Land-cover change analysis in 50 global cities by using a combination of Landsat data and analysis of grid cells

We selected 50 global cities from five continents, including ones from developed (Europe, North America and Japan) and developing regions (Asia, Africa and South America). Converting land to urban land use leads to increases of greenhouse gas emissions (Angel et al 2011). For further investigate the relationship between urban land cover changes and greenhouse gas emissions, out of the 50 cities, 44 (not including Amsterdam, Berlin, Bern, Rome, Ottawa and Lisbon) are studied by Kennedy et al (2011) to estimate urban greenhouse gas emissions. Hence, if we have detailed urban growth patterns of these cities, it may allow us to explore relationships between the greenhouse gas emissions and urban growth patterns, which are critical to evaluate land use-driven greenhouse gas emissions from urban regions. This was not the focus of our work here. However, we will investigate this point in our future research.

Urbanization is characterized by notable gains in urban/built-up areas at the expense of green or open spaces. In this study, the urban/built-up land-cover class includes all non-vegetative, human-constructed elements such as buildings, asphalt and concrete, including residential, commercial, industrial, office use and transportation space. Other types of urban land use, such as golf courses, urban green parks and nature areas, are not included in the urban/built-up class.

The MODIS Land Cover Type product (MCD12Q1: Land Cover Type Yearly L3 Global 500 m SIN Grid V005) for 2005 was used to delineate the urban boundaries of all 50 cities to further refine the study landscape instead of using the administrative boundary of the city or metropolitan area (figure 1). First, we extracted the urban/built-up land-cover class and then buffered urban/built-up pixels up to a distance of 10 km to define each urban boundary.

Zoom In Zoom Out Reset image size

Table 1 lists the 50 cities and the corresponding Landsat images used for land-cover classification. We acquired Landsat TM and ETM+ images to interpret land-cover change on two separate dates (nominally around 1985 and 2010). Landsat data from these approximate time-points covered nearly all 50 cities with 73 and 72 Landsat scenes per time step (145 Landsat scenes overall). However, owing to cloud coverage over Cape Town (South Africa), Glasgow (United Kingdom), Delhi and Kolkata (India) around these times, we used Landsat data from 1995 and 2000 for Glasgow, 1985 and 2000 for Cape Town, 1999 and 2009 for Delhi and 2000 and 2007 for Kolkata (table 1).

All Landsat standard data products are processed using the Level 1 product generation system with the following parameters applied: Cubic Convolution resampling method, 30 m (TM, ETM+) pixel size, universal Transverse Mercator map projection, World Geodetic System 84 datum and North-up image orientation. The standard terrain correction, which provides systematic radiometric and geometric accuracy by incorporating ground control points while employing a Digital Elevation Model for topographic accuracy, is used (http://landsat.usgs.gov).

In both developing and developed countries, cities from different regions are characterized by different land-cover types and urban sprawl patterns. They vary in shape and size (e.g., overstretched blocks and low-density development) and are surrounded by a variety of landscapes (e.g., agriculture, forest, grassland and desert) (Schneider and Woodcock 2008). Therefore, a key requirement for this study was to create land-cover classes to standardize the classifications and therefore facilitate comparisons. We adopted the six broad land-use categories described by the Intergovernmental Panel on Climate Change (IPCC) (IPCC 2003) for land-cover classes (table 2).

Very high resolution satellite data (from, e.g., QuickBird, IKONOS and WorldView-2), acquired from web-based resources such as Google Earth and visual interpretation of the remote sensing data were used to acquire the ground truth dataset. Land-cover classification accuracy at different times of the year is not constant due to seasonal variations in spectral characteristics of different land cover classes. Hence, knowledge about suitable seasons and timing for land-cover mapping is important for generating accurate land cover maps (Sinha et al 2012). In consideration of the weather conditions, to maximize the vegetation information content for each monitoring date, only images from the green vegetation season and with low cloud cover were used. All analyzes were based on the optical and thermal infrared bands of the TM and ETM+ data, and panchromatic bands were excluded. If the images contained cloud cover and associated shadows, we isolated the clouds and their associated shadow portions by visual image interpretation, and then substituted the selected portions with clear pixels from Landsat imagery of the same area from the nearest dates. The MLC method was used to produce land-cover maps from the Landsat images. MLC is a classification technique widely used in the remote sensing community. With the MLC method, probability density functions are built for each class based on the training samples' spectral values. During classification all pixels are assigned class membership based on the relative likelihood (probability) of that pixel occurring within each class' probability density function (Lillesand et al 2008). MLC is easy to apply and widely available in image processing and statistical software packages. The newly developed classifier may be improve classification accuracy when compared with the MLC method. However, gathering specific knowledge is an enormous task and tuning parameters is very costly. Therefore, the newly developed classifier are not universally applicable (Xie et al 2008).

