Evaluation of High-Resolution Land Cover Geographical Data for the WRF Model Simulations

Abstract

Increased computing power has made it possible to run simulations of the Weather Research and Forecasting (WRF) numerical model in high spatial resolution. However, running high-resolution simulations requires a higher-detail mapping of landforms, land use, and land cover. Often, higher-resolution data have limited coverage or availability. This paper presents the feasibility of using CORINE Land Cover (CLC) land use and land cover data and alternative high-resolution global coverage land use/land cover (LULC) data from Copernicus Global Land Service Land Cover Map (CGLS-LC100) V2.0 in high-resolution WRF simulations (100 × 100 m). Global LULC data with a resolution of 100 m are particularly relevant for areas not covered by CLC. This paper presents the method developed by the authors for reclassifying land cover data from CGLS-LC100 to MODIS land use classes with defined parameters in the WRF model and describes the procedure for their implementation into the model. The obtained simulation results of the basic meteorological parameters from the WRF simulation using CLC, CGLS-LC100 and default geographical data from MODIS were compared to observations from 13 meteorological stations in the Warsaw area. The research has indicated noticeable changes in the forecasts of temperature, relative humidity wind speed, and direction after using higher-resolution LULC data. The verification results show a significant difference in weather predictions in terms of CLC and CGLS-LC100 LULC data implementation. Due to the fact that better results were obtained for CLC simulations than for CGLS-LC100, it is suggested that CLC data are first used for simulations in numerical weather prediction models and to use CGLS-LC100 data when the area is outside of CLC coverage.

1. Introduction

The Earth’s surface, heated by the sun, reflects and transfers the absorbed solar radiation to both the deeper layers and the air. Thus, the air heats up from the surface of the Earth. However, the Earth does not heat evenly. Everyone has certainly noticed that on a bright warm day, stones and sand on the shore of a lake sometimes heat up so strongly that they become hot. In contrast, the water in the lake never reaches that temperature. Uncovered fields or slopes of hills facing the sun heat up much more strongly than a meadow covered with lush grass or a forest. Moreover, the air above these places will not be equally heated. Though it is hot in a concrete-covered city center, the air over a nearby green meadow is much cooler. The vegetation, soils, land use, presence of water bodies, and topography all influence heat flux exchanges between city surfaces and the overlying air. The proper description of the exchange of heat between a surface and the overlying atmosphere in terms of cities is critical for diurnal variation in boundary layer temperature, humidity, depth, and turbulence [1]. Studies of climate elements in urban areas prove that urbanization modifies almost all meteorological parameters in relation to their natural courses [2,3,4,5,6,7]. Recognizing the significant influence of locally occurring surfaces on the distribution of temperature and humidity in a city, the researchers decided to adapt land cover and land use in a WRF numerical weather prediction model [8,9,10] to the more recent 100 × 100 m resolution data and evaluate the impact of this implementation on the forecast verification of basic meteorological parameters. The team developing the Weather Research and Forecasting (WRF) model emphasizes that a more accurate consideration of the local conditions of the study area allows for the elimination of significant errors in forecasts [11]. The representation of the land use/land cover (LULC) data in the model was improved through the representation of the surface parameters (described in the model by such parameters as albedo, emissivity, roughness, soil heat capacity, surface thermal inertia, and soil moisture availability) so that more realistic results of atmospheric processes can be obtained [12]. The physical properties of the surface, as well as its color and degree of wetness, determine the value of its albedo, which is a measure of the reflectivity of a surface, i.e., the ability of a surface to reflect electromagnetic radiation, including solar radiation. The surface can also differ in terms of its emissivity, which is a measure of the ability of a surface to emit electromagnetic radiation in the infrared range. Surfaces with higher emissivity release more heat into the atmosphere [1]. The rate at which a surface reacts to changes in temperature is described by surface thermal inertia. Soil heat capacity affects the rate at which soil absorbs and releases heat. Soil moisture availability is a measure of water in the soil, and it contributes to the ability of heat storage and air humidity. Roughness length affects the interaction of heat, moisture, and momentum from the surface to the atmosphere [13].

The LULC data have a wide range of applications from monitoring land cover and land use changes, among others, to land use planning or tracking progress in sustainable development [14] by monitoring changes in forests, croplands, surface water, and urbanization. The LULC data can vary in resolution, coverage, compilation method, and validity [15]. These data can be obtained through various methods such as field surveys and satellite or aerial imageries and are then classified into an adopted category. These categories can vary considerably by region, which can result in inconsistencies between classifications. Each class can also differ in terms of its physical properties [16,17]. The result of satellite imagery classification in land cover and land use classes and its thematic accuracy can vary due to the method adopted for classification. The classification can be conducted entirely manually (visual interpretation) [18] or semi-automatic (i.e., digital image processing and machine learning) [19,20,21]. The current research shows that machine and deep learning methods are effectively used not only in LULC semi-automatic classification or deriving various land use attributes, but also in the successful prediction of their changes [15,19,21], which is important for sustainable development (resources managing), land use planning, and management decisions.

Due to the physical properties of LULC data that influence the formation of weather processes, these data are used to improve numerical weather prediction models. The authors of this paper, in their previously conducted research [22], implemented land use and land cover data from Corine Land Cover (CLC) [18,23,24,25] into the WRF model. However, the coverage of these data for Europe ends at the eastern border of Poland; therefore, in the research presented here, the authors tested the application of high-resolution global LULC data from Copernicus Global Land Service Land Cover 100 m (CGLS-LC100) product V2.0 [26,27,28] in WRF simulations. To the authors’ knowledge, no studies have yet been conducted on the impact of implementing high-resolution global land cover data at 100 m resolution on the results of a WRF model of the same resolution. The aim of this study is to evaluate the impact of the implementation of CLC and CGLS-LC100 data on the forecast results of basic meteorological parameters (temperature, relative humidity, wind speed and direction) to select a data source that achieves the highest forecast accuracy.

