1 Introduction

The issue of terrain impact on air flow in the atmospheric boundary layer is often considered in meteorological research, especially in the context of mesoscale and microscale anemological phenomena (Barry 2008). Orographically induced air flow modifications, observed in mountainous or hilly regions of the Earth, became a subject of studies related to numerical modelling, including GIS techniques. Within this scope, two general approaches may be distinguished. The first is based on indirect evaluation of potential anemological conditions according to land surface parameters (Böhner and Antonić 2007). The second is direct modelling, which requires considerable computational power and detailed input data.

Direct modelling of the wind field may be conducted in mesoscale and microscale, regardless of the model type. It is important to distinguish between meteorological models and computational fluid dynamics (CFD) models. Dynamic mesoscale models like HIRLAM (Undén et al. 2002) and WRF (Skamarock et al. 2005) provide a more realistic physical description of the atmosphere; they are capable of simulating synoptic processes and the interactions between the Earth’s surface and atmosphere (Truhetz 2010). On the other hand, the complexity of these models requires extensive computational resources. Thus, their spatial resolution usually does not exceed 5 km. CFD models, often defined as diagnostic models (Ratto et al. 1994; Truhetz 2010), have been developed since the 1970s (Jackson and Hunt 1975). Models from this group are capable of simulating only steady-state flows. In consequence, they are suited to simulate orographic speed-up effects, but cannot be expected to solve any dynamic process like flow splitting, vortex shedding or thermal induced circulation (Truhetz 2010). The continuous growth of available computational power facilitates the development of more sophisticated CFD-models which are usually based on Navier–Stokes equations but differ in numerical approaches and turbulence models (direct numerical simulation, Reynolds-averaged Navier–Stokes, large eddy simulation) (Uchida and Ohya 2003; Undheim 2005). Regardless of the technical details of a model, any simulation strongly depends on wind fields generated from observational data, which are used to initialize the model and thereby determine the model’s performance (Lopes 2003, 2011; Truhetz 2010). Currently, a popular approach to achieve more detailed initial wind fields assumes the use of data from re-analysis driven dynamic mesoscale models; a good example is the application of MM5 (dynamic) and CALMET (diagnostic) models (Barna and Lamb 2000; Chandrasekar et al. 2003; Truhetz 2010).

Diagnostic models, while relatively simplified, are rather used to acquire high-resolution phenomena simulation (Undheim 2005). The studies on diagnostic modelling of wind velocity are usually executed within a resolution higher than (or equal) 1 km. The results of the evaluation of models vary depending on resolution, terrain complexity and time span. For example, evaluation of CALMET at the test-site in Nevada (1-km grid spacing) provided a mean velocity RMSE = 1.7 m/s (Cox et al. 2005); the application of the same model (and resolution) in long-term simulations in the south-central region of Chile provided RMSE—1.03 m/s (Morales et al. 2012). On the other hand, short-term simulations in the Alpine region, using CONDOR and MINERVE models (500-m grid spacing) provided a mean RMSE of 0.68 and 0.65 m/s, respectively (Desiato et al. 1998). Experimental evaluation of Meteodyn WT and WindMap models, conducted on four different test sites and using relatively high spatial resolution (50-m grid spacing) resulted in a RMSE ranging from 0.75 to 1.02 m/s (Beaucage and Brower 2012).

Modelling of wind field using the aforesaid techniques requires topographic information in addition to the necessary meteorological input data. In this case, the most relevant are digital elevation and vegetation (land use) data (Emeis and Knoche 2007). Both types of data may be provided by remote sensing technology. Freely accessible elevation models developed on the basis of satellite imagery, such as Global Land Survey Digital Elevation Model (GLS DEM), Shuttle Radar Topography Mission (SRTM) or Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model (ASTER GDEM), cover the majority of the Earth’s surface (Nelson et al. 2007). The same situation applies to land cover information derived from satellite imagery (Landsat TM and ETM+). In consequence, the accuracy and spatial resolution of the aforementioned data may affect the results of the modelling. Accordingly, experimental studies on remote sensing data usage seem to be indispensable in order to obtain an information about the roughness patterns of the land surface; the precision of this type of input data should be appropriate to the spatial resolution of the model (Wakes et al. 2010).

