Abstract

Operational analyses of 2-m temperature, 2-m humidity, and 10-m wind speed are verified independently against observations obtained from the WegenerNet, an extremely high-density, grid-type surface station network in southeastern Austria with an average distance between stations of 1.4 km. The Integrated Nowcasting through Comprehensive Analysis (INCA) system provides high-resolution analyses in space (1 km) and time (1 h) over the eastern Alpine region and has been specially designed for use in complex terrain. The quality of the system is investigated within a small domain with gentle topography ranging in elevation from 250 to 500 m. A comprehensive validation of INCA relative to WegenerNet for a 3-yr period from December 2007 to November 2010 indicates high analysis skill during all seasons. A sensitivity study reveals the importance of a sufficiently dense station network used by the system and, even more important, the relevance of adequate representativeness of the observation data.

1. Introduction

The analysis of meteorological surface fields at high resolution is becoming increasingly important for a variety of applications such as nowcasting, forecast downscaling, or verification. Analysis systems may be based on variational methods (Tyndall et al. 2010), on distance interpolation techniques (Haiden et al. 2011), or on spline interpolation and climatological pattern matching (Steinacker et al. 2006). These systems try to create continuous, gridded fields by combining irregularly distributed point observations with auxiliary information about spatial patterns. This auxiliary information can be based on a numerical weather prediction (NWP) model, on a theoretical model, on climatology, or on grid-type observations (radar, satellite).

An assessment of the quality of an analysis system is commonly made by cross validation, in which individual observations are withheld and the analyzed value at a location is compared with the observation. One limitation of this approach is that it gives a realistic estimate of the true analysis error at an arbitrary location only if the stations are distributed more or less randomly with regard to the underlying topography. The extremely high-density station network known as WegenerNet, operated by the Wegener Center for Climate and Global Change of the University of Graz (WEGC), consists of 151 stations positioned on an (almost) regular grid. It contains valley, slope, and hilltop locations and provides an opportunity for independent verification of the 1-hourly, high-resolution analyses produced operationally at the Austrian weather service. One advantage of using the dense WegenerNet measurement network is the possibility of verifying small-scale features that are usually not measured by the limited number of irregularly distributed (automated) stations. Thus, the ability of the Integrated Nowcasting through Comprehensive Analysis (INCA) system to capture those small-scale features can be analyzed systematically over a longer validation period. The INCA system is briefly described in section 2. Section 3 presents the WegenerNet high-density station network. Verification results as well as sensitivity analyses are given in section 4. Section 5 provides a summary and conclusions.

2. Analysis system

The 1 km × 1 km surface analyses used in this study are provided by the INCA system. INCA generates hourly analyses of temperature, humidity, and wind from a combination of NWP model output, surface station data, and high-resolution topographic data. Spatial interpolations are based on distance weighting in physical and potential temperature space. A detailed description of the analysis method is given in Haiden et al. (2011).

The Austrian operational INCA domain covers an area of 600 km × 350 km, centered over the eastern Alps. In the vertical direction a z system is used, where z is the height above the valley-floor surface. This is a spatially slowly varying reference surface that connects major valley floors. It has been designed for the downward extrapolation of three-dimensional NWP forecast fields into subgrid-scale valleys (Haiden et al. 2011).

NWP fields are provided by the Austrian operational version of the limited-area Aire Limitée Adaptation Dynamique Développement International (ALADIN) model, as described by Wang et al. (2006). It has a horizontal resolution of 9.6 km, has 60 levels in the vertical direction, and is run 4 times per day out to a forecast range of 72 h.

The Central Institute for Meteorology and Geodynamics (ZAMG) operates a network of ~250 real-time surface weather stations in Austria. The average distance between stations is 18 km. The Austrian network of hydrometeorological stations provides additional real-time observations at about 100 locations.

