the pre- and post-summer monsoon periods (Table 3), while the Indian summer monsoon period (June–September) shows smallest RMSD values (\(<\)18 %) at 24-h.

Fig. 4
figure 4

Temporal distribution of RMSD (circle) and BIAS (triangle) in 3-hourly surface relative humidity (RH) forecast

Full size image

4.3 Wind Speed

The prediction of surface wind is one of the most important but least accurate model derivatives over land. As with the forecast verification of surface temperature and relative humidity, a gross quality control is also applied for wind-speed forecast validation (Fig. 5). If the difference between two different wind speeds at any coincident time is \(<\)30 m s\(^{-1}\), the corresponding points (\(\approx \)40 % points are discarded) are considered for error calculation. RMSD values of \(<\)3.9 m s\(^{-1}\) are found at all forecast lengths based on one year’s data, while BIAS values in the wind-speed forecast lie between 1 and 3 m s\(^{-1}\). A slight increase is observed in 48-h and 72-h forecasts compared to 24-h forecasts. The maximum BIAS and RMSD values occur at around 1800 UTC and the minimum RMSD values occur at 0600 UTC. The WRF model overestimates the tower-observed wind speed in all months. RMSD values of approximately 4 m s\(^{-1}\) are found in 24-h wind-speed forecasts in June and July (Table 3), with minimum RMSD values found in October.

Fig. 5
figure 5

Temporal distribution of RMSD (circle) and BIAS (triangle) values in 3-h surface (10-m) wind-speed forecast

Full size image

4.4 Surface Pressure

RMSD and BIAS values for the surface pressure analysis and forecasts are shown in Fig. 6a. RMSD values of \(<\)11 hPa are found in the surface pressure analysis and subsequent forecasts. RMSD values decrease by approximately 5 % around 1200 UTC but higher RMSD values are found for all forecast lengths. BIAS values of less than 2 hPa are observed at all the forecast lengths. RMSD values of approximately 11 hPa are found in 24-h surface pressure forecasts for all months (Table 3). WRF model grids (5000 m) represent the average elevation using 1000-m elevation data from the United States Geological Survey, whereas, tower observations represent the point location. Errors in WRF model surface pressure forecasts may thus partly be due to these elevation mismatches. For 4–35 m deviation in elevation, surface pressure errors are approximately 0.5–5 hPa. One station located at Dehradun (the high topographic region) shows a large difference between the elevations (202 m), which represents a deviation of approximately 25 hPa in surface pressure.

Fig. 6
figure 6

Temporal distribution of RMSD (circle) and BIAS (triangle) in 3-h a surface pressure and b soil temperature forecast

Full size image

4.5 Soil Temperature

The RMSD and BIAS values for soil temperature analyses and forecasts are shown in Fig. 6b. A slightly higher error occurs in soil temperature forecasts compared to surface temperature forecasts. The WRF model simulation is obtained for the soil temperature within the zero to 0.1 m depths beneath the soil surface, which compares with tower-observed soil temperature within zero to 0.05 m. In this soil layer, the vertical temperature gradient is greater during the day than at night and varies according to soil type, soil moisture status and cover types. During the noon to afternoon hours in cloudless skies this gradient generally remains higher than the vertical gradient in air temperature. In addition, the measurement footprint of soil sensors at towers is much less than that of atmospheric sensors. The above factors may attribute to greater errors in soil temperature than in surface air temperature. Soil temperature forecasts have maximum and minimum errors at 0900 UTC and 0000 UTC, respectively. Negligible shifts are seen in 48-h and 72-h forecasts compared to the 24-h forecast. The impact of soil moisture errors on the diurnal simulation of the 0.05-m soil temperature, from a simple one-layer slab model, were evaluated for model initialization with observed soil temperatures at Watonga in the Okhlahoma mesonet site (Godfrey and Stensrud 2008). It was reported that negative soil moisture BIAS values alone accounted for more than 1.6 K increase (or decrease) in the maximum (or the minimum) soil temperatures depending on the time of initialization. Positive soil moistureBIAS values accounted for a more modest reduction of almost 0.9 K in the amplitude of the diurnal soil temperature cycle. While underestimates of soil moisture may contribute to the sign of the soil temperature errors, soil moisture alone apparently cannot account for the magnitude of the errors in soil temperature.