Improvements in forecast accuracy and studies of the impact of LULC changes on the forecast of basic meteorological parameters have forced researchers to increase the resolution of the model’s computational grids. The research has shown that merely increasing the resolution of the model grids had a positive effect on forecast results [22,29]. In the WRF model, 30 arc sec. resolution data can be used as default LULC data (900–1000 m) from the USGS or MODIS data [10]. The USGS LULC data were developed by the United States Geological Survey, and are based on high-resolution global imagery from April 1992 to March 1993 [20]. MODIS data, on the other hand, were developed by the University of Boston based on global Moderate Resolution Imaging Spectroradiometer data from January to December 2001 and are divided into 21 categories [16,30]. CLC data are most often used in research on the implementation of alternative LULC sources in the simulation of numerical weather prediction models for the European region [31,32,33,34,35,36,37,38]. Research related to Western Europe shows that meteorological variables differ significantly between WRF simulations depending on the land cover dataset [36]. The results of studies for northern Italy [31] show the sensitivity of mean wind speed forecasts (at 10 m), temperature (at 2 m), and the detection probability of precipitation events in the WRF simulations on two different sets of LULC data: CLC with SRTM and USGS data. That study shows that for CLC and SRTM simulations, wind speed forecasts are more consistent with observations due to the fact that CLC and SRTM data lower the wind speed by introducing more friction and higher roughness in the simulation domain. Furthermore, the probability of precipitation detection was higher and more consistent with observations. Due to the CLC implementation, the sensible and latent heat flux over the simulation domain changed and influenced temperature forecasts, which caused its overestimation. Studies focusing on different LULC data for Berlin [38] and two Austrian regions [39] have shown that due to the high heterogeneity of LULC data in urban areas, higher-resolution LULC data simulations perform better in air temperature forecasts.

In studies on the results of WRF simulations using default and CLC data, the authors compared the results of default simulations from the USGS to CLC simulations [31,32,33,34,35,36,37,38]. It is worth noting that since the WRF model version 3.8 was released in 2016, more recent MODIS data have been included as default LULC data in the WRF model [40]. According to the paper’s authors, the USGS LULC database is quite outdated, especially for simulations in urban or suburban areas where there has been significant urban development in the last 30 years. Therefore, comparing the results of WRF model simulation with USGS LULC data to simulations with high-resolution LULC data from CLC improves the results of the model more than the case when comparing the outputs of the WRF model simulations received on MODIS LULC data to those received on LULC data from CLC.

Most articles on the influence of local conditions on NWP (Numerical Weather Prediction) results focus on the effect of changing LULC on basic parameters such as temperature [38,41,42] and humidity. The subject of wind speed and direction analysis has been addressed less frequently [31,43,44,45,46,47]. The forecast of wind direction is largely dependent on the conditions of placement of the meteorological station. Theoretically, a given meteorological station can meet the WMO’s requirements for the location of a meteorological station [48], but the corridors that aerate the city can significantly disturb the distribution of wind and the data obtained from a mesoscale model, with a resolution of 4 to 1 km, will be highly generalized. Even though forecasting the surface wind speed and direction in an urban area can be quite problematic, there have been recent articles whose authors have focused on analyzing the impact of changing the LULC on the forecast of even more complex phenomena, such as convective phenomena [37]. The influence of LULC on the prediction of the distribution and concentration of urban air pollutants is also not insignificant [29,49].

As reported by [13], the use of alternative global LULC data in the form of Global Land Cover (GLC2015) data [50] with a spatial resolution of 300 m for the Xinjiang region in China helped to improve the results of basic meteorological parameters. The simulation results using the alternative data were compared to those obtained from the simulations using the default data from the USGS. Furthermore, the simulation results were compared with observations from meteorological stations. In this paper, the following meteorological parameters were studied for the summer period: air temperature at 2 m, ground temperature, wind speed, and humidity. The obtained verification results of the WRF model forecast of the basic meteorological parameters indicate that the representation of the coverage and use of the study area is better in the GLC2015 data than in the USGS data [13].

In this article, the authors described the implementation of LULC CGLS-LC100 data into the WRF model according to MODIS classes. Moreover, the impact of using LULC LC100 data from CGLS and CLC on the results of 100 m resolution simulations for the Warsaw region was evaluated. Simulations using the above high-resolution data were compared with observations and simulation results obtained using default LULC data from MODIS.

2. Materials and Methods

2.1. Geographical Data and Methodology

The authors used data from CLC 2018 [18,23,24,25] and CGLS-LC100 V2.0 [26,27,28] to analyze the impact of implementing high-resolution geographic data on LULC into the WRF numerical weather prediction model. To perform simulations with high-resolution LULC data (CLC, CGLS-LC100), high-resolution terrain topography data from the Shuttle Radar Topography Mission (SRTM) were used [51,52]. The results of the high-resolution simulations were compared with observations from weather stations and contrasted with the results obtained in simulations with default data, i.e., LULC from MODIS (the modified IGBP 21-category, 30 arc-seconds) and GMTED2010 topography (The USGS Global Multi-resolution Terrain Elevation Data with a horizontal resolution of 30 arc-seconds) [53].

To implement the CGLS-LC100 data into the WRF model, the authors proposed a method for reclassifying the CGLS-LC100 data into MODIS classes defined in the WRF model. The process of implementing data on LULC into the WRF model can be divided into three stages, as shown in Figure 1. These stages for the implementation of CGLS-LC100 data were as follows: the first stage involved the acquisition of global data on LULCs from CGLS; the second stage consisted of the reclassification of CGLS-LC100 data into MODIS classes; and, finally, the third stage involved the conversion of raster data into a binary format supported by the WRF model. For stage two of the implementation of high-resolution global CGLS-LC100 data into the WRF model, the authors developed a method to reclassify LULC classes from CGLS-LC100 to classes defined in the WRF model. In the WRF model, individual LULC classes have defined albedo, emissivity, roughness length, soil heat capacity, thermal inertia, and soil moisture [17].