Within the scope of wind velocity modelling, the area of the Sudetes proves particularly interesting due to high intensity of orographically induced anemological phenomena. The highest wind speed values are typical for the ridges in the Karkonosze, the Śnieżnik Massif and Hruby Jesenik where, besides the impact of orography, a significant role is played by low roughness length values determined by the existing types of vegetation (Sobik and Błaś 2010).

Based on the current state of climatologic studies in the Sudetes, it is apparent that the anemological conditions of the selected mountain range, the Śnieżnik Massif, are not as thoroughly investigated as they probably deserve (Piasecki 1996). Within the mentioned area, any potential attempts of numerical approximation of wind field encounter an additional obstacle: lack of long-term meteorological data series. This is in part a consequence of low density and irregular distribution of working synoptic stations inside the entire area of the Sudetes. As a result, in the current conditions and within the selected area, a feasible approach of anemological studies (using GIS and remote sensing techniques) should be based on series of short-term field measurement. Such studies may be valuable as they contribute to extending knowledge on the local meteorological phenomena.

This study constitutes an attempt of wind-field modelling in a part of the Śnieżnik Massif. The main objectives of the study involved:

  1. 1.

    Validation of the chosen CFD solver’s performance in terms of limited input data based on wind velocity field measurements taken at height of 2 m above ground.

  2. 2.

    Validation of usefulness of GLSDEM and Corine Land Cover 2006 data in the process of modelling based on input data referred to in item 1.

2 Study Area

The Śnieżnik Massif is the second-highest mountain range in the Eastern Sudetes. Its highest peak is Śnieżnik (1,425 m). The massif lies in two countries: the north-western part in Poland and the south-eastern part, also known as Králický Sněžník (Fig. 1), in the Czech Republic. The climatic subregion of the Śnieżnik Massif is separated from other geographical-climatic units by areas of intermontane valleys and basins; it represents a distinct orographic barrier for inflowing atmospheric masses (Piasecki and Sawiński 2009). The massif’s form causes mesoscale and local deformations in the air flow. Depending on the atmospheric circulation type, the air masses flowing into the valleys adopt a direction of movement along valley axes. These valley air-currents were termed anemo-orographic systems of air flow (Jenik 1961). In favourable conditions, these systems extend over and include valleys located on the leeward, usually northern, side of the massif (occurence of foehn winds). To date, three anemo-orographic systems have been mentioned in the Śnieżnik Massif: Wilczka/Bogoryja—Czarna Woda/Kleśnica, Velka Morava—Kleśnica/Kamienica and Krupa—Morawka. Prevailing wind directions, observed by J. Kwiatkowski throughout 1973–1975, range from W to S (SSE) depending on the positioning of the measurement station (Piasecki 1996; Piasecki and Sawiński 2009) (Fig. 2).

Fig. 1
figure 1

Position of the study area (marked by the red rectangle)

Full size image
Fig. 2
figure 2

Mean annual frequency of wind directions in the Śnieżnik Massif in relation with local anemo-orographic wind systems: Wilczka—Pogonna (a), Velka Morava—Kleśnica, Kamienica (b), and Krupa—Morawka (c) [according to Piasecki and Sawiński (2009)]

Full size image

In the presented case, a rectangular research polygon was delimited in the northwestern part of the massif, covering an area of 120 sq km (Fig. 1). It contains the peak of Śnieżnik, the main northern ridge (peaks: Żmijowiec, Średniak, Smerkowiec, Czarna Góra), and the Wilczka and Kleśnica valleys (Fig. 3). Within the presented area, the altitude ranges from 421 m in the western part, to 1,425 m at the top of Śnieżnik. A majority of the presented area is covered by coniferous, mixed and broad-leaved forests. The one exception is the western part dominated by agricultural forms of land use.

Fig. 3
figure 3

Distribution of wind measurement points inside the study area

Full size image

3 Methods

The modelling process was carried out using WindStation 2.0.10 software, which is based on CFD solver Canyon (Lopes 2003, 2011). It solves for mass conservation, momentum conservation (Navier–Stokes equations), energy conservation and turbulence quantities (k–ε model) (Lopes 2011). The model can be initialized by point data (measurement stations) providing wind speed and direction readings from two heights above the land surface; at least one point is necessary. Values of wind velocity and direction are interpolated for each node of the computational domain using the inverse of the distance as the weighting factor (Lopes 2011). This results in an initial wind field, that subsequently undergoes further processing depending on the model settings. The final result of the modelling is a three-dimensional grid, whose nodes contain information on xyz coordinates, velocity components, direction and turbulence intensity. The first versions of WindStation, presented in 2003, were validated using data obtained from the Askervein Hill site and two test areas in Portugal (Lopes 2003). Later versions were used in a study on the estimation of wind energy resources of the El-Kef region in Tunisia (Abbes and Belhadj 2012).

