Impacts of land cover data quality on regional climate simulations

Authors

  • Elif Sertel,

    Corresponding author
    1. Department of Geodesy and Photogrammetry, Istanbul Technical University, Maslak, Istanbul, Turkey
    2. Department of Environmental Sciences, Rutgers University, New Brunswick, New Jersey, USA
    • Department of Geodesy and Photogrammetry, Istanbul Technical University, Maslak, Istanbul, 34469, Turkey.
    Search for more papers by this author
  • Alan Robock,

    1. Department of Environmental Sciences, Rutgers University, New Brunswick, New Jersey, USA
    Search for more papers by this author
  • Cankut Ormeci

    1. Department of Geodesy and Photogrammetry, Istanbul Technical University, Maslak, Istanbul, Turkey
    Search for more papers by this author

Abstract

The land surface influences local, regional and global climate across many time scales. Accurate representation of land surfaces is an important factor for climate modelling studies because land surfaces control the partitioning of available energy and water. Here we introduce new, up-to-date and accurate land cover data for the Marmara Region, Turkey derived from Landsat Enhanced Thematic Mapper (ETM+) images into the Weather Research and Forecasting (WRF) model. We used several image processing techniques to create accurate land cover data from Landsat sensor images obtained between 2001 and 2005. By comparing the new land cover data with the default WRF land cover data, we found that there are two types of error in WRF land cover data that caused misrepresentation of the study region. WRF uses Global Land Cover Characteristics (GLCC) data created from images acquired during 1992 and 1993 and it does not reflect current land cover. And the GLCC includes misclassifications. As a result of these errors, GLCC data do not represent urban areas in the cities of Istanbul, Izmit and Bursa and there are spectral mixing problems between classes, e.g. croplands, urban areas and forests. We used WRF land cover and our new land cover data to conduct numerical simulations. Using meteorological station data within the study area, we found that simulation with the new land cover dataset produces more accurate temperature simulations for the region, thus demonstrating the importance of accurate land cover data. Copyright © 2009 Royal Meteorological Society

1. Introduction

The climate system is complex and interactive, including the land surface, atmosphere, oceans and other water bodies, the cryosphere and the biosphere. Accurate simulation of the climate system requires accurate representation of each component and the physical interactions between the components. The climate over the land surface is extremely important to us because humans are land-dwelling creatures. The land surface controls the partitioning of available energy at the surface between sensible and latent heat and controls partitioning of available water between evaporation and runoff. The way heat, moisture, momentum, dust and pollutants move upward from the surface into the atmosphere is affected by interactions between land surface and the overlying atmosphere. Land has considerable heterogeneity because of the existence of different land cover types such as bare soil, water, urban land, trees and snow, which vary over small areas. This surface variability not only determines the microclimate but also affects mesoscale atmospheric circulation (Hartmann, 1994; Weaver and Avissar, 2001; Yang, 2004).

Land surfaces have changed as a result of human and natural processes, such as urbanization, deforestation, desertification and natural disasters. The impacts of land cover change on warm season climate over different spatial and temporal scales have been studied by Pielke et al. (1999, 2001, 2007), Weaver and Avissar (2001), Pitman et al. (2003) and Ezber et al. (2007). For example, Pielke et al. (1999) investigated the possible impacts of twentieth century land cover change on Florida peninsula's near-surface temperature and convective rainfall for July–August. Pitman et al. (2004) used three high-resolution mesoscale model configurations forced at the boundaries to simulate July climates for natural and current land cover. Gero et al. (2006) investigated the impact of land cover change on storms over the Sydney Basin. They simulated the storm patterns using pre-European settlement land cover and land cover representing Sydney's current land-use pattern. They found that synoptically forced storms were unresponsive to a changed land surface, whereas local convective storms were highly sensitive to the triggering mechanism associated with land surface influences. Shepherd (2005) reviewed the current investigations of urban-induced rainfall and gave recommendations for the future.