4.6 Radiation Fluxes

4.6.1 Downward Shortwave

The scatter plot of the WRF model predicted downward shortwave radiative fluxes \((S^{\downarrow })\) and micrometeorological tower-observed incoming shortwave radiative fluxes at 0300, 0600, 0900 and 1200 UTC shows large (Fig. 7) positive BIAS (88 W m\(^{-2}\)) and RMSD (167 W m\(^{-2}\)) values over a limited period (November and December 2011). Four data clusters representing morning, midday, early afternoon and late afternoon hours, respectively are found in Fig. 7, and RMSD values of approximately 81, 205, 228 and 60 W m\(^{-2}\) are found in morning, midday, early afternoon and late afternoon hours, respectively. High positive BIAS values are found at midday (128 W m\(^{-2}\)) and early afternoon (165 W m\(^{-2}\)) compared to morning (\(-\)1 W m\(^{-2}\)) and late afternoon (46 W m\(^{-2}\)). The low insolation regime may be either due to low shortwave fluxes in the morning in cloudless situations, or due to cloudy conditions. The uncertainties in predicting cloud and fog dynamics may be the major reason for this scatter. The aerosol load generally remains higher during the latter part of the morning for weekdays due to higher pollution loads (Backman et al. 2012). However, observations of surface aerosol concentration do not necessarily relate to total column aerosols. For example, the boundary-layer height is lower in the morning compared to the afternoon. So, equivalent amounts of aerosols will produce larger surface concentrations in the morning even though their radiative impact is the same. In Fig. 7, points lying along the diagonal represent the smallest aerosol loading in the morning (0300 UTC). During midday (0600 UTC) and early afternoon (0900 UTC), points cluster along a line (above the diagonal) with a shallow slope, where aerosol attenuation is larger compared to 0300 UTC, and it is larger when the incoming radiation is weaker (for the same time of day, this corresponds to a shallower sun angle and hence a longer atmospheric path length). In late afternoon (1200 UTC), the departure from the diagonal is the largest, though the net radiative effect is small due to the weak radiative intensities. The lack of representation of aerosol characterization is a major drawback of the WRF model, and so a lack of attenuation of shortwave radiative flux due to aerosol loading in the insolation regime could lead to higher predicted fluxes compared to tower measurements.

Fig. 7
figure 7

Scatter plot of downward shortwave radiative flux \((S^{\downarrow })\) forecast from the WRF model and micrometeorological tower observed \(S^{\downarrow }\) radiation during November and December 2011

Full size image

The incoming shortwave radiative flux is the major energy input in the surface radiation and energy balance. The WRF downward shortwave radiative flux forecasts depend on the cumulus parametrization scheme, which in turn is sensitive to surface relative humidity. Moreover, the roles of aerosol and Rayleigh scattering, and direct–diffuse irradiance partitioning have not been accounted for the WRF model for computing downward shortwave radiation. The prediction of cloud type and cloud-cover fraction from the atmospheric humidity profile is also a great source of uncertainty. The sensitivity of incoming shortwave radiation to aerosol optical depth in cloudless skies is shown by Bhattacharya et al. (2013b). In cloudy-sky conditions, the cloud-top albedo is the major determinant for attenuation of irradiances for the wavelength range of 300–750 nm (Stephens 1978), but both reflection and cloud absorption attenuate irradiances for wavelengths beyond 750 nm.

4.6.2 Downward Longwave

The 24-h WRF model predicted longwave radiative fluxes are compared with tower-observed incoming longwave fluxes in Fig. 8. A diurnal pattern is noted with RMSD (BIAS) values that gradually increase (decreasing) from 3 h and attain a peak (lowest point) at 6 h followed by a decrease (increase) towards 24 h. The minimum RMSD values gradually increase from day 1 to day 3, while the maximum RMSD values change very little. The 1:1 scatter plots of the WRF model predicted \(L^{\downarrow }\) and tower-observed \(L^{\downarrow }\) for November and December 2011 show RMSD values of 33 and 40 W m\(^{-2}\) and BIAS of \(-22\) and \(-21\) W m\(^{-2}\). The cyclic patterns of BIAS and RMSD values of \(L^{\downarrow }\) are quite similar to relative humidity but are opposite in nature. It is known that incoming longwave radiative flux depends on sky temperature and air emissivity. The latter is greatly influenced by actual vapour pressure. Relative humidity is the ratio of the actual vapour pressure and air temperature-dependent saturation vapour pressure, while \(L^{\downarrow }\) is directly proportional to the product of sky temperature and air emissivity. Therefore, the nature of errors for both temperature and actual vapour pressure may influence differently values of \(L^{\downarrow }\) and RH.