According to the circumstances mentioned in the introduction, the assumed manner of obtaining necessary input information was field measurements, conducted in short (6 or 7 h every day) series between November 2011 and June 2012. Due to Canyon’s limitations, typical for a diagnostic model, measurements were planned for periods of constant wind direction and relatively high wind velocities (forced by horizontal pressure gradient) within and around the study area. These initial decisions concerning the choice of measurement periods were based on the analysis of the data from weather forecast web services (provided by wetteronline.de, new.meteo.pl). Following the measurements, data from the nearest synoptic stations placed in a distance of 23–37 km from the center of the study area (Polom, Serak, Kłodzko) (Fig. 1) were analyzed for the purpose of detecting temporal inconsistencies of wind field near the Śnieżnik Massif. Further processing of field measurements data was conducted only if these stations recorded steady wind velocity and direction during field measurements. Datasets from the mentioned stations were provided by the National Oceanic and Atmospheric Administration (ftp://ftp.ncdc.noaa.gov/pub/data/noaa/isd-lite, access date, June 10, 2012).

The foregoing procedure limited the experimental dataset to 3 days of measurements: 26 November 2011 and 25–26 May 2012. On 26 November 2011 (Fig. 4a), the Sudetes area was within the northeastern periphery of vast high-pressure area “Afra” (the names of pressure systems are quoted according to the Institute for Meteorology at the Free University of Berlin; http://www.met.fu-berlin.de/wetterpate). An inflow of maritime polar air masses occurred after the passing of the cold front connected to low “Xaver” with a centre in the northern Scandinavia. As a result, throughout the observation period, the wind conditions inside and near the study area were temporarily constant. Wind direction ranged from W to SSW, depending on the topographic position of the observation points. Mean wind velocity ranged from 3 m/s on the lowest altitudes to 9 m/s in the case of stations with windward exposure placed on the highest peaks (Fig. 5a). On 25 and 26 May 2012, the study area was placed within the southern periphery of high-pressure area “Petermartin” extending above northern Europe. This situation caused an increment of horizontal pressure gradient and resulted in a moderate advection of continental polar air masses from the NE direction (Fig. 4b, c). Again, the wind conditions were temporarily constant during both observation periods. The wind direction ranged from N to ENE and wind velocity ranged from 4 to 10 m/s on 25 May (Fig. 5b) and from 3 to 8 m/s (Fig. 5c). It should be noted that on 26 May, the pressure gradient started to decrease and wind velocity values were slightly lower. Nevertheless, the general wind direction did not change.

Fig. 4
figure 4

Atmospheric pressure field over Europe: a 26 November 2011, b 25 May 2012, c 26 May 2012 [according to the data from the Austrian Central Institute for Meteorology and Geodynamics (ZAMG)]

Full size image
Fig. 5
figure 5

Wind speed and direction observed within study area and in Serak and Kłodzko synoptic stations during measurement time-periods: a 26 November 2011, b 25 May 2012, c 26 May 2012. The selected observation points represent the ridge area and foothills of the Śnieżnik Massif (Fig. 3; Table 1)

Full size image

During the measurement periods, the velocity readings were taken at a height of 2 m above ground level at 5-min intervals. The measurement instruments were Kaindl Windmaster 2 anemometers, for which the estimated measure error equals ±4 % within a range from 0.7 to 42 m/s, according to the manufacturer’s specification (http://www.windtronic.com/02-e.html). The estimation of wind direction was based on the observation of banners mounted on poles at wind speed measurement points. The direction was estimated to the nearest of the 16 points of the compass, in accordance with the statement of the Commission for Instruments and Methods of Observations (CIMO) concerning simple instrumental methods of surface wind measurement (WMO 2008). Due to technical and organizational reasons, the measurements were taken in 6-h periods, starting at 9:00 and ending at 15:00.