Accurate representation of the land surface is important to precisely model the effects of past, current and future land cover. Land cover products used in most climate models were initially compiled from maps and ground surveys until global scale land cover products generated from remote-sensing images became available. These remotely sensed derived global land cover products, such as Global Land Cover Characteristics (GLCC, Loveland et al., 2000), University of Maryland land cover classification (UMD, Hansen and Reed, 2000) and Global Land Cover 2000 (Bartholome et al., 2002) were implemented into various land surface schemes and climate models. Satellite remote sensing of land cover has proven very useful for management of environmental and natural resources, sea and coastline studies and observing land use/land cover changes on global and regional scales for use in weather forecasting and climate modelling (e.g. Pielke et al., 1999; Schweiger et al., 2005; Sertel et al., 2007a; Kaya et al., 2008). However, most of the remotely sensed derived land cover datasets are not 100% accurate, even if developed from the most advanced satellite images (Matthews, 1983; Sellers et al., 1996a, 1996b; Friedl et al., 2002; Yang, 2004; Ge et al., 2007). Ge et al. (2007), for example, showed how the classification accuracy of a land cover dataset employed in a land surface scheme affects simulated cumulative precipitation in a regional climate model. Here we make use of new remote-sensing techniques to investigate the effects of land cover data quality on modelling the climate of the Marmara region of Turkey (Figure 1).

Figure 1.

Location of the study area and distribution of the transects used in Figures 2 and 3

Current land cover data used in regional climate models such as Regional Atmospheric Modelling System (RAMS, Walko and Tremback, 2000), the Fifth-Generation NCAR/Penn State Mesoscale Model (MM5, Grell et al., 1994) and Weather Research and Forecasting (WRF, Skamarock et al., 2005) were obtained from the GLCC database. These data were created using 1-km Advanced Very High Resolution Radiometer (AVHRR) satellite images spanning April 1992 through March 1993 with an unsupervised classification technique.

The GLCC data are not up-to-date and are not accurate for all regions and some land cover types, such as urban areas. Ezber et al. (2007) showed that the MM5 model uses a very old land cover/vegetation map that does not reflect the current urban boundaries in Istanbul. They made manual changes based on topographic maps to fix the land cover data. Pitman et al. (2003) used vegetation data from the Atlas of Australian Resources (AUSLIG, 1990) in their studies to better represent the land surface; however, this source of data will not be appropriate to identify some land cover classes, especially urban areas. These methods are neither convenient nor accurate enough to derive multi temporal land cover data for larger regions representing all types of land cover classes. Here we introduce new, up-to-date and accurate land cover data for the Marmara Region, Turkey derived from Landsat Enhanced Thematic Mapper (ETM) images into the WRF regional climate model to overcome the above-mentioned problems and produce better numerical simulation results.

The objective of this research is to answer the following questions: Are global land cover datasets used in regional climate models representing the land surface accurately? Can Landsat satellite images be used to improve land surface datasets for regional climate models? Do these new land cover data produce improved climate simulations?

2. Study area and data

The Marmara Region occupies the northwest corner of Turkey with a surface area of 67 000 km2 and represents approximately 8.6% of the Turkish national territory (Figure 1). It is the smallest but most densely populated of the seven geographical regions of Turkey. This region includes 11 cities: Istanbul, Bursa, Edirne, Kocaeli, Balikesir, Kirklareli, Tekirdag, Canakkale, Bilecik, Sakarya and Yalova (Sertel, 2008).

The Marmara Region forms a passage between the Balkan Peninsula and Anatolia, and connects Europe and Asia. As a result of being on the edge of Europe, of the Bosphorus and Dardanelles Straits as a passage from the Black Sea to the Aegean Sea, and of ports on the Black and Aegean Seas, this region is highly developed in industry, commerce, tourism, and transportation. Istanbul, Bursa and Kocaeli cities are the centres of large industrial establishments and produce processed food, textile, cement, paper, petrochemical products, automobiles, furniture, leather and ships. The Marmara region is the most industrialized region in Turkey and one third of the country's industry is situated in this region (Sertel, 2008).