These parameters are defined for USGS classes and MODIS classes. CGLS-LC100 data are the CGLS Land Cover Map at 100 m data based on the PROBA-V satellite for epoch 2015 with 23 classes (22 discrete classes and 1 class for areas wherein no input data are available). The authors proposed their own reclassification method of GLC CGLS 23 classes defined according to the Land Cover Classification System (LCCS) scheme of the Food and Agriculture Organization [54,55] into 21 MODIS IGBP [16] classes (21 discrete classes and 1 class of unassigned areas) based on the relationship of the overlap between the analyzed MODIS and GLC CGLS classes. The process of the reclassification of the CGLS-LC100 classes to MODIS classes is presented in

Table 1. Reclassification procedure of CGLS-LC100 classes to MODIS classes.

LC-100 Code	Land Cover Class (CGLS-LC100)	MODIS Code	MODIS Class
111	Closed forest, evergreen needle leaf	1	Evergreen Needleleaf Forest
112	Closed forest, evergreen broad leaf	2	Evergreen Broadleaf Forest
113	Closed forest, deciduous needle leaf	3	Deciduous Needleleaf Forest
114	Closed forest, deciduous broad leaf	4	Deciduous Broadleaf Forest
115	Closed forest, mixed	5	Mixed Forest
116	Closed forest, unknown	3	Deciduous Needleleaf Forest
121	Open forest, evergreen needle leaf	1	Evergreen Needleleaf Forest
122	Open forest, evergreen broad leaf	2	Evergreen Broadleaf Forest
123	Open forest, deciduous needle leaf	3	Deciduous Needleleaf Forest
124	Open forest, deciduous broad leaf	4	Deciduous Broadleaf Forest
125	Open forest, mixed	5	Mixed Forest
126	Open forest, unknown	3	Deciduous Needleleaf Forest
20	Shrubs	7	Closed Shrublands
30	Herbaceous vegetation	10	Grasslands
40	Cultivated and managed vegetation/agriculture (cropland)	12	Croplands
50	Urban/built up	13	Urban and Build-Up
60	Bare/sparse vegetation	16	Barren or Sparsely Vegetated
70	Snow and Ice	15	Snow and Ice
80	Permanent water bodies	21	Lakes
90	Herbaceous wetland	11	Permanent Wetlands
100	Moss and lichen	16	Barren or Sparsely Vegetated
200	Open Sea	17	Water Bodies
0	No input data available	255	Unclassified

.

To implement stage three, the reclassified LULC CGLS-LC100 data were transformed into binary data [46]. The data format was changed using GDALL [12]. In the WRF model, the geogrid data concerning LULC are processed in the WRF Processing System (WPS). In the WPS, the geogrid module interpolates the geographic data to the resolution of the computational domains. The nearest neighbor method was used to interpolate the CGLS-LC100 categorical data of LULC classes to the model grids. To initialize the WRF model data calculation with CGLS-LC100, CLC, or default data, it is necessary to declare this at the WPS list. The order of the declaration of the geographical data to be used by the WPS is important because when the area is out of the coverage of the first set of data, the model is using the second one, or if the areas are adjacent, the module completes the missing area with the data from the second set [22]. The process of implementing CLC data into the WRF model was presented in a previous article [22] by the authors. The visualization of the overclassified LULC CLC and CGLS-LC100 classes to MODIS classes is shown in Figure 2. It is worth highlighting that different classification methods were used to compile those LULC data from the satellite data. CLC data are based on manual classification, and CGLS-LC100 data are based on pixel-based (spectral) classification. The thematic classification accuracy of CLC data is above 85%, and CGLS-LC100 is above 80%. The area of Figure 2 includes the area of the city of Warsaw, which was taken as the study area.

For the analyzed synoptic situations, a detailed analysis of the forecast errors at individual measuring stations and aggregate analyses of the errors for individual days were carried out. The values of the average forecast errors for all analyzed synoptic situations and meteorological measuring stations were determined. The data extraction and calculation of the verification indices characterizing the quality of the forecasts were performed in Python language [56].

To verify the quantitative meteorological forecasts, the following statistics were calculated: mean error (ME), mean absolute error (MAE), root-mean-square error (RMSE), mean squared error (MSE), bias error (BIAS), Pearson correlation coefficient (R), and normalized root-mean-square error (nRMSE) (Table 2). To verify the results of the numerical modelling of weather forecasts, hourly data from a network of road meteorological stations [49] and airport meteorological stations within Warsaw were used (Figure 3).

Table 2. Definition of the statistical indicators [57,58,59].

Indicators	Definition
Mean error (ME)	ME = $\frac{\sum _{i=1}^n{(f}_i{-o}_i)}{n}$
Mean absolute error (MAE)	MAE = $\frac{\sum _{i=1}^n{|f}_i{-o}_i|}{n}$
Root-mean-square error (RMSE)	RMSE = $\sqrt{\frac{\sum _{i=1}^n{(f_i-o_i)}^2}{n}}$
Mean squared error (MSE)	MSE = $\frac{\sum _{i=1}^n{{(f}_i{-o}_i)}^2}{n}$
Pearson correlation coefficient (R)	R = $\frac{\sum _{i=1}^n(f_i-\stackrel{-}{f})(o_i-\stackrel{-}{o})}{\sqrt{\sum _{i=1}^n{(f_i-\stackrel{-}{f})}^2}\sqrt{\sum _{i-1}^n{(o_i-\stackrel{-}{o})}^2}}$
Bias error (BIAS)	BIAS = $\frac{\frac{1}{n}\sum _{i=1}^n{f}_i}{\frac{1}{n}\sum _{i=1}^n{o}_i}$
Normalized root-mean-square error (nRMSE)	nRMSE = $\frac{RMSE\ast 100\%}{\frac{1}{n}\sum _{i=1}^n{o}_i}$

Table 2. Definition of the statistical indicators [57,58,59].