The list of measurement points is presented in Table 1, while their spatial distribution is presented in Fig. 3. The first objective of positioning anemometers in the presented way was to record wind velocity and direction at various altitudes. The second was to conduct measurements in conditions of variable relative exposure to wind direction (windward slopes, leeward slopes, flat land). Setting measurement points inside the areas of potentially high roughness (Czarna Góra) or surrounded by topographic obstacles (Międzygórze 2) was premeditated. The author did not include data from these points as input data for the model, but used them exclusively to validate results of the modelling. The locations of other points were also carefully selected following field reconnaissance to minimize the potential impact of topographic obstacles—the position of every measurement point was determined by the actual topographic situation, not by remote sensing land-use data.

Table 1 Measurement points used in the wind velocity modelling case study
Full size table

The next step of data pre-processing involved a reading-based calculation of hourly mean velocity values and prevailing directions in order to use them as an input data to the model. Information about wind conditions in the upper parts of the atmospheric boundary layer was based on the upper air soundings performed in stations nearest to the study area: Praha-Libus, Prostejov and Wrocław. The results of the soundings were provided by the Department of Atmospheric Science at the University of Wyoming (http://www.weather.uwyo.edu/upperair/sounding.html, access date: June 10, 2012).

The second part of the input dataset contained information on land surface (elevation and roughness length) and was based on remote sensing data. Two digital elevation models were used: Global Land Survey Digital Elevation Model (GLSDEM) and Digital Terrain Elevation Data (DTED) Level 2. The former is a 90-m resolution model, developed by USGS (United States Geological Survey) and provided by GLCF (Global Land Cover Facility http://www.glcf.umiacs.umd.edu/data/glsdem, access date: 13 March 2012). The latter is a 30-m resolution model, based on military 1:50,000 topographical maps, and was used for comparison. Roughness length information was based on Corine Land Cover 2006 (version 15) 100-m resolution raster dataset, provided by the European Environmental Agency (EEA, http://www.eea.europa.eu/data-and-maps/data/corine-land-cover-2006-raster-1, access date: 13 March 2012). The land use map of the study area based on the CLC 2006 data is presented in Fig. 6. To avoid discordances which could affect the model’s performance, CLC 2006 data were compared to Polish and Czech orthophotomaps (provided by GEODIS BRNO, spol. s.r.o. and geoportal.gov.pl web service) and, where necessary, corrected. The biggest discordances manifested themselves in the upper tree line (example on Fig. 7) and were probably a consequence of differences in spectral response caused by terrain parameters such as aspect or slope. In order to detect the impact of corrections on the model’s performance, the subsequent simulations were conducted using original and corrected roughness data in an alternating manner. Roughness length assignment was mainly based on the Finnish Wind Atlas classification (http://www.tuuliatlas.fi/modelling/mallinnus_3.html, access date: March 15th, 2012). In one problematic case (CLC class “transitional woodland-shrub” is not included in the Finnish Wind Atlas) an assignment problem was also considered according to the proceedings of European Wind Energy Conference 2007 (Silva et al. 2007). The final assignment results are presented in Table 2. Both elevation and roughness data were preprocessed in ArcInfo software and exported to ASCII format, required by WindStation.

Fig. 6
figure 6

Land use of the study area based on Corine Land Cover 2006 raster version

Full size image
Fig. 7
figure 7

Example of discordances between CLC 2006 classification results and actual land use on the peak of Śnieżnik. CLC 2006 polygons are indicated by numbers: 24 coniferous forest, 26 natural grasslands, 29 transitional woodland-shrub (orthophotomap source: GEODIS BRNO, spol. s.r.o.)

Full size image
Table 2 Roughness length classes based on Corine Land Cover 2006
Full size table

Modelling was performed using various resolutions of the calculation domain (100 and 250 m) and various input data (elevation model, constant—0.05 m or CLC 2006 based-on roughness). The model top was defined as 2,500 m; this value was a minimum for the model to perform properly. Finally, the computational grid had 146 × 126 × 20 nodes (100-m resolution) or 59 × 51 × 20 nodes (250-m resolution). The first node was placed at a height of 4 m. To enable comparison of results with previous studies (Lopes 2003), an assumption of neutral atmospheric stability was made. Thus, the further conclusions of the study do not apply to the modelling in stable or unstable atmosphere.

Raw output data were exported to ArcInfo, converted to a shapefile vector format and subsequently converted to a raster format using the spline interpolation method.