Rapid demographic and economic development, industrialization and urbanization occurred in the Marmara Region after the 1980s and the population of the Marmara Region increased dramatically as a result of intense migration from other regions. The rapid industrialization and population increase have caused two changes in landscape characteristics of the Marmara region. Urbanization increased and several agricultural and forest areas have been transformed into urban and built-up areas. Istanbul has been affected because of huge immigration; the city population was 3 million in the 1970s, 7.4 million in 1990s and is currently around 12 million (TSI, 2008). Bursa, another important city of the Marmara region had a population of 275 953 in 1970, but it reached 1 194 687 in 2000.

We used Landsat ETM+ images with 30 m spatial resolution, obtained between 2001 and 2005, to create land cover maps of the study region. The 2005 Landsat ETM+ image included Istanbul and the neighbouring region where most of the land cover changes occurred. The study region covers six Landsat images and each image was evaluated individually to derive accurate information. We also used a 1992 Landsat Thematic Mapper image comprising Istanbul, Bursa, Izmit and their surroundings to investigate the accuracy of GLCC data over this part of the study region. Ground photographs taken during 2001 and 2005, a digital elevation model generated from 1/25 000 scale topographic maps, field samples collected in the Marmara Region showing the types and distribution of vegetation, 1/25 000 scaled topographic maps, forest maps generated by Ministry of Forestry and socioeconomic data (e.g. population, immigration rates) were used in this study. Most of the maps, photographs and field samples were used to select training sites and perform accuracy assessment for the classification. We also collected daily minimum, maximum and average temperature data for 29 meteorological stations from the State Meteorological Office of Turkey.

GLCC data were generated by the U.S. Geological Survey National Center for Earth Resources Observation and Science, the University of Nebraska-Lincoln, and the European Commission's Joint Research Centre (available at http://edcsns17.cr.usgs.gov/glcc/), on a continent-by-continent basis. All continental databases are available in two different map projections, Interrupted Goode Homolosine and Lambert Azimuthal Equal Area, on a 1-km nominal spatial resolution. The dataset was derived from 1-km AVHRR data spanning April 1992 through March 1993 (Loveland et al., 2000). Table I shows the GLCC land use/land cover types.

Table I. U.S. Geological Survey Land Use/Land Cover System Legend (Modified Level 2) (Loveland et al., 2000)
ValueCodeDescription
1100Urban and built-up land
2211Dryland cropland and pasture
3212Irrigated cropland and pasture
4213Mixed dryland/irrigated cropland and pasture
5280Cropland/grassland mosaic
6290Cropland/woodland mosaic
7311Grassland
8321Shrubland
9330Mixed shrubland/grassland
10332Savanna
11411Deciduous broadleaf forest
12412Deciduous needleleaf forest
13421Evergreen broadleaf forest
14422Evergreen needleleaf forest
15430Mixed forest
16500Water bodies
17620Herbaceous wetland
18610Wooded wetland
19770Barren or sparsely vegetated
20820Herbaceous tundra
21810Wooded tundra
22850Mixed tundra
23830Bare ground tundra
24900Snow or ice

3. Methodology

3.1. Remote sensing

The use of satellite imagery has made the mapping of land cover, from global to local scales, much easier and faster. The following remote-sensing techniques were used to create current and accurate land cover map of the study area.

3.1.1. Atmospheric and radiometric correction

All images were atmospherically and radiometrically corrected to minimize contamination effects of atmospheric particles (scattering and absorption effects due to the atmosphere) and systematic errors. Atmospheric correction of the satellite images was performed using the dark object subtraction method, because this approach is a widely used by many scientists and it is simple (Liang et al., 2001; Liang, 2004). Dark objects having very small surface reflectance, namely lakes, seas or other water bodies, were selected for each Landsat frame. It is assumed that the minimum reflectance values in the histograms from the entire scene are from the effects of the atmosphere. The correction is applied by subtracting the minimum observed value, determined for each specific band, from all pixel values in each respective band. After dark object subtraction, haze was removed and the surface features blocked by haze were recovered.