The three-dimensional analysis of temperature in INCA has been described by Kann et al. (2009) and Haiden et al. (2011). The analysis starts with an NWP short-range forecast as a first guess, which is corrected based on observation–forecast differences. The corrections are spatially interpolated using inverse-distance-squared weighting (IDW) in the horizontal plane and IDW in potential temperature in the vertical direction. A “surface-layer index” ensures that corrections derived at a certain type of location (e.g., valley floor) have low or zero weight at other types of location (e.g., slope). An analogous procedure is applied to the specific humidity field.

The IDW interpolation does not contain explicit distance scales. For a given number of nearest stations used, the radius of influence is determined by the local station density. The average distance between stations in the Austrian INCA domain is 15 km. Using eight nearest stations, the average distance scale would be .

The INCA wind analysis is also three-dimensional and is based on an NWP field as the first guess. The interpolation of observation corrections does not produce a mass-consistent field, however, and the NWP wind forecast does not fit to the high-resolution INCA topography. An iterative relaxation algorithm is applied to obtain a mass-consistent field. Wind vectors at grid points closest to station locations are kept at the observed values during the relaxation procedure (Haiden et al. 2011).

3. Surface station network

The WegenerNet climate station network is a pioneering experiment by the WEGC for high-resolution observation of weather and climate. The network comprises 151 meteorological stations within an area of approximately 20 km × 15 km, located in the region of Feldbach (~46.93°N, 15.90°E) in southeastern Austria. Results of previous studies underline the sensitivity of the region to changes in climate conditions (Heinrich 2008; Kabas et al. 2011a). WegenerNet has been providing regular measurements since the beginning of 2007. As shown in Fig. 1, the study area is orographically characterized by the Raab River valley (from northwest to east) and a hilly landscape with station altitudes between 257 and 520 m. The stations are arranged on a quasi-regular 1.4 km × 1.4 km grid and measure air temperature, relative humidity, and precipitation amount. Selected stations additionally provide measurements of wind and soil parameters. Air pressure and net radiation are observed at a reference station situated near the center of the area. All meteorological quantities are sampled at 5-min intervals.

Fig. 1.

WegenerNet study area in the district of Feldbach, Austria, and the station locations in the station grid. Different symbols indicate analyzed parameters of stations: air temperature (circles), air temperature and relative humidity (squares), and air temperature and wind speed (triangles). Within the WegenerNet, two synoptic stations of the ZAMG, Feldbach (11298) and Bad Gleichenberg (11245), are marked by stars. (Base map is taken from openstreetmap.org; hill-shade source is GIS-Steiermark).

Fig. 1.

WegenerNet study area in the district of Feldbach, Austria, and the station locations in the station grid. Different symbols indicate analyzed parameters of stations: air temperature (circles), air temperature and relative humidity (squares), and air temperature and wind speed (triangles). Within the WegenerNet, two synoptic stations of the ZAMG, Feldbach (11298) and Bad Gleichenberg (11245), are marked by stars. (Base map is taken from openstreetmap.org; hill-shade source is GIS-Steiermark).

The network was mainly set up during 2006 (Kirchengast et al. 2008), followed by the development of an automatic processing system that includes data transfer, quality control, product generation, and presentation at a Worldwide Web portal (Kabas et al. 2011b). In the quality control system, the raw data are tested for technical and physical plausibility. Data values of highest quality are used for the generation of data products in which station data and gridded data (1 km × 1 km) are produced on various weather and climate time scales (from 5 min and half-hourly up to annual). The products are available in near–real time (data latency is less than 1–2 h in standard operation) at the WegenerNet data portal (online at http://www.wegenernet.org).

The study presented here includes 150 observational stations for temperature and 27 stations with continuous measurements for humidity analyses. For wind speed, 7 of 12 stations have been selected according to representativeness, that is, the absence of local-scale flow effects and their ability to represent the wind field at the station distance scale (Fig. 1).