In order to evaluate the model’s performance, a modified leave-one-out cross-validation was executed. For every repetition of the model’s performance, the original wind velocity/direction dataset was divided into two subdatasets: the input (training) dataset and the validation dataset. In theory, the data obtained from seven observational sites working at the same time, allowed for the creation of 119 different combinations of input datasets (containing from 2 up to 6 points) per hour. However, due to the limited time and computer resources, it would be impracticable to use all of them (the study would have ended after 5,712 runs of the model). Instead, it was decided that 20 different training datasets would be used per every observational hour. It was assumed that the half of these sets had to contain at least five measurement points while the rest included at least two measurement points. Within the aforesaid conditions, the datasets were chosen randomly as the deliberate choice of training datasets could affect the results of the experiment. As a result, 120 runs of Canyon were performed for every day of measurements and for every specific model setup. This enabled a calculation of the following performance indices: velocity Bias (B v), root mean square error of velocity (RMSEv), index of wind speed (I v) direction mean absolute error (MAEdir) and index of wind direction (I dir). The equations for the indices are presented in Table 3. Figure 8 contains additional plots for I v and I dir which were used for comparison of the model’s performance with Poiares and Trevim sites (Lopes 2003).

Table 3 Error measures used in evaluation of overall wind field modelling results
Full size table
Fig. 8
figure 8

a Relationship between direction mismatch and index of wind direction (I dir), b relationship between velocity mismatch and index of wind velocity (I v)

Full size image

The next step involved the execution of Student’s t test for paired samples in order to determinate statistically whether the changes of the resolution, roughness or DEM setup affect significantly the results of the modelling. Finally, fields of the mean velocity and the velocity range were calculated for selected hours and model setups. This calculation was conducted using the focal statistics tools; it was based on the raster representations of the wind velocity at a height of 2 m above ground level, created during the consecutive runs of the model (described in the previous paragraph). The main purpose of the aforementioned calculation was to present graphically some examples of the spatial variability of the modelled wind velocity field (mean velocity) and differences in the modelled wind field which occur if various sets of input measurement points are used in the modelling (velocity range, defined as V max − V min).

4 Results

As a final result of the modelling and data post-processing, it was possible to present graphically the simulated wind field in the study area. The examples of maps displaying mean modelled wind velocity and modelled velocity range are given in Figs. 9, 10 (26 November 2011; 13:00) and Figs. 11, 12, 13 and 14 (26 May 2012; 13:00). Images generated for 25 May were very similar to those for the next day, due to the same prevailing wind direction. As is noticeable, in both presented episodes, the highest wind velocity values at the considered height of 2 m are typical for peaks and ridges above 1,000 m a.s.l. (Śnieżnik, Średniak, Żmijowec, Czarna Góra and Jaworowa Kopa). Relatively high velocities can also be noticed on pastures located on windward slopes (example: pastures placed south from Międzygórze on Fig. 11) and in windward areas at low altitudes, which are mainly arable lands.

Fig. 9
figure 9

Spatial distribution of mean modelled wind velocity (26 November 2011; 13:00). The values were calculated from the results of 20 simulations based on various combinations of the input measurement points (containing at least two points). The model’s setup included roughness length information derived from CLC 2006

Full size image
Fig. 10
figure 10

Spatial distribution of modelled wind velocity range (26 November 2011; 13:00). The values (V max − V min) were calculated from the results of 20 simulations based on various combinations of the input measurement points (containing at least two points). The model’s setup included roughness length information derived from CLC 2006

Full size image
Fig. 11
figure 11

Spatial distribution of mean modelled wind velocity (26 May 2012; 13:00). The values were calculated from the results of 20 simulations based on various combinations of the input measurement points (containing at least two points). The model’s setup included roughness length information derived from CLC 2006

Full size image
Fig. 12
figure 12

Spatial distribution of modelled wind velocity range (26 May 2012; 13:00). The values (V max − V min) were calculated from results of 20 simulations based on various combinations of the input measurement points (containing at least two points). The model’s setup included roughness length information derived from CLC 2006

Full size image
Fig. 13
figure 13

Spatial distribution of mean modelled wind velocity (26 May 2012; 13:00). The values were calculated from the results of 20 simulations based on various combinations of the input measurement points (containing at least two points). A constant roughness length (z 0 = 0.05 m) was assumed

Full size image
Fig. 14
figure 14

Spatial distribution of modelled wind velocity range (26 May 2012; 13:00). The values (V max − V min) were calculated from results of 20 simulations based on various combinations of the input measurement points (containing at least two points). A constant roughness length (z 0 = 0.05 m) was assumed