3.1.2. Geometric correction

Geometric correction was performed for each image to eliminate geometric distortions, correct errors in the relative positions of pixels and define images in a common coordinate system (Sertel et al., 2007b). Six Landsat ETM + images were geometrically corrected using the affine transformation model. The nearest neighbourhood method was used for the resampling to preserve the radiometric properties of the images. The affine model parameters required were estimated by least square adjustment. Each geometric correction has a root mean square error (RMSE) less than 0.5 pixel. All images were transformed into a latitude and longitude coordinate system with the World Geodetic System WGS 84 Datum (Sertel et al., 2007b). Locations that can be easily and accurately identified such as highway and road intersections, ports and bridges and approximately homogeneously distributed on the images were selected as Ground Control Points (GCPs). Ground coordinates of these points were provided from 1:25 000 topographical maps and other satellite sensor images.

3.1.3. Classification

The objective of classification is to assign all pixels in the image to particular classes or themes. U.S. Geological Survey land use/land cover system themes presented in Table I were used in the study. The resulting classified image comprised a mosaic of pixels, each of which belongs to a particular theme, and is essentially a thematic map of the original image. Both unsupervised and supervised classification techniques were utilized to form the most accurate land cover dataset for the study region. Each image was analysed independently; in some regions, especially for urban and sandy areas, pilot regions were created and were classified separately to minimize the mixed pixel problem.

There were spectral mixing problems between crop and urban areas and between crops and some forest areas. To eliminate this effect, pilot areas determined by semivariograms and spatial profiles were subsetted from the images and evaluated separately. Therefore, urban areas were acquired more accurately and spectral mixing problems in many areas were eliminated. Also, using semivariograms and spatial profiles, regions with significant land cover changes were determined easily and accurately.

3.1.4. Selection of pilot areas

We used the semivariogram approach to find pilot regions that experienced significant land cover change. A semivariogram describes the spatial relationship between the sample values by using the variation of samples with distance and direction. Semivariogram parameters can be used to identify abrupt land cover changes and derive textural information for the classification (Sertel et al., 2007a). Because of the spatial variations caused by land cover changes, the semivariogram parameters of range, nugget and sill can be used to quantify the form of this spatial variability. The range captured coarse-scale spatial variability, the nugget captured fine-scale spatial variability and the sill captured overall variability in the landscape.

We found that the semivariogram shape was different for different years if the area faced significant land cover change, but similar if the area did not change. The semivariogram range was used to quantify coarse spatial variability. Transects covering areas of significant land cover change had even larger increases in range whereas transects covering unchanged areas had similar ranges for different years. In some cases, the semivariograms obtained did not pass through the origin and had nugget effects, which provide information about fine-scale spatial variation. Transects covering changed areas had a much larger increase in nugget than the transects of unchanged areas.

We also conducted a study to analyse the changes in different regions of Istanbul using spatial profiles (Kaya et al., 2008). Spatial profiles of the regions have different values for different years, emphasizing the impact of land cover changes within the regions. Figure 2 is an example that shows the spatial profiles of a transect obtained from Kilyos–Karaburun site and they clearly revealed the land cover change in the area. The 1984 spatial profile has lower pixel values because of being on the sea in that year. However, the Istanbul-Black Sea coastline has been changing after 1990s as a result of open mining activities filling the Black Sea with open mining residue and conversion of some forest areas into barren and sparsely vegetated areas in this region. The 2005 profile reveals the changing areas in coastline with higher pixel values especially in the regions 3800–5000 m, 5800–8600 m, and 10 345–11 685 m.

Figure 2.

Spatial profiles and location of the transect. Figure 1 gives location of the transect (Transect 1) at a larger spatial scale

The overall accuracy of the classification obtained by classifying each image was around 65% because of spectral mixing problem and inadequate training sites. We classified many sub-regions separately to increase overall classification accuracy, minimize spectral mixing problem and identify urban areas more clearly. This could be done by selecting sufficient and appropriate training sites and defining sub-regions to be handled separately. We used semivariograms and spatial profiles to determine the location of changed regions and quantify the spatial variation. Based on the information obtained from semivariograms and spatial profiles, sub-regions were selected on Istanbul-Black Sea coastline, urban areas of Istanbul, Bursa, Izmit and Adapazari, forest areas in Istanbul located between the Bosphorus and Black Sea, crop areas around Adapazari and Izmit and forest areas in Canakkale, and these regions were classified separately. This procedure improved the classification accuracy 18%, minimized the spectral mixing and identified the urban areas.