Indicators	Definition
Mean error (ME)	ME = $\frac{\sum _{i=1}^n{(f}_i{-o}_i)}{n}$
Mean absolute error (MAE)	MAE = $\frac{\sum _{i=1}^n{|f}_i{-o}_i|}{n}$
Root-mean-square error (RMSE)	RMSE = $\sqrt{\frac{\sum _{i=1}^n{(f_i-o_i)}^2}{n}}$
Mean squared error (MSE)	MSE = $\frac{\sum _{i=1}^n{{(f}_i{-o}_i)}^2}{n}$
Pearson correlation coefficient (R)	R = $\frac{\sum _{i=1}^n(f_i-\stackrel{-}{f})(o_i-\stackrel{-}{o})}{\sqrt{\sum _{i=1}^n{(f_i-\stackrel{-}{f})}^2}\sqrt{\sum _{i-1}^n{(o_i-\stackrel{-}{o})}^2}}$
Bias error (BIAS)	BIAS = $\frac{\frac{1}{n}\sum _{i=1}^n{f}_i}{\frac{1}{n}\sum _{i=1}^n{o}_i}$
Normalized root-mean-square error (nRMSE)	nRMSE = $\frac{RMSE\ast 100\%}{\frac{1}{n}\sum _{i=1}^n{o}_i}$

Figure 3. Network of meteorological stations of the City Cleaning Board within the city of Warsaw accepted for the study (in black circles) and airport meteorological stations (based on reference [60]).

Figure 3. Network of meteorological stations of the City Cleaning Board within the city of Warsaw accepted for the study (in black circles) and airport meteorological stations (based on reference [60]).

The network of road meteorological stations consists of automatic meteorological measurement stations located in the strip of national roads, which allow the forecasting of atmospheric phenomena occurring in the road area, especially the phenomenon of glaze, to support seasonal road maintenance decision-making processes. The number of stations installed in each road section depends strictly on the microclimatic conditions of the road sections by maintaining, as far as practical, standards of representativeness of the meteorological measurements; hence, they characterize local-to-microscale conditions [61,62]. The network of measuring stations in the Warsaw agglomeration consists of eighteen measuring stations, as shown in

Table 3. Coordinates list of locations of automatic weather stations of the City Cleaning Board.

No.	Station Name	Coordinates	No.	Station Name	Coordinates
1	Bielany	52°17′32.675″N 20°58′22.367″E	10	Most Polnocny	52°20′7.439″N 20°57′41.219″E
2	Czecha	52°13′32.231″N 21°9′57.347″E	11	Plochocinska	52°18′23.88″N 20°57′05.84″E
3	Drewny	52°7′38.892″N 21°5′48.407″E	12	Pulawska	52°19′38.963″N 21°1′45.227″E
4	Izbicka	52°10′45.551″N 21°13′58.368″E	13	Radzyminska	52°9′12.6″N 21°1′5.231″E
5	Jelonki	52°13′6.995″N 20°54′33.912″E	14	Siekierkowska	52°16′45.797″N 21°4′9.691″E
6	Jerozolimskie	52°12′14.615″N 20°56′26.555″E	15	Torwar	52°12′9.252″N 21°3′57.887″E
7	Krakowska	52°10′30.719″N 20°56′39.803″E	16	Wilanow	52°13′23.412″N 21°2′39.623″E
8	Lazurowa	52°14′46.5″N 20°53′52.044″E	17	ZOO	52°15′14.147″N 21°1′19.596″E
9	Modlinska	52°20′7.439″N 20°57′41.219″E	18	Zolnierska	52°15′31.56″N 21°8′32.885″E

. Because not all measuring stations carry out measurements of wind speed and direction, for this study, we used eleven stations, from which measurements of all four analyzed meteorological parameters were available. Other data used to verify the results of numerical modeling of weather forecasts were obtained from airport meteorological stations Okecie (EPWA) with coordinates: 52°9′46″N, 20°57′40″E [49] and Babice (EPBC) with coordinates: 52°16′0″N, 20°55′0″E.

2.2. Study Area and Synoptic Situations

The area of the capital city of Warsaw was taken as the study area. This city is located in the central-eastern part of Poland (Figure 4). Built-up and urbanized area account for almost 60% of the total city’s area. This influences the formation of conditions characteristic for urban weather in Warsaw. Among other things, ten ventilation corridors of the city are observed in the Warsaw area. The impact of anthropogenic factors is important in weather modeling. Urban forecasting requires increasing the computational resolution of the model and, thus, also provides appropriate geographic data that will be used by the model. As for the larger computational domains of the WRF model for Warsaw, their area often includes the non-CLC area as well.

For the area of Poland, CLC data are available, yet, unfortunately, their coverage ends at the eastern border of Poland. This motivated the article’s authors to research the use of alternative sources of high-resolution CGLS-LC100 data, whose coverage also includes the area in the east of Poland.

For the analysis, a random period was selected, during which there was eastern advection in Warsaw conditioned by the distribution of baric systems (Figure 5). This was the period from 22 June 2020, 00:00 UTC, to 27 June 2020, 00:00 UTC. During the first part of the term (22–24 June 2020), the weather in Warsaw was shaped between the eastern part of the anticyclone over France and the bordering rear part of the cyclone over the Black Sea. The occurring unstable air mass, additionally aided by convergent airflow, favored the formation of a convergence line (22 June 2020). During the day, Cu and Cb clouds were present over Warsaw.

In the following days (23–24 June 2020), the formation of a cold front of an occluding system could be observed, along which thunderstorms commonly occur in summer. In the filling gulfs and away from the center of the low, the cloud systems of the occluded fronts have less vertical extension and are strongly stratified. It is worth noting that this front was associated with an upper low whose center moved from the Baltic Sea over western Poland to the Czech Republic.

In the second part of the analyzed period (25–27 June 2022), the high-pressure center from Western Europe moved over the northern Baltic Sea and farther through the Baltic countries to western Russia. During this period, Warsaw was under the influence of the western edge of the mentioned anticyclone. In summer, at the western edge of the anticyclone with high temperature, humidity, and thermodynamic instability, storms often occur in the formation of convergence lines. Furthermore, in diffuse pressure fields and on the western edges of inactive anticyclones, when the air mass in the ground layer heats up, and the humidity of this mass increases, with significant air mass instability, when there are conditions for strong daytime heating of the ground air layer, and high humidity persisting at an altitude of several kilometers and the absence of inversion and isothermal layers, thunderstorms form in homogeneous air masses.