Full size image

The spatial distribution of mean wind velocity tends to vary due to differences between the main wind directions (WSW on 26 November and NNE on 25 and 26 May). The model of the November episode (Fig. 9) presents a very well developed western windward area of higher velocities and the eastern leeward area, where wind velocity is significantly lower. On the other hand, in the situation observed on 26 May (Fig. 11) this division is more explicit. The main reason for this difference is the relation of the direction of atmospheric circulation to the orientation of the main convex terrain forms (mountain ridges)—almost perpendicular on 26 November and slantwise on 26 May. Another noteworthy aspect is the changes in spatial variability of the modelled velocity range—simulations for the May episode are significantly more diverse in the area of western foothills of the massif (Fig. 12), than in case of the November episode (Fig. 10). Nevertheless, within the greater part of study area, velocity range does not exceed 1.0 m/s. The aforementioned remarks relate to the results of modelling with the use of input variable roughness information. The assumption of uniform roughness completely changes the wind field image (Figs. 13, 14)—it becomes a strict reflection of the local relief and its morphometric parameters. It should be noted that in this case the values of velocity range are higher as compared to the results achieved by the model’s performance using CLC 2006 roughness (Fig. 12).

Lowering the spatial resolution of the model produces a less-detailed wind-field simulation (Fig. 15). Differences can be observed in the whole area, although a tendency to estimate lower velocities above isolated peaks and ridges is the most evident—the area near the top of Średniak seems to be a good example.

Fig. 15
figure 15

Example of wind velocity modelling results calculated within various spatial resolutions (a 250 m, b 100 m); the modelling was based on data from Śnieżnik 1 and Łąki Myśliwskie measurement points (26 November 2011; 13:00) with roughness length information derived from CLC 2006

Full size image

The analysis of model performance indices reveals that, within the given domain resolution, using different DEMs does not change significantly the accuracy of the modelled wind velocity field. The DTED model provides slightly better index values, however GLS DEM may be successfully used instead. Including the roughness length information (derived from CLC 2006), significantly improves accuracy of the model in case of a 100 m resolution. It is expressed by decreasing RMSEv and increasing I v values (Fig. 16)—respectively to RMSEv = 1.1 m/s and I v = 80 for GLSDEM, for DTED RMSEv = 1.0 m/s and I v = 82. However, including the same information in a lower (250-m) model resolution causes a slight decrement of both RMSEv and I v. This indicates that the mean absolute error is decreased, but simultaneously the model’s relative mean performance is also affected. This is probably a consequence of relatively worse results for locations characterized by low wind velocity (Międzygórze 2) or in locations surrounded by complex roughness patterns which could not be properly presented in a resolution lower than 100 m (Mariańskie Skały, Puchaczówka). Significant changes of B v values (Fig. 16) are also evident. In the presented case, the model tends to overestimate velocity if a constant (but relatively low) value of roughness is set. It also tends to underestimate velocity if CLC 2006 roughness information is added (the underestimation is greater in coarser domain resolution). According to the modelled wind direction, changes of input data configuration and resolution did not affect markedly the accuracy of the model. Mean MAEdir values ranged from 28° to 32° (Fig. 16).

Fig. 16
figure 16

Impact of domain’s resolution on the overall values of wind field modelling performance indices: a mean B v, b mean RMSEv, c mean I v, d mean MAEdir. Bar fill indicates various input elevation and roughness data configuration: white GLS DEM with uniform roughness, grey DTED with uniform roughness, patterned white GLS DEM with CLC 2006 roughness, patterned grey DTED with CLC 2006 roughness

Full size image

A more detailed insight into performance indices’ values, calculated for particular measurement points and pertaining to the domain resolution and the input data, is presented in Tables 4 and 5. Differences between B v and RMSEv values, calculated for a 100-m and 250-m resolution, show that including variable roughness length information improves the results of modelling in a resolution consistent with the land use data source. This is particularly noticeable in the case of points placed in forested areas (Czarna Góra, Mariańskie Skały, Żmijowa Polana, Średniak); a particular example is the Czarna Góra location, for which RMSE decreased from 3.9 to 0.4 m/s. The results of an additional experimental simulation, which included roughness information from the original CLC 2006 datasets, clearly show the influence of incorrect input data on the model’s performance (Table 6). The difference between the calculated RMSEv is especially high in the case of Śnieżnik 2 (3.2 m/s).