3.1.5. Classification accuracy

A confusion matrix and some common measures derived from this matrix, namely overall accuracy, user's accuracy, producer's accuracy and Kappa coefficient, were used to assess the accuracy of derived land cover data (Sertel, 2008). The confusion matrix is currently at the core of the accuracy assessment literature in remote sensing. This matrix is a simple cross-tabulation of the mapped class label against that observed in the ground or reference data for a sample of cases at specified locations (Foody, 2002). Detailed information about the calculation of overall accuracy, user's accuracy, producer's accuracy and Kappa coefficient can be found in Foody (2002). Accuracy assessments of classified images were performed from aggregated images, because these data were used for land cover change determination, land cover comparison with global datasets, and climate modelling. We selected 250 ground points and compared the classification result and reference land cover based on ground truth data. Our 2005 classified image has an overall accuracy of 83% and Kappa coefficient of 82%.

3.2. Numerical modelling experiment

One main and two nested domains were formed to conduct numerical simulations using the WRF modelling system. To analyse the larger scale circulation of the atmosphere, a main domain was created to cover latitudes 25–60°N and longitudes 15°W–55°E. The first nested domain includes the western part of Turkey and its near surroundings and the second nested domain includes the Marmara Region, which is the main focus of this research from the land cover change perspective. The main domain has a 27 km spatial resolution, whereas the nested domains have 9 and 3 km, respectively (Figure 3).

Figure 3.

Locations of the main and nested domains for the WRF model runs

The vertical regime of both main and nested domains extends over 28 vertical levels, from the surface to 50 mb. The main domain (labelled 1 in Figure 4) has 180 × 142 grid points, the first nested domain (2) has 130 × 115 grid points and the innermost domain (3) has 166 × 130 grid points. Initial and boundary conditions were obtained from National Centers for Environmental Prediction-Department of Energy Reanalysis II and the Noah model was selected as the land surface model (LSM). Two numerical experiments were conducted; one with the standard WRF land cover data (control run) and the other with our newly created land cover data (new land cover run). Elevation values of changed land cover types were updated based on a current Digital Elevation Model of the study area. Simulations were conducted for summer (June, July, August) of 2004, with the simulations starting on 1 June and ending on 31 August. After extensive experimentation with WRF to get it to do a reasonable job of simulating the current climate, the following configuration was used for the experiments: the ARW version of WRF, NOAH LSM, Kain-Fritsch cumulus parameterization scheme, Lin et al. microphysics scheme, Yonsei University planetary boundary layer scheme, and Rapid Radiative Transfer Model (RRTM)/Dudhia longwave/shortwave radiation scheme. More details on all these WRF model choices are given in Skamarock et al. (2005).

Figure 4.

New and old (GLCC) land cover classification. Selected meteorological stations for analyses in Figure 9 are black circles. Locations of detailed analyses in Figures 6–8 are black boxes. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

4. Results

4.1. Comparison of GLCC data with new land cover data

We compared GLCC data with new land cover data and found two types of error in GLCC causing inaccurate representation of the current land cover types in the study region. GLCC land cover data are not up-to-date (derived images collected between 1992 and 2003) and this is the cause of the first type of error. There have been some changes in the Marmara Region after the 1990s; therefore, these areas cannot be found in GLCC data. The second type of error is caused by misclassification of the images. We have found that GLCC data do not represent urban areas in the cities of Istanbul, Izmit and Bursa and have spectral mixing problems between some classes like croplands, urban areas and forests. Figure 4 shows new and old (GLCC) land cover classification.