4. Verification

a. Overall skill

The skill of the INCA 2-m temperature, 2-m humidity, and 10-m wind speed results is quantified at the station locations of the WegenerNet based on standard point verification. Mean error (bias), mean absolute error (MAE), and root-mean-square error (RMSE) (Wilks 2006) are calculated on an hourly basis for the 3-yr period from 1 December 2007 to 30 November 2010. The INCA values at the WegenerNet locations are calculated by bilinear interpolation from the INCA grid. In general, the MAE of the INCA 2-m temperature analysis is below 1 K at all times of the day. The RMSE is about 1 K during daytime and is slightly higher during the night (approximately 1.3 K). The analyses show almost no systematic error from late afternoon until early morning but show a negative bias of about 0.5 K during daytime, especially in the (late) morning (Fig. 2a). Similarly, the performance of 2-m relative humidity shows a slight negative bias during daytime and almost bias-free values during the night. MAE and RMSE are about 5% and 7%, respectively, with the lower skill in the late afternoon (Fig. 2b). For 10-m wind speed, the MAE varies between 0.5 and 0.8 m s−1 and the RMSE varies between 1 and 1.4 m s−1. The mean observed wind speed is about 1.2 m s−1, leading to a relative MAE of ~50%. During daytime, a small negative bias occurs (Fig. 2c). The results generally confirm the high analysis skill reported in Haiden et al. (2011). Because the verification here is performed using data from an independent, high-resolution observation network that was not used in the development of the analysis system, however, the results also indicate that there has been no overfitting. The absolute values of the analysis error are close to the limits imposed by observation accuracy and representativeness, which are estimated to be on the order of 1°C, 5% relative humidity, and 1 m s−1.

Fig. 2.

Bias (dashed–dotted lines), MAE (solid lines), and RMSE (dotted lines) of (a) 2-m temperature, (b) 2-m relative humidity, and (c) 10-m wind speed.

Fig. 2.

Bias (dashed–dotted lines), MAE (solid lines), and RMSE (dotted lines) of (a) 2-m temperature, (b) 2-m relative humidity, and (c) 10-m wind speed.

On the basis of 3 yr of data, the seasonal variability of the temperature error appears to be fairly small (Fig. 3a), although the monthly mean errors (not shown) vary for distinct seasons. In other words, the seasonally averaged skill of the temperature analysis remains largely constant throughout the year. For relative humidity, a negative bias is found during the summer months that might be introduced by unrepresentative station values. MAE and RMSE are of similar magnitude in both summer and winter (Fig. 3b). The seasonal variation of the wind speed error reveals a general negative bias that is larger in the winter months (Fig. 3c), although the mean observed wind speeds of both summer and winter periods are of similar magnitude. This could be due to the weakness of the NWP model in simulating elevated inversions with a well-mixed layer below. Instead, the model tends to produce surface inversions with a decoupled surface layer, where wind speeds become low.

Fig. 3.

Seasonal distribution of bias (gray), MAE (black), and RMSE (hatched) for (a) 2-m temperature, (b) 2-m relative humidity, and (c) 10-m wind speed (DJF: December–February; MAM: March–May; JJA: June–August; SON: September–November).

Fig. 3.

Seasonal distribution of bias (gray), MAE (black), and RMSE (hatched) for (a) 2-m temperature, (b) 2-m relative humidity, and (c) 10-m wind speed (DJF: December–February; MAM: March–May; JJA: June–August; SON: September–November).

The frequency distribution of temperature errors shows that about 50% of the values are found between −0.5 and +0.5 K. Errors exceeding ±1.5 K occur with a probability of about 5% (Fig. 4a). The frequency distribution of the humidity error shows a slight negative bias and more than 50% probability for errors between −5% and +5% (Fig. 4b). Similarly, the error distribution of wind speed shows a slight negative bias, combined with a general skill of approximately 50% of values within ±0.5 m s−1 (Fig. 4c).

Fig. 4.

Error frequency distribution for (a) 2-m temperature, (b) 2-m relative humidity, and (c) 10-m wind speed.

Fig. 4.