With reference to the previously mentioned eastern inflow over the Warsaw area, it is worth noting that from the practical use of numerical weather forecast models by the synoptics, it appears that with eastern inflows over the Polish area, the models are characterized by lower forecasting capabilities than with western inflows, which may be caused, among other reasons, by a smaller number of observations and radiosonde data from Eastern Europe. It is worth noting that the formation of convergence lines is the most surprising weather pattern and is generally not well forecasted. In general, convergence lines are not always predicted in numerical weather forecast models, especially those with low spatial resolutions. The local factors that favor convergence lines (landforms, local pressure differences, small changes in atmospheric conditions at different altitudes and sea breezes) are difficult to account for in numerical models, leading to imperfections in their forecasting. Forecasting convergence lines requires synopticians to keep a close eye on the development of atmospheric processes to be able to warn the concerned institutions in time about the possibility of their occurrence because if there are thunderstorms on the convergence line, then it is called a squall line.

It is interesting to note that on 26 June 2020 (Figure 5 and Figure 6), there was a line of convergence over Warsaw, which was not forecasted by the WRF model. Therefore, the model results deviated significantly from the actual conditions. According to the model, the daytime temperature was expected to have a normal diurnal distribution, and the observational data show that there was a pronounced drop in temperature after the convergence line passed. Squall lines are narrow zones with thunderstorms, precipitation, and squalls with characteristics like those of cold fronts, but which are local and short-lived results of thunderstorm activity. They often form in the warm and unstable sectors of low-pressure systems before cold fronts or on the edge of high-pressure systems where incoming cold air pushes warmer air upwards, causing convective phenomena to develop. In other words, the term describes an unstable air mass characterized by increased turbulence and the development of convection with corresponding cloudiness.

2.3. WRF Model

The WRF model is a non-hydrostatic mesoscale weather forecast model. The model has been designed and developed continuously for more than 22 years, among others, by the National Oceanic and Atmospheric Administration (NOAA), National Centers for Environmental Prediction (NCEP), National Center for Atmospheric Research (NCAR), Air Force Weather Agency (AFWA), Forecast Systems Laboratory (FSL), Federal Aviation Administration (FAA), and the United States Air Forces (USAF) [8]. The WRF model output accuracy depends on the accuracy of the input data, the topography of the terrain, geographical data, the quality of the parametrization of the physical processes it models, the model version, and synoptic situations. For the purposes of this study, WRF model version 4.2.1 was used. The WRF model set-up for the conducted studies is presented in

Table 4. The WRF model configuration used in the experiment for computational domains (D).

Parameters	D1	D2	D3
Spatial resolution	2.5 × 2.5 km	500 × 500 m	100 × 100 m
Surface layer scheme	Revised MM5	Revised MM5	Revised MM5
Planetary boundary layer (PBL) scheme	YSU	none	none
Radiation scheme	RRTMG	RRTMG	RRTMG
Microphysics	WSM6	WSM6	WSM6
Cumulus	none	none	none
Land surface	Noah	Noah	Noah
Topographic datasets	SRTM/GMT2010	SRTM/GMT2010	SRTM/GMT2010
Land use	CLC/CGLS-LC100/MODIS	CLC/CGLS-LC100/MODIS	CLC/CGLS-LC100/MODIS

. The WRF model was run on the data (first approximation fields and boundary conditions) of the superior Global Forecast System (GFS) model with a resolution of 0.25° (~28 km) [64]. Due to the higher spatial resolution of the regional and mesoscale WRF simulations compared to the GFS, downscaling was carried out using multiple nesting.

It is worth noting that the spatial resolution of the global model is sufficient to provide data for regional, mesoscale, and even microscale models. For comparison, synoptic stations in Poland are spaced approximately every 50 km. The spatial resolution of the model used for the simulation was 2.5 km for the first domain (D1), 500 m for the second (D2), and 100 m for the third (D3). The forecasts for each day started at 00:00 UTC, the spin-up time was 6 h, and the forecast time was 24 h with one-hour output data. The selection of parametrization schemes of the WRF physical processes was conducted based on the studies for the Central European area [64,65] and other recent studies related to this subject [66,67]. Based on that, the revised MM5 Monin–Obukhov surface layer scheme [9,65,66,68] was used in this study (for all domains). For planetary boundary layer (PBL) parameterization, the Yonsei University Scheme (YSU) [9,64,65,69] was applied for the first domain. In the second and third domain, explicit vertical diffusion was used directly (non-PBL parameterization) [9,65,70,71,72]. For those domains, the horizontal and vertical eddy viscosities were determined using 3D Smagorinsky turbulence [9]. For all domains, microphysics parametrization WSM6 [7] and RRTMG longwave and shortwave radiation schemes [8] were used.

3. Results

As a result of simulating the WRF model in three variants, the distributions of basic meteorological parameters (temperature (Figure 6), relative humidity, wind direction and speed) were obtained.

The results were compared with the observations and then, to determine the quality of the forecasts of the meteorological parameters, they were verified based on the values of statistical indicators (Table 2).

To show that not only the terrain but also land cover influence the wind simulation, the example results of different geographical data setups for the third domains are summarized in Figure 7. The forecast of wind speed and direction is influenced, as are the forecasts of the other elements analyzed, by both the topography of the terrain and its cover. Land cover roughness and heat capacity significantly influence wind formation, distribution, and intensity.

The analysis of the forecast error values of individual parameters was divided into the analysis of forecast errors for individual days for all analyzed stations and the analysis of errors at a given measuring station for the complete study period. The value of the forecasting errors of individual parameters averaged over stations and days was also calculated (

Table 5. Average statistical evaluation of the simulations error values: ME, MAE, RMSE, MSE for five-day (22–27 June 2020) 24 h simulations for all stations (13 stations: 11 automatic and 2 airport stations) with 6 h spin-up time.