Figures 5 and 6 show the difference between GLLC and new land cover data and these figures are examples of the classification errors. Both ground truth and new land cover data show the expansion of urban areas in the Istanbul metropolitan area, but in the GLCC data only a limited area along the Bosporus is shown as urban (Figure 6). Moreover, ground truth data obtained for 1992 showed that Istanbul had more urban areas in this region than what GLCC illustrated. GLCC data cannot represent either 1992 or current urban areas in Istanbul. Landsat derived new land cover data indicate that the northern part of Istanbul is covered by evergreen and deciduous forest (verified with current ground truth data), but the GLCC data indicated that most of this region is covered with croplands (Figure 6). Also, in the northern part of the Marmara Region towards the Black Sea coast, there is now bare ground as a result of open mining activities. This barren land class (19) can be identified in the new land cover data, whereas the GLCC data indicated this region as woodland (11) and crop (2) (top parts of Figure 6). The difference in GLCC is a result of misclassification but not land cover change because the northern part of Istanbul has been forest for a long period of time verified with 1975, 1987, 1992 and 2005 satellite images. The coastline part was also barren land in 1992 ground truth data.

Figure 5.

Comparison of (a) WRF GLCC land cover data, (b) new Landsat-derived land cover data for urban areas in Istanbul and (c) their difference. The numbers in panels (a) and (b) refer to the land cover types in Table I, and in panel (c); 1 means change and 0 means no change. Figure 5 gives box location. Each pixel has a size of 1 km × 1 km

Figure 6.

Comparison of (a) WRF GLCC land cover data, (b) new Landsat-derived land cover data for forest areas in Istanbul and (c) their difference. The numbers in panels (a) and (b) refer to the land cover types in Table I, and in panel (c); 1 means change and 0 means no change. Figure 5 gives box location. Each pixel has a size of 1 km × 1 km

Figure 7 includes two types of error together. Urban areas in Bursa increased after the 1990s; therefore, it is not expected that GLCC would represent the current land cover because it is an older dataset. Bursa city has a very small portion of urban area in GLCC but based on 1992 ground truth data, the total area of urban areas is twice the total urban area in GLCC. Bursa has an even larger urban area in the current land cover dataset. New land cover data represent the current land surface accurately, but most of the urban areas were misrepresented and classified as woodlands or croplands in GLCC data.

Figure 7.

Comparison of (a) WRF GLCC land cover data, (b) new Landsat-derived land cover data for urban areas in Bursa and (c) their difference. The numbers in panels (a) and (b) refer to the land cover types in Table I, and in panel (c); 1 means change and 0 means no change. See Figure 5 for box location. Each pixel has a size of 1 km × 1 km

Accuracy assessments for the GLCC data give classification accuracies of 45, 40 and 60% for the areas shown in Figures 5, 6 and 7, respectively.

4.2. Numerical modelling experiment

A comparison was made between the results obtained with new land cover run and control run to find out if the new land cover data can improve the numerical simulations and if the land cover changes in the region impact the climate. Four meteorological stations were selected and comparisons were made between observations and the results of the new land cover run and control run for minimum, maximum and average temperature values. We compared the results of the model by using a 3-km square grid box that included the station. The locations of the meteorological stations are shown in Figure 4, Edirne in the Ergene Section, Florya in Istanbul, Sakarya in the Catalca-Kocaeli Section and Bursa in the Southern Marmara Section of the Marmara Region. Table II shows the land cover types around each station for each land cover dataset. In these regions, different types of land cover changes have occurred, like conversion from crop to urban, woodland to urban and forest to urban. Each of these land cover types has different parameter settings within LSM. Table III shows the vegetation and land use-related parameters in LSM. These parameters and governing equations of the LSM were explained in detail by Chen and Dudhia (2001). Changes in land cover cause changes in vegetation and land-use-related parameters like albedo, surface roughness, green vegetation fraction and stomatal resistance, which lead to differences in sensible heat flux, latent heat flux and many other variables dependent on vegetation and land-use parameters.