Error frequency distribution for (a) 2-m temperature, (b) 2-m relative humidity, and (c) 10-m wind speed.

b. Sensitivity study

The skill of the INCA is determined by the quality of the NWP model background and the number and representativeness of observations entering the system. In principle, the more stations that are available for correcting the first-guess field, the more realistic the spatial distribution of the considered quantity should be. In the study area selected here the observations of primarily two automatic stations, Feldbach and Bad Gleichenberg, both located within a distance of 10 km, largely determine the INCA (Fig. 1). The question arises as to what extent the two stations influence the spatial mean error and how representative they are of the network area. The following four experiments have been carried out for two 3-month periods (July–August 2009 and December 2009–February 2010) for 2-m temperature and 2-m relative humidity:

  1. EXP-1 is an operational version including both Feldbach and Bad Gleichenberg,

  2. EXP-2 is reanalyses with Feldbach and without Bad Gleichenberg,

  3. EXP-3 is reanalyses without Feldbach and with Bad Gleichenberg, and

  4. EXP-4 is reanalyses that exclude both Feldbach and Bad Gleichenberg.

For EXP-4, the skill of the analyses mainly depends on the quality of the NWP background fields, corrected by station data from outside the area (Haiden et al. 2011). As expected, EXP-4 shows the poorest performance (Fig. 5a). The MAE of the model background, corrected by stations farther away, is about 20% higher than the operational one that includes both stations (EXP-1). Without the Feldbach station (EXP-3), the quality is slightly worse than the operational version, especially during late morning. Unexpectedly, the experiment without Bad Gleichenberg (EXP-2) performs slightly better than the operational version with both stations (EXP-1), especially between 0500 and 1000 UTC. A comparison of the mean daily temperature evolution reveals that this station, which is located in a small basin, behaves slightly atypically relative to the mean exposition of the WegenerNet stations. It suffers from limited representativeness, which is most pronounced during the night and in the morning, when the near-surface stratification is strongest. For relative humidity, similar results are obtained. EXP-4, excluding the two stations, performs worst. The inclusion of Bad Gleichenberg (EXP-3) leads to slightly better results than the operational version, particularly during the second half of the night (Fig. 5b). On the other hand, the inclusion of the humidity measurements at Bad Gleichenberg has a positive impact on the analysis during daytime. The main reason appears to be the spatial distribution of stations used in the humidity verification. Most of them are clustered around the Bad Gleichenberg station and only a few are located in the vicinity of Feldbach (Fig. 1). Thus, Feldbach is affected more by limited representativeness than is Bad Gleichenberg in this case.

Fig. 5.

MAE of experiments EXP-1 (solid lines), EXP-2 (dashed lines), EXP-3 (dotted lines), and EXP-4 (solid lines with triangles) for (a) 2-m temperature and (b) 2-m relative humidity.

Fig. 5.

MAE of experiments EXP-1 (solid lines), EXP-2 (dashed lines), EXP-3 (dotted lines), and EXP-4 (solid lines with triangles) for (a) 2-m temperature and (b) 2-m relative humidity.

The fact that at certain times of day it is beneficial to drop a station from the analysis is a result of the imperfection of the analysis system. The first guess has 10-km resolution, and even at the 1-km high resolution of INCA there are meteorologically significant but unresolved topographic features. Thus, a physical or statistical “station model” could be applied that translates a point observation into a grid-scale value before it enters the analysis. Also, the current procedure of parameterizing surface inversions in INCA with a constant-surface-layer index field (Haiden et al. 2011) needs to be improved.

5. Summary and conclusions

Operational analyses of the INCA system were validated using the WegenerNet, an extremely dense station network in southeastern Austria. Verification results generally indicate a high skill of the INCA system with respect to 2-m temperature and 2-m relative humidity and lower skill for 10-m wind speed (small biases, and RMSEs of about 1–1.3 K, 5%–7%, and 1–1.4 m s−1, respectively). The diurnal variability of the error is larger than the seasonal variations, and the number of outliers (e.g., greater than ±3 K) is small. A sensitivity study shows the impact of including observations that are representative of the areal mean. The inclusion of more stations in the analysis does not always produce better results; the analysis may degrade in cases of low representativeness. For the analysis of temperature, humidity, and wind in complex terrain, additional skill can be gained by selectively withholding data from certain stations with limited representativeness (or by reducing their weight), depending on the time of day. One possible approach would be a station model that translates a point observation into a grid-scale representative value before use in the analysis. High-resolution networks like the WegenerNet are ideally suited for guiding the construction of such models and for their validation.