We created 1 km² grid square cells for the 50 cities. The grid square cells enable us to represent the land-cover maps in 1  × 1 km grid square cells. To evaluate the spatio-temporal changes of land-cover categories and compare them, the classified images were intersected with empty 1 km² grid square cells and the percentages of the six land-cover types within each cell were computed. The results were stored in a new attribute table. To calculate the percentage of a land-cover type within a cell, we divided the sum of the land-cover-type area by the area of the cell. For a pixel located on a grid line (or spanning multiple grid lines), we assigned the pixel to the cell that covered the largest part of the pixel. We then calculated the correlation coefficients of the land-cover categories based on the 1 km² grid square cells to investigate the relationships among changes in the land-cover categories.

Thus, the grid cells enable us to aggregate the categories for each map and to calculate their proportions. Furthermore, they enable us to evaluate the spatio-temporal changes in land-cover categories to allow a much easier statistical comparison of the land-cover changes.

Classification by the MLC method was performed for the 50 cities. Overall accuracy for the resulting land-cover maps for each year ranged between 77 and 95%. The lowest classification accuracy was for New York City in 2009 (77.05%) and the highest was for Seoul in 2006 (94.92%). However, Cropland classes tended to be more strongly affected by changes in observation time than other classes. Therefore, the Cropland class was the most difficult among all categories to classify and was confused mainly with Grassland. Using these maps, we assessed the land-cover changes in the cities caused by urban expansion over the study period, and then examined the relationship between the expansion of urban area and other changes in the land-cover categories.

There are several major trends evident in the changes of land cover over the approximate 1985–2010 period. The suburban areas surrounding these core cities experienced enormous increases in Settlements area, and there was a marked decrease in green space. In Beijing (figure 2), the area of Settlements increased from 21.3% of the study area in 1984 to 34.5% in 2010, whereas Cropland area decreased from 43.6% in 1984 to 18.4% in 2010 and Forest decreased from 25.6% in 1984 to 9.8% in 2010. In Brussels, Settlements area doubled over the two decades, increasing from 17.7% of the study area in 1987 to 33.9% in 2010, whereas Cropland decreased from 26.1% in 1987 to 16.4% in 2010 and Grassland decreased from 25.1% in 1987 to 18.3% in 2010. In Seoul, Settlements area increased more than two-fold over the two decades, increasing from 14.1% of the study area in 1986 to 38.9% in 2006, Grassland decreased from 28.8% in 1986 to 17.6% in 2006 and Wetlands decreased from 13.0% in 1986 to 9.4% in 2006. In Los Angeles, Settlements area grew from 32.8% in 1986 to 46.0% in 2009, while Grassland decreased from 26.2% in 1986 to 14.8% in 2009.

Zoom In Zoom Out Reset image size

Figure 3 shows the rates of increase in Settlements area during the study period. More than 32 of the cities grew rapidly by more than 30% between around 1985 and 2010. The fastest growing city was Bangkok, Thailand, (259% increase) followed by Seoul, Republic of Korea, (173% increase) and Tianjin, China (147%). Of the cities included in this study, the fastest growing were found predominantly in Asia, whereas the slower growing cities were almost all located in North America and Western Europe.

Zoom In Zoom Out Reset image size

As shown in figure 3, by 2006 the Settlements area of Bangkok was already more than 2.5 times what it was in early 1988. In fact, over the past several decades, rural-urban migration has been a major contributor to urban growth in Bangkok, where the population increased from 2.1 million in 1960 to 6.3 million in 2000 (Choiejit and Teungfung 2005).

To investigate the relationships among the land-cover changes caused by urban expansion, we calculated the correlation coefficients between Settlements and the other land-cover categories (i.e., Forest, Cropland, Grassland and Wetlands) based on the 1 km² grid square cells. Table 3 presents a summary of the linear correlation coefficient matrix for the changes of land-cover categories in the 50 cities during the study period.

Table 3 shows that the Settlements changes had a strong negative correlation with changes of greenfields (i.e., Forest, Cropland and Grassland) and Wetlands, which implies that the enormous expansion of Settlements area resulted in rapid decreases in green spaces and wetland. Note the following in particular:

• (a)  
  Settlements changes had a negative correlation with Forest changes in 30 cities. A strong negative correlation was observed in New York City, where the value of the correlation coefficient was −0.83, followed by Seattle, Porto and Helsinki with −0.56, −0.52 and −0.51, respectively.
• (b)  
  Settlements changes had a strong negative correlation with Cropland changes in 24 cities, particularly in Tokyo and Beijing, where the linear correlation coefficients were −0.79 and −0.73, respectively.
• (c)  
  Settlements changes had a strong negative correlation with Grassland changes in 30 cities, particularly in Brussels, Toronto and Rio de Janeiro, where the linear correlation coefficients were −0.78, −0.75 and −0.75, respectively.
• (d)  
  Settlements changes had a weak negative correlation with Wetlands in seven cities. Relatively strong negative correlations were found in Bangkok, Seoul, and Beijing where the linear correlation coefficients were −0.30, −0.24, and −0.23, respectively.