Table II. Land cover types, averaged for a 3-km circle centred on each station. Figure 7 gives the station locations and the circles
StationModel (GLCC) land coverLandsat (new) land cover
Florya60% crop, 40% woodland90% urban, 10% crop
Bursa50% forest, 50% urban30% forest, 70% urban
Edirne100% crop100% crop
Sakarya40% urban, 60% crop60% crop, 40% urban
Table III. Vegetation-related parameters in the LSM, which include roughness length (Z0) in m, green vegetation fraction (SHDFAC) and stomatal resistance (RS) in s m−1
ValueLand cover typeAlbedoZ0SHDFACRS
1Urban and built-up land0.151.000.1200.
2Dryland cropland and pasture0.190.070.840.
3Irrigated cropland and pasture0.150.070.840.
4Mixed dryland/irrigated cropland and pasture0.170.070.840.
5Cropland/grassland mosaic0.190.070.840.
6Cropland/woodland mosaic0.190.150.870
7Grassland0.190.080.840
8Shrubland0.250.030.7300
9Mixed shrubland/grassland0.230.050.7170
10Savanna0.200.860.570
11Deciduous broadleaf forest0.120.800.8100
12Deciduous needleleaf forest0.110.850.7150
13Evergreen broadleaf forest0.112.650.95150
14Evergreen needleleaf forest0.101.090.7125
15Mixed forest0.120.800.8125
16Water bodies0.190.000100
17Herbaceous wetland0.120.040.640
18Wooded wetland0.120.050.6100
19Barren or sparsely vegetated0.120.010.01999
20Herbaceous tundra0.160.040.6150
21Wooded tundra0.160.060.6150
22Mixed Tundra0.160.050.6150
23Bare ground tundra0.170.030.3200
24Snow or ice0.700.0010999

Figure 8 shows the minimum, maximum, and average temperatures of Florya station from observations, the control run and the new land cover run. Florya has 60% crop- and 40% woodland in the model land cover, but 90% urban areas and 10% cropland in the new, more accurate, land cover dataset (Table II). The minimum, maximum and average temperatures from the simulation with the new land cover have values closer to observations. As shown in Table IV, RMSE values were calculated by comparing observed temperature values with temperature values obtained from the new land cover and control runs. RMSE values for the new land cover run are between 2.1 and 3 °C, whereas RMSE values obtained for the control run are between 2.9 and 7.1 °C. Because the dominant land cover around Florya is urban, introducing new land cover data improved the results and better RMSE values were obtained with the new land cover. With the conversion of crop- and woodland to urban areas, albedo values decreased from 0.19 to 0.15, roughness length increased from 0.07 m (cropland) and 0.15 m (woodland) to 1.00 m and stomatal resistance increased from 40 sm−1 (crop) and 70 sm−1 (woodland) to 200 sm−1. The Bowen ratio changed as a result of land cover change near the meteorological station because of albedo, roughness length, greenness fraction and stomatal resistance changes. Conversion from woodland to urban decreased the amount of latent heat flux emitted from the surface, which leads to an increase of maximum temperature. Conversions of land cover type from cropland to urban land and from woodland to urban land cause warming in the summer.

Figure 8.

Minimum, maximum and average temperatures for Florya, from regional climate model simulations with standard model land cover (control) and new one. Figure 5 gives station location. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

Table IV. Root mean square error (RMSE) values obtained from the comparison of the new land cover run with observations and the control run with observations
Station nameRMSE (min temperature, °C)RMSE (max temperature, °C)RMSE (avg temperature, °C)
Florya (new land cover vs obs)2.23.02.1
Florya (control vs obs)2.97.14.6
Bursa (new land cover vs obs)2.13.62.4
Bursa (control vs obs)2.55.73.4
Sakarya (new land cover vs obs)2.83.13.3
Sakarya (control vs obs)3.23.23.9
Edirne (new land cover vs obs)2.33.11.8
Edirne (control vs obs)2.43.11.9

Figure 9 shows minimum, maximum and average temperatures at Bursa obtained from model simulations and observations. Minimum temperature values obtained from the control and new land cover data runs are similar to each other, but better maximum and average temperatures were obtained with the improved land cover data. The Bursa region has 50% urban areas and 50% forest land in the model land cover data, but 70% urban areas and 30% forest in new land cover data, which leads to better maximum and average model temperatures. RMSE values obtained for new land cover run are between 2.1 and 3.6 °C, whereas RMSE values obtained for control cover run are between 2.8 and 5.7 °C (Table IV). Forest (deciduous) to urban conversion resulted in increases in albedo, roughness length and stomatal resistance.