Acknowledgments

We are grateful to three anonymous reviewers for their suggestions and constructive comments. The WegenerNet work that supports this study was funded by the Provincial government of Styria, Austria; the University of Graz; and the City of Graz, which also provide the main funding for WegenerNet operations and maintenance. The WegenerNet setup was also cofunded by the Austrian government, the 27 municipalities of the station region, and further regional sponsors.

REFERENCES

REFERENCES
Haiden
,
T.
,
A.
Kann
,
C.
Wittmann
,
G.
Pistotnik
,
B.
Bica
, and
C.
Gruber
,
2011
:
The Integrated Nowcasting through Comprehensive Analysis (INCA) system and its validation over the eastern Alpine region
.
Wea. Forecasting
,
26
,
166
183
.
Heinrich
,
G.
,
2008
:
Klimawandel und Trockengefährdung in der Landwirtschaft: Eine Analyse für die Steiermark (Climate change and drought danger in agriculture: An analysis for Styria)
.
Wegener Center for Climate and Global Change Sci. Rep. 20-2008, University of Graz, Graz, Austria, 132 pp. [Available online at http://wegc26.uni-graz.at/extern/pdf/WCV-WissBer-Nr20-GHeinrich-Jan2008.pdf.]
Kabas
,
T.
,
U.
Foelsche
, and
G.
Kirchengast
,
2011a
:
Seasonal and annual trends of temperature and precipitation within 1951/1971-2007 in south-eastern Styria, Austria
.
Meteor. Z.
,
20
,
277
289
.
Kabas
,
T.
,
A.
Leuprecht
,
C.
Bichler
, and
G.
Kirchengast
,
2011b
:
WegenerNet climate station network region Feldbach, Austria: Network structure, processing system, and example results
.
Adv. Sci. Res.
,
6
,
49
54
.
Kann
,
A.
,
C.
Wittmann
,
Y.
Wang
, and
X.
Ma
,
2009
:
Calibrating 2-m temperature of limited area ensemble forecasts using high-resolution analysis
.
Mon. Wea. Rev.
,
137
,
3373
3387
.
Kirchengast
,
G.
,
T.
Kabas
,
S.
Binder
,
A.
Leuprecht
, and
C.
Bichler
,
2008
:
Pionierexperiment WegenerNet Klimastationsnetz: Ein neuartiges Messnetz in der Region Feldbach (Steiermark/Österreich) zur Beobachtung von Wetter und Klima mit sehr hoher Auflösung [WegenerNet climate station network pioneering experiment: A novel monitoring network in the Feldbach region (Styria/Austria) for the observation of weather and climate with very high resolution]
.
Wegener Center for Climate and Global Change Sci. Rep. 23-2008, University of Graz, Graz, Austria, 76 pp. [Available online at http://www.uni-graz.at/en/igam7www_wcv-wissber-nr23-gkirchengastetal-sep2008.pdf.]
Steinacker
,
R.
, and
Coauthors
,
2006
:
Downscaling meteorological information over complex terrain with the fingerprint technique by using a priori knowledge
.
Mon. Wea. Rev.
,
134
,
2758
2771
.
Tyndall
,
D. P.
,
J. D.
Horel
, and
M. S. F. V.
de Pondeca
,
2010
:
Sensitivity of surface air temperature analyses to background and observation errors
.
Wea. Forecasting
,
25
,
852
865
.
Wang
,
Y.
,
T.
Haiden
, and
A.
Kann
,
2006
:
The operational limited area modelling system at ZAMG: ALADIN-AUSTRIA
.
Austrian Contributions to Meteorology and Geophysics. Vol. 37, Zentralanstalt für Meteorologie und Geodynamik (ZAMG), Vienna, Austria, 33 pp
.
Wilks
,
D. S.
,
2006
:
Statistical Methods in the Atmospheric Sciences
.
2nd ed. Academic Press, 627 pp
.