Average Statistics	Model Simulation	Temperature (°C)	Relative Humidity (%)	Wind Speed (m/s)	Wind Direction (°)
	Default	0.98	−11.29	2.15	36.46
ME	CLC	0.79	−9.75	2.01	21.79
	CGLS	0.85	−10.31	2.19	31.52
	Default	1.80	15.07	2.28	80.74
MAE	CLC	1.65	13.97	2.22	76.43
	CGLS	1.68	14.09	2.35	76.20
	Default	2.28	17.51	2.39	99.12
RMSE	CLC	2.11	16.38	2.35	90.99
	CGLS	2.14	16.51	2.46	92.49
	Default	6.03	368.93	6.57	12,415.11
MSE	CLC	5.21	324.56	6.31	10,047.97
	CGLS	5.33	331.32	6.93	10,831.89

and

Table 6. Average statistical verification results (Pearson’s correlation coefficient (R), BIAS, and nRMSE) of the simulations for five-day (22–27 June 2020) 24 h simulations for all stations (13 stations: 11 automatic and 2 airport stations) with 6 h spin-up time.

Average Statistics	Model Simulation	Temperature	Relative Humidity	Wind Speed	Wind Direction
	Default	0.81	0.63	0.46	0.23
R	CLC	0.82	0.63	0.36	0.25
	CGLS	0.82	0.64	0.52	0.23
	Default	1.05	0.87	2.52	1.36
BIAS	CLC	1.04	0.89	2.47	1.27
	CGLS	1.02	0.88	2.56	1.33
	Default	10.58	22.75	105.54	67.72
nRMSE (%)	CLC	9.77	21.28	103.67	62.17
	CGLS	9.91	21.46	108.45	63.19

). Analyses of the values of the temperature forecast errors for individual days average over stations (Figure 8) showed that in 80% of the situations studied, the use of CLC or CGLS-LC100 data improved the obtained results. This improvement was greater for simulations using CLC data than those using CGLS-LC100.

Analysis of the RMSE values for the individual measurement stations (Figure 9) averaged over days showed the greatest improvement in temperature forecast results when using CLC data (the matrix mesh near ideal value has lighter color). The improvement over the simulation in the default configuration also occurred when CGLS-LC100 data were used. The analysis of BIAS values for individual measuring stations showed that in only at two stations (Plochocinska and Wilanow) of the thirteen analyzed stations, the value for the default simulation was better than for the CLC or CGLS-LC100 simulations. The value of the Pearson correlation coefficient (R) of temperature forecast, an average over stations and days, was the best for the simulation on CLC data, and it was 0.82 (Table 4).

The verification of the relative humidity forecasts for the synoptic situations analyzed revealed (Table 6, Figure 10) that the near-ideal value of forecasts, as determined by Pearson’s correlation coefficient (R), was obtained by the model simulations using CLC. BIAS indicated higher agreement between forecasts and observations for CLC simulations. Moreover, ME, MAE, RMSE, and MSE showed the highest agreement with observations for the simulations plotted using CLC. For 80% of the analyzed cases, the use of CLC or CGLS-LC100 data effectively improved relative humidity forecasts.

The verification of the relative humidity forecasts at individual gauging stations (Figure 11) showed that at 69% of stations, the use of CLC or CGLS-LC100 data improved the obtained weather forecast results compared to simulations using default geographical data. Despite the fact that the average value (for all surveyed stations and for the whole analyzed five-day period of the simulation) of the RMSE for relative humidity forecast was better (the lowest) for simulations using CLC when analyzing the average value of the correlation coefficient (R), the best result was achieved for simulations on CGLS-LC100 data, and it was 0.64 (Table 6).

The verification of the wind speed forecasts for the examined synoptic situations in the three different variants showed that the values of the mean errors (ME, MAE, MSE, and RMSE) were the lowest for the simulation using the CLC data (Figure 12). BIAS also showed the highest compliance for CLC simulations (Table 6). For 60% of the analyzed cases, the use of CLC data enhances the capability of receiving more accurate wind speed forecasts.

A detailed analysis of the errors at individual stations showed that for 54% of the stations studied, the BIAS value for CLC and CGLS-LC100 simulations was closer to the ideal value than for simulations of the default parameters (Figure 13). The value of the Pearson correlation coefficient (R) of the wind speed forecast was the best for the simulation on CGLS-LC100 data, and it was 0.52 (Table 6).

Analysis of wind direction forecasts for the examined synoptic situations showed that the value of mean error (ME), mean square error (MSE), and BIAS reached the most desirable value for simulations using CLC (Table 5 and Table 6, Figure 14). Furthermore, simulations using CGLS-LC100 showed higher accuracy in terms of wind direction forecasts than simulations on default data.

A detailed analysis of the errors at individual stations showed that for 77% of the stations studied, the BIAS value for CLC and CGLS-LC100 simulations was closer to the ideal value than for simulations on default parameters (Figure 15). The value of the Pearson correlation coefficient (R) of the wind speed forecast was the best for the simulation on CGLS-LC100 data, and it was 0.25 (Table 6). RMSE and R were more sensitive to large forecast errors, so the errors shown for several stations outweighed the final result of the statistical analyses for all weather stations (Figure 15).

Table 5 and Table 6 show the average forecast errors of forecast parameters for the whole analyzed period and all of the analyzed meteorological stations (road network and airport stations). The average forecast error in this study is described as the error average for stations and days.

Table 7. Average statistical evaluation of error values: ME, MAE, RMSE, MSE for five-day (22–27 June 2020) 24 h simulations for two airport stations (spin-up time, 6 h).