Fig. 8
figure 8

a Temporal distribution of RMSD (circle) and BIAS (triangle) values in the 3-h downward longwave radiative flux \((L^{\downarrow })\) and b scatter plot of 24-h WRF \(L^{\downarrow }\) forecast and micrometeorological tower observed \(L^{\downarrow }\) during November and December 2011

Full size image

4.7 Ground Heat Flux

Both RMSD and BIAS values of the WRF model predicted \(G_\mathrm{S}\) show prominent cyclic patterns (Fig. 9). The inward and outward nature of \(G_\mathrm{S}\) change with respect to day and night soil hydro-thermal regimes. The sign of the BIAS values depends on the sign of the ground heat flux. There is practically no increase or decrease in the magnitude of the minimum and maximum values of BIAS and RMSD as the forecast length increases up to 72 h. Depending on the time of day, RMSD values range from 47 to 113 W m\(^{-2}\) and BIAS ranges from \(-\)103 to 62 W m\(^{-2}\). The scatter plots of the WRF model predicted and tower-observed \(G_\mathrm{S}\) at 24 h in the months of November and December 2011 show that there is a general overestimation by the WRF model leading to positive BIAS values of 31, 30 W m\(^{-2}\) and RMSD values of 48, 50 W m\(^{-2}\), respectively. The scatter plots are broad in nature as compared to the narrow-spread scatter plots for \(L^{\downarrow }\).

Fig. 9
figure 9

a Temporal distribution of RMSD (circle) and BIAS (triangle) in 3-h \(G_\mathrm{S}\) and b scatter plot of 24-h WRF model \(G_\mathrm{S}\) forecast and micrometeorological tower-observed \(G_\mathrm{S}\) during November and December 2011

Full size image

4.8 Comparison of Errors of \(G_\mathrm{S}\) in Relation to Other Parameters

The nature of variation of RMSD and BIAS values with forecast lengths are similar in both \(G_\mathrm{S}\) and \(L^{\downarrow }\), the latter shows negative BIAS values for all forecast lengths and the former shows BIAS values ranging from negative to positive. On the other hand, predicted soil temperature consistently shows positive BIAS values. At equilibrium, the ground heat flux is in balance with the soil temperature gradient, and the surface radiation budget in terms of net radiation, sensible and latent heat fluxes. Both longwave and shortwave radiative fluxes govern net radiation with more contribution of the latter during daytime hours than the former.

The errors are relatively high in \(G_\mathrm{S}\) compared to \(L^{\downarrow }\). The latter is generally governed by atmospheric humidity, sky conditions, and atmospheric temperature, which have less spatial variability within a model grid resolution of 5 km. In contrast, the \(G_\mathrm{S}\) is mainly influenced by the soil hydro-thermal regimes depending on thermal conductivity and heat capacity based on soil composition and structure, and also fractional vegetation cover above the soil. These factors might produce a greater spatial variability in \(G_\mathrm{S}\) within a WRF model grid. The error statistics for \(G_\mathrm{S}\) are expected to show more variability within the WRF model grid than \(L^{\downarrow }\), but due to the relatively static nature of soil thermal conductivity and heat capacity over shorter times scales no shift was noticed in the minimum and maximum values of BIAS and RMSD in \(G_\mathrm{S}\), but a small shift is noticed for \(L^{\downarrow }\) with increase in the forecast length. Moreover, very little to no shift in the predicted soil temperature with increase in forecast length, but with increasing BIAS and RMSD values, are noticed for surface air temperature that might lead to no shift in the errors for \(G_\mathrm{S}\) but a shift in the errors for \(L^{\downarrow }\).

5 Conclusions

Short-range weather forecasts from the WRF model are compared with both surface weather and boundary-layer quantities such as shortwave, longwave and soil heat fluxes from micrometeorological towers over the Indian landmass for the first time. The tower network is important for quantifying the errors in the forecast of both weather and surface fluxes thus establishing the need to improve the model deficiencies. Large errors in the incoming shortwave radiation flux suggest a need for improved representation of clouds and the inclusion of aerosol loading for attenuation of the shortwave flux in cloudless situations. The diurnal pattern of errors in relative humidity is found to be oppositely correlated to that of the incoming longwave radiation flux. The BIAS values in ground heat fluxes during daytime hours are possibly dominated by the BIAS values in incoming shortwave radiation flux apart from other possibilities. During nighttime, the BIAS values in ground heat fluxes are possibly influenced by BIAS values in incoming longwave radiation and soil temperature. The sources of forecast errors in WRF model could be many. One source of error is a gross mismatch between representations in land-cover types by the WRF model and the micrometeorological towers. The latter of course represent local-scale cover type while the WRF model considers aggregated cover type corresponding to the model grid resolution (5 km). In addition, there are elevation differences between the WRF model grids and the collocated towers that could contribute to errors in surface temperature and surface pressure. Moreover, systematic and direct measurements of sensible and latent heat fluxes from a network of eddy-covariance systems or scintillometers could be used to evaluate model-predicted turbulent heat fluxes at the present scale of study.