Figure 9.

Minimum, maximum and average temperatures for Bursa, from regional climate model simulations with standard model land cover (control) and new one. Figure 5 gives station location. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

For the Sakarya meteorological station, model land cover within the 3 km buffer of Sakarya is 40% urban and 60% crop, whereas it is 60% urban and 40% crop in the new land cover. Because the land cover types are very similar and the change is very small for model and new land cover data, simulated temperature values were close to each other. RMSE values obtained for the new land cover run are between 2.5 and 3.3 °C, whereas RMSE values obtained for the control cover run are between 3.2 and 3.9 °C (Table IV). The last comparison was conducted for Edirne station where both model and new land cover data have 100% cropland. As a result of this, both simulations gave similar results for minimum, maximum and average temperature values. Comparison of two simulations with observations gave very similar RMSE values (Table IV).

Precipitation results obtained with model land cover and new land cover data were very similar to each other. The simulated precipitation showed that the model was able to capture the main features of the observed precipitation except around the Bosphorus. Precipitation change as a result of land cover change was not significant for the region.

Modifying the model land cover by correcting the land cover classification to reflect current conditions improved the simulation of minimum, maximum and average 2 m air temperature results significantly. The results showed that temperature is more sensitive to local land cover change than precipitation, for this location and time of year.

5. Conclusions

Accurate representation of land surface is important to conduct reliable and accurate numerical modelling experiments. Because different land cover types have different vegetation and land use related parameters, such as albedo, surface roughness, green vegetation fraction and stomatal resistance, inaccurate representation of land cover will lead to differences in simulating sensible heat flux, latent heat flux and many other variables depending on vegetation and land use parameters. Remote sensing provides accurate representation of Earth's surface on different spatial and temporal scales and is an attractive source for creating land cover data.

We used Landsat images to create land cover data for regional climate modelling and investigated the accuracy of a land cover dataset (GLCC) that has been used in several atmospheric modelling systems for the Marmara Region of Turkey. Comparisons of GLCC data with our new land cover dataset derived from Landsat images showed that the GLCC data are not up-to-date and accurate for some land cover types, especially for urban and forest areas. We were able to show that Landsat images can be successfully used to create improved land cover data for regional climate modelling. These new land cover data improved the numerical simulation results, especially for minimum, maximum and average temperatures, and emphasized the importance of accurate land cover for regional climate modelling.

The results of this work are evidence that land cover change has impacted the regional climate of the Marmara Region. Because the GLCC data were determined in the early 1990s and our new dataset was produced in the early 2000s, we can interpret the differences in simulation results also as a study of land cover change impact on climate. Urbanization increase in Istanbul and Bursa resulted in statistically significant changes in urban climatology of the region, with the increase especially of maximum and average summer temperatures.

In our results, temperature was more sensitive to local land cover changes than precipitation. Conversions of land cover type from cropland to urban land, woodland to urban land, and forest to urban land caused warming within the region. Conversions from woodland to urban land and forest to urban land decreased the amount of latent heat flux emitted from the surface, which leads to an increase of maximum temperature. However, because this was a dry season and there was not much precipitation, and because we only simulated one season for one particular year, we were not able to demonstrate significant changes in precipitation.

Acknowledgements

We would like to thank the State Meteorological Office of Turkey for providing meteorological station data. This research was supported by a TUBITAK (Turkiye Bilimsel ve Teknoloji Arastirma Kurumu-The Scientific and Technological Research Council of Turkey) Ph.d. Fellowship and a Fulbright Fellowship to ES. AR was supported by U.S. NSF grant ATM-0450334.

Ancillary