Table 4 Mean error measures calculated for particular validation points within selected model’s resolutions and with an assumption of constant roughness length (above—for DTED elevation model; below—for GLS DEM)
Full size table
Table 5 Mean error measures calculated for particular validation points within selected model’s resolutions and with an assumption of roughness length derived from CLC 2006 dataset (above—for DTED elevation model; below—for GLS DEM)
Full size table
Table 6 Comparison of RMSEv errors calculated for selected validation points with an assumption of roughness length derived from the original CLC 2006 dataset and the corrected one (simulation used GLS DEM and was performed within 100 grid space)
Full size table

In some locations, relatively high error values tend to occur regardless of roughness length information (Śnieżnik 2, Jaworek, Idzików). This was probably the consequence of two tendencies of the model: the first to overestimate velocity in points placed at the highest altitudes (Śnieżnik 2) and the second to underestimate velocity in points placed at low altitudes or inside landforms with a spatial span exceeding the calculation domain range (Międzygórze 2, Jaworek, Idzików). In the second case a potential solution might be to expand the domain area, however this shall be a subject of further experiments.

The results of Student’s t test indicate that changing the particular parameters of Canyon’s setup may affect the results of the modelling with varied significance (Table 7). Within a 100 m resolution, the inclusion of the roughness information derived from CLC 2006 instead of a constant value of 0.05 m implied changes in the model’s performance detected by 11 of 15 validation points. The same change, but within the coarser (250 m) resolution, affected 10 points and the overall impact was weaker. On the other hand, changes of the domain resolution caused significant changes in nine points (CLC 2006 roughness) and only in seven points if the constant roughness value was assumed. Finally, the exchange of DEMs (GLS DEM to DTED), caused the least noticeable changes in the modelled wind fields (only five points were affected).

Table 7 Results (probability values) of the Student’s t test for paired samples
Full size table

The accuracy of the presented results may be considered as mediocre as compared with the results of other studies presented in the Introduction section. The achieved mean RMSEv values at best were no lower than 1.0 m/s. Nevertheless, some improvement was accomplished comparing the results of this study to the modelling conducted in test sites of Poiares and Trevim (Lopes 2003) (Table 8). However, it should be emphasized that the mentioned study was based on longer simulation periods, thus it presented rather an average model performance. Still, it is encouraging that in the mentioned studies input wind data were obtained from readings taken on masts higher than 5 m. In this case it is justified to expect smaller sensitivity to the effects of roughness length spatial variability and, in consequence, higher values of performance indices. However, it should be noted that the I v index is relative to the measured velocity; as such it should be used with caution as a basis for any comparisons.

Table 8 Comparison of Canyon CFD solver performance indices calculated for various test sites
Full size table

5 Conclusions

The presented case study shows how remote sensing data may be used in the process of diagnostic modelling of a wind velocity field. Under the assumed methodical approach (limited number of observations taken by handheld anemometers), a very careful selection of modelled time-periods is crucial for obtaining feasible results of the modelling. It should be emphasized that diagnostic models run well only during the occurrence of specific meteorological conditions (connected with strong and constant synoptic forcing of the air flow). Hence, all of further conclusions are limited to the aforementioned circumstances.

The results of the study clearly demonstrate the significance of roughness length information derived from remote sensing (in this case: CLC 2006) data in this type of modelling. The quality of the used data was high enough to achieve results comparable with other studies which present applications of diagnostic models. However, using raw CLC 2006 data is not recommended (in mountainous terrain), due to the aforementioned inaccuracies. Before generating roughness length information, where practicable, this data should be compared to higher resolution aerial photos or orthophotomaps.

Surface roughness resolution has a marked impact on the modelled wind fields. In case of the resolution of the model this impact seems to be rather weaker. However, this issue cannot be definitely settled based on short-term, empirical results. It may be assumed that increasing the domain resolution further may improve the accuracy of the results, although this should be tested. Another way to improve the performance of the model would be probably the application of roughness data derived directly from orthophotomaps. This, in consequence, may necessitate the use of a different approach in roughness length class assignment. Moreover, an experiment involving the data from re-analysis driven dynamic models should also be conducted to further improve the initial wind fields.

The presented method of providing the input wind velocity and wind direction data proved to be relatively useful, especially given the lack of data from other sources. In this case, the issue of defining measurement locations (the most representative in the considered model and roughness data resolution) seems to be still open and ready for further studies and discussion.