Average Statistics	Model Simulation	Temperature (°C)	Relative Humidity (%)	Wind Speed (m/s)	Wind Direction (°)
	Default	1.14	−8.31	0.14	3.00
ME	CLC	0.93	−6.78	−0.16	−19.00
	CGLS	0.94	−7.14	0.20	−1.60
	Default	1.72	12.71	0.62	30.60
MAE	CLC	1.64	12.00	0.68	32.00
	CGLS	1.65	12.02	0.74	33.00
	Default	2.17	14.99	0.79	39.66
RMSE	CLC	2.07	13.95	0.92	43.09
	CGLS	2.08	14.07	0.93	40.63
	Default	5.52	264.35	0.62	1611.40
MSE	CLC	4.91	219.66	0.85	1931.20
	CGLS	4.97	223.48	0.90	1653.00

and

Table 8. Average statistical verification results (Pearson’s correlation coefficient, BIAS, and nRMSE) of the simulations for five-day (22–27 June 2020) 24 h simulations for all stations (13 stations: 11 automatic and 2 airport stations) with 6 h spin-up time.

Average Statistics	Model Simulation	Temperature	Relative Humidity	Wind Speed	Wind Direction
	Default	0.79	0.62	0.47	0.94
R	CLC	0.80	0.64	0.35	0.88
	CGLS	0.79	0.64	0.52	0.92
	Default	1.06	0.90	1.04	1.02
BIAS	CLC	1.05	0.92	0.98	0.90
	CGLS	1.05	0.92	1.06	0.99
	Default	10.24	19.93	18.33	21.29
nRMSE (%)	CLC	9.75	18.54	21.36	23.13
	CGLS	9.82	18.70	21.55	21.81

present the verification results (average for stations and days) exclusively for airport meteorological stations, which meet the WMO requirements for location and, thus, should yield more objective analysis results.

For temperature and relative humidity forecasts for both the airport weather stations and all of the analyzed weather stations, the smallest average forecast errors (RMSE, ME, MAE, and MSE) were obtained by the CLC simulations. The errors in the CGLS-LC100 simulations were slightly higher. Simulations using default data introduced the worst results. Furthermore, the BIAS and R-factor values showed the highest accuracy in terms of temperature and relative humidity forecasts for simulations for airport stations and temperature forecasts for simulations for all of the analyzed weather stations using CLC (Table 6 and Table 8). The average value of the temperature forecast improvement for simulations using CLC data compared to the default data was 0.19 °C. The average value of the difference between the observed and modeled temperatures was 0.79 °C in the CLC simulation, 0.85 °C in the CGLS-LC100 simulation, and 0.98 °C in the simulation using the default data (Table 5). According to ICAO annex 3, the temperature for aviation purposes should be predicted with an accuracy of ±1 °C. Therefore, even such a seemingly small improvement in model forecast capability is important for synoptics. The statistical verification of the predicted parameters showed (Table 6) that for the 2 m temperature prediction, the use of CLC data allowed the nRMSE to be decreased by 0.81% and by 0.67% when using CGLS-LC100 data compared to the results obtained in default geographical data simulations.

The greatest improvement in the average relative humidity forecast value was recorded for the CLC simulation. For the three simulation variants analyzed, the average relative humidity forecast error for the CLC simulations compared to the default simulations decreased by 1.54% (Table 5). For relative humidity forecasts, the decrease in nRMSE for the simulations using CLC was 1.47% and 1.29% for CGLS-LC100 compared to the default simulations (Table 6). The value of the average difference between the observed and modeled relative humidity was −9.75% for the CLC simulation, −10.31% for the CGLS-LC100 data, and −11.29% for the default data (Table 5). It is notable that for the analyzed period, the model overestimated the average value of the forecast temperature by about 0.8 °C to 1.1 °C (Table 5 and Table 7). As for relative humidity, it underestimated its value by 6.8% to 11.3% (Table 6 and Table 8). These significant errors may have been caused by the presence of squall lines during the analyzed period, which are not always properly forecasted by the WRF model (Figure 6).

As far as the wind speed analyses for airport stations and all of the analyzed meteorological stations are concerned, the lowest value of the mean forecast error (ME) was obtained by the CLC simulation, a slightly higher error was obtained by the simulation using default data, and the worst result was noted for the simulation using CGLS-LC100 (Table 5 and Table 7). An increase of 2.90% in nRMSE was recorded for the wind speed simulation using CGLS-LC100 data compared to the simulation using default data (Table 6 and Table 8). The value of the improvement in wind speed forecasts was below the acceptable wind speed measurement error (0.5 m/s). According to ICAO annex 3, the wind speed reading required for aviation should be predicted with an accuracy of ±2.5 m/s and a wind direction of ±20°. The average value of the difference in wind speed measured to result from the model was 2.15 m/s for the simulation on the default data, 2.01 m/s for the CLC simulation, and 2.19 m/s for the CGLS-LC100 simulation. The mean value of the improvement after applying the CLC data was 0.14 m/s (Table 5 and Table 6). It is also worth noting that in the case of the wind speed forecast, the value of the average forecast error (for stations and days) was about 2 m/s higher for stations from the network of road meteorological stations (Table 5 and Table 7) in comparison with the values obtained for airport meteorological stations (or stations situated in open areas, such as Most Polnocny). This may be caused by the failure to maintain a 30 m separation of the measuring station from terrain obstacles.

As for the wind direction analysis for all analyzed meteorological stations, the lowest mean forecast error (for stations and days) was obtained for simulations with CLC. Moreover, for the simulations created with CGLS-LC100, lower values of the mean error of the wind direction forecast were obtained than for the simulations on the default data.

For wind direction, a decrease in nRMSE of 5.55% was recorded for the simulation using CLC and a decrease of 4.53% for the CGLS-LC100 simulation compared to the simulation using default data (Table 8). The average difference (for all stations and days) between the wind direction calculated by the model and measured was 36.46° for the default simulations, 21.79° for the CLC simulation, and 31.52° for the CGLS-LC100 simulation (Table 5). The use of CLC simulations resulted in an average improvement of 14.68°.