As shown in table 3, the differences in correlation coefficients of Settlements changes between Forest, Cropland and Grassland suggest that the availability of land for urban expansion varied amongst cities. However, some of the unrealistically high or low correlation coefficient values were possibly due to the fact that the land cover classification results affected by changes in observation time.

Figure 4(a) shows the grid-cell-based spatial change of Settlements area from 1984 to 2010 in Beijing. The value of each grid square cell was calculated by subtracting the Settlements area of 1984 from that of 2010 in each grid cell and then dividing the changed area by the cell area. As figure 4(a) illustrates, the Settlements area rapidly expanded mainly in three directions: north, east and south. Figure 4(b) shows the grid-cell-based spatial change of Cropland area from 1984 to 2010 in Beijing, which was calculated in the same way as for figure 4(a). The Cropland area decreased rapidly around the core city, typically in three directions: north, east and south. As shown in figure 2, and figures 4(a) and (b), the Cropland area has been mainly converted to the urban/built-up area and Grassland. Results from our correlation analyzes also show that Cropland changes was negatively correlated with Grassland changes (r = −0.59) in Beijing. The conversion of Cropland to Grassland was partly due to several government initiatives such as the Desertification Combating Program around Beijing and Tianjin, and the Conversion of Cropland to Forest and Grassland Program (also known as the 'Grain for Green' Program) (Liu et al 2010). Upon completion of these programs, part of the Cropland was converted to urban land use, grassland and shrubs (Hu 2007). Another reason is that Cropland is temporarily converted to bareland when the land is cleared to prepare the site for construction and this bareland would usually be covered with weed grass during the transition period.

Zoom In Zoom Out Reset image size

As figures 4(a) and (b) illustrate, in Beijing, the Settlements area rapidly expanded to the surrounding suburban area (north, east and south parts) where it was mainly flat; in contrast, the urban/built-up area decreased in the centre of the city. In fact, many high-rise (e.g., office or commercial) buildings replaced the dense low-rise buildings in the city centre concurrently with the development of housing estates in the suburbs. Meanwhile, the greatest decreases in agricultural land area took place in areas suitable for urban land use where the demand for land for urban purposes is high. This suburbanization process expanded urban areas enormously over a short time. This finding is consistent with previous studies of Tokyo, Japan, where urban areas expanded at the expense of agricultural land (Bagan and Yamagata 2012).

Figures 4(c) and (d) show the spatio-temporal dynamics of Settlements and Grassland changes in Los Angeles, and figures 4(e) and (f) shows them for Settlements and Wetlands changes in Seoul. Figures 4(c)–(f) are also calculated in the same way as for figure 4(a). In Los Angeles, urban area rapidly expanded to suburban area causing either fragmentation or a complete loss of Grassland (figures 4(c) and (d)). Conversions to the developed area from grasslands/shrublands in the Central California Valley are also reported by Sleeter et al (2011). In Seoul from 1986 to 2011, the Wetland area rapidly decreased around the human development area (figures 4(e) and (f)).

Figure 5(a) shows the relationship between Settlements and Cropland changes from 1984 to 2010 in Beijing. We found a strong, negative linear relationship (r = −0.73), suggesting that a vast area of Cropland had been converted to urban/built-up area during the last two decades in Beijing. Figure 5(b) shows a strong negative relationship (r = −0.83) between Settlements and Forest changes in New York City during 1986–2009, suggesting a vast area of forest conversion to developed area. Figures 5(c) and (d) show the relationships between Settlements and Grassland changes in Brussels and Los Angeles, respectively. The strong negative relationship between Settlements and Grassland changes indicates that the urban expansion causes loss of Grasslands near urban areas.

Zoom In Zoom Out Reset image size

Figures 5(e) and (f) show the relationships between Settlements and Wetlands changes in Seoul and Bangkok, respectively. The correlations are very weak in both cities, implying that Wetlands have been converted to other land-cover types. In fact, in Seoul, the correlation between Wetlands and Grassland changes is −0.45, which implies that part of Wetlands has been converted to Grassland.

As figures 3–5 illustrate, the loss of green spaces was spatially concentrated in human development areas. These results confirm that the tendency to continuously increase the amount of land with residential and industrial facilities and rapid urban development directly cause loss of farmland and other vegetation on the urban periphery.

The rapid growth in global urban land cover is likely to continue as long as urban populations continue to grow. The future expansion of cities into arable lands causes concern that the displacement of agricultural land by urban land use will require, where possible, bringing new land into cultivation as well as increasing land productivity (Angel et al 2011).