4. Discussion

The correctness of forecasting atmospheric phenomena is determined by the accuracy of the basic meteorological parameters forecast by the numerical model; therefore, it is important to improve the accuracy of forecasting basic meteorological parameters even slightly, for example, by updating the geographical data (LULC, topography) used by numerical weather prediction models. Measurements, as well as the prognostic values of meteorological parameters in numerical weather prediction models, are used to develop forecasts of atmospheric phenomena (e.g., glaze or fog) or prognoses of the amount of energy obtained from renewable sources. WMO requirements [50] for the accuracy of measurements of individual meteorological parameters indicate the need to provide highly accurate measurements of individual meteorological parameters for both aviation and synoptic meteorology. Based on that, the accuracy of the meteorological measurements used for synoptic, aviation, or agrometeorological purposes should be as follows: temperature, ±0.1 °C; relative humidity, ±1%; wind speed, ±0.5 m/s; and wind direction, ±5°. According to this article’s authors, those values could be equated with the accuracy requirements of an ideal weather forecast model. According to the received values of forecast errors, it could be seen that the received values of the prognosed parameters meet the ICAO requirements for an operationally desirable accuracy of forecasts, which should be as follows: temperature, ±1.0 °C; wind speed, ±2.5 m/s; and wind direction, ±20°, for at least 80% of cases.

Based on conducted studies, it can be seen that the implementation of higher-resolution geographical data has allowed for increased forecast accuracy for Warsaw. The implementation of higher-resolution and more up-to-date LULC data resulted in receiving more consistent values in terms of temperature and relative humidity forecasts. The greatest improvement was received in terms of CLC data implementation. For CGLS-LC100 data implementation, improvements also occurred compared to the default simulations. The results obtained after the implementation of the CLC data did not confirm the temperature overestimation effect observed in Northern Italy [31], but confirmed the results received after global LULC data implementation for Xinjang in northwest China [13]. The improvement in temperature forecast after CLC data implementation to the WRF model was also confirmed for Berlin [38] and Austrian regions [39]. The use of USGS data as a default for studies concluded that for Berlin, the temperature is underestimated [38], whereas for the use of MODIS data as a default, there was overestimation in temperature for Warsaw. In addition, for the analyzed period and the stations studied, the forecasted temperature value after the implementation of the high-resolution LULC data showed a decrease, which is caused by a better representation of small green areas in the city.

For wind speed forecasts after higher-resolution geographical data implementation for the analyzed areas, a reduction in wind speed was visible. The wind speed decrease was not as high as noted for other regions in Europe [36], which could be caused not only by the roughness length of land use and land cover classes, but also by the better representation of friction in higher-resolution topography. The received differences in wind speed forecasts between airport and road meteorological network stations indicate that there are microscale environmental influences in the urban area in terms of the measurement of wind speed and direction by the road meteorological stations. The BIAS and RMSE values illustrated that the simulated near-surface parameters improved when using the highest and more-updated land use and land cover data with a topography adequate to its resolution. The received results indicate that for the analyzed area, a better representation of LULC data is achieved through CLC data than CGLS-LC100, which confirmed its higher thematic accuracy.

Overestimation of the wind speed forecasts at the WRF model during spring and summer was also emphasized in other works [73]. Some steps for an additional increase of wind speed forecast verifiability can be taken; for example, the use of additional measurements of wind speed in the planetary boundary layer from the Doppler wind LiDARs (Light Detection and Ranging) [74] can provide high-accuracy reference data to validate the NWP model. Moreover, that data assimilation can improve the WRF accuracy of wind simulations [73,75].

Our research shows the possible effect of basic meteorological parameter forecast improvement when using more accurate LULC data in very high-resolution WRF modeling. It is worth emphasizing that the WRF model results can have different accuracy for different seasons, synoptic situations, and localizations. Therefore, to provide a more convincing statistical assessment, the study will be continued considering different synoptic situations over a longer period for different seasons and locations. Moreover, due to the development of increasingly up-to-date land cover and land use data, e.g., from the Sentinel-2 mission [76], research into their impact on numerical weather modeling results will be carried out. The currently available 10 m spatial resolution of land cover and land use data from Sentinel 2 Global Land Cover (S2GLC) provide several opportunities for microscale urban weather modeling.

5. Conclusions

Research studies on the impact of land use and land cover on the results of WRF numerical weather prediction model simulations show that providing more up-to-date LULC data with a higher level of detail can effectively improve numerical weather forecasting of basic meteorological parameters. Most of the studies known to the authors on this topic for the European area have investigated the effect of CLC data implementation on WRF simulation results. Due to the limited coverage of CLC data, the authors developed and tested global land use and coverage data from CGLS at the WRF model. The authors analyzed the impact of implementing high-resolution (100 × 100 m) data on land cover and land use from two independent sources, which are CLC and CGLS-LC100.

For the analyzed period, the use of CLC data allowed for an average improvement in the forecasting of temperature by 0.2 °C, relative humidity by 1.5%, wind speed by 0.1 m/s, and wind direction by 14.7° over the default data. The use of CGLS-LC100 data allowed for an average improvement of 0.1 °C in temperature forecast, 1.0% in relative humidity, 0.04 m/s in wind speed, and 4.9° in wind direction compared to the results obtained on the default data. The results indicate the possible effects of basic meteorological parameter forecast improvement when using higher-accuracy LULC data in high-resolution WRF modeling. Nevertheless, forecasts of meteorological parameters are affected in general by different synoptic situations at each season. Therefore, to ensure the statistical significance of the results, long time series are needed to verify of the new LULC datasets in various synoptic conditions for various seasons and localizations to determine whether CLC and CGLS-LC100 can successfully improve numerical weather forecasts.

Based on the results, it is expected that the use of CLC data more effectively improves numerical weather forecasting capability for Warsaw than CGLS-LC100 data, and, hence, the authors suggest that for high-resolution simulations of the WRF model, if CLC data are available for the area under analysis, these data should be considered to be used first, and in the cases where the analyzed area is outside the range of the CLC, data from the CGLS-LC100 from Copernicus should be used.