Effect of Land Use and Cover Change on Air Quality in Urban Sprawl

Due to the frequent urban air pollution episodes worldwide recently, decision-makers and government agencies are struggling for sustainable strategies to optimize urban land use/cover change (LUCC) and improve the air quality. This study, thus, aims to identify the underlying relationships between PM10 concentration variations and LUCC based on the simulated PM10 surfaces in 2006 and 2013 in the Changsha-Zhuzhou-Xiangtan agglomeration (CZT), using a regression modeling approach. LUCC variables and associated landscape indexes are developed and correlated with PM10 concentration variations at grid level. Results reveal that the overall mean PM10 concentrations in the CZT declined from 106.74 μg/m3 to 94.37 μg/m3 between 2006 and 2013. Generally, variations of PM10 concentrations are positively correlated with the increasing built-up area, and negatively correlated with the increase in forests. In newly-developed built-up areas, PM10 concentrations declined with the increment of the landscape shape index and the Shannon diversity index and increased with the growing Aggregation index and Contagion index. In other areas, however, the reverse happens. These results suggest that LUCC caused by urban sprawl might be an important factor for the PM10 concentration variation in the CZT. The influence of the landscape pattern on PM10 concentration may vary in different stages of urban development.


Introduction
Urban sprawl, one of the most significant causes of the increasingly severe air pollution in the world [1][2][3], has made recent headlines in peer-reviewed journals of economics, urban planning, and public health [4][5][6][7]. As a direct result of urban sprawl, land use/cover changes (LUCC), as well as their spatial distribution (i.e., landscape pattern) variations may affect pollutants emission indirectly through industrial layout, travel behavior, and other human activities [3,8,9]. In view of the widespread health impacts of air pollution [10,11], studies have increasingly focused on the association between LUCC and air pollution variation caused by urban sprawl, including research on numerical simulations and empirical statistical modeling.
Numerical simulations refer to coupled modelling systems dealing with emissions from plants and traffic, and their dispersion-related meteorological and terrain factors. They are able to retrieve detailed air pollution distribution characteristics and address the complex links between LUCC, pollution discharge, weather patterns, and atmospheric chemistry. In this field, an early investigation was carried out by Civerolo, et al. [12]. They evaluated the influence of the increased urban land cover

Study Domain
The Changsha-Zhuzhou-Xiangtan agglomeration (CZT) is located in the northeast of Hunan province in China, comprising the cities of Changsha, Zhuzhou and Xiangtan (Figure 1). The CZT covers an area of 28,087 km 2 , and has a population of 13.78 million. In terms of transportation, the CZT is a national traffic hub, linked together by a comprehensive railway system, as well as major arterial road networks. The Xiangjiang River flows through this area in a way that promotes efficient goods movement. The CZT is the most important industrial base and the economic growth hub of Hunan Province with an urbanization rate of 61.0% in 2013, 17.7% higher than the rest of Hunan and 7.3% greater than the national average. However, with these achievements, these three cities are the major pollution emission sources in Hunan province.

Study Domain
The Changsha-Zhuzhou-Xiangtan agglomeration (CZT) is located in the northeast of Hunan province in China, comprising the cities of Changsha, Zhuzhou and Xiangtan (Figure 1). The CZT covers an area of 28,087 km 2 , and has a population of 13.78 million. In terms of transportation, the CZT is a national traffic hub, linked together by a comprehensive railway system, as well as major arterial road networks. The Xiangjiang River flows through this area in a way that promotes efficient goods movement. The CZT is the most important industrial base and the economic growth hub of Hunan Province with an urbanization rate of 61.0% in 2013, 17.7% higher than the rest of Hunan and 7.3% greater than the national average. However, with these achievements, these three cities are the major pollution emission sources in Hunan province.

Land Use/Cover Classification
Materials used for monitoring land use include cloud-free Landsat TM (for 2006) and Landsat OLI images (for 2013) in the same time phase. Data were downloaded from the United States Geological Survey (USGS) database [31]. Taking the land cover features of the CZT into account, a modified version of the land use classification was developed [3,16]. The categories include: (1) water, (2) bare soil, (3) built-up, (4) agricultural land, (5) forest, and (6) green land. Green land refers to the open forest land, grass land, and urban green land (i.e., parks, lawns, and other urban open green spaces). They tend to have a disconnected and scattered distribution and share similar spectra characteristics in the study area, which means it is difficult to separate them from each other. And the land use classification was implemented by the maximum likelihood algorithm in ENVI 5.2 [32]. The results are raster maps with a resolution of 30 m. These grid maps of area proportion of each

Land Use/Cover Classification
Materials used for monitoring land use include cloud-free Landsat TM (for 2006) and Landsat OLI images (for 2013) in the same time phase. Data were downloaded from the United States Geological Survey (USGS) database [31]. Taking the land cover features of the CZT into account, a modified version of the land use classification was developed [3,16]. The categories include: (1) water, (2) bare soil, (3) built-up, (4) agricultural land, (5) forest, and (6) green land. Green land refers to the open forest land, grass land, and urban green land (i.e., parks, lawns, and other urban open green spaces). They tend to have a disconnected and scattered distribution and share similar spectra characteristics in the study area, which means it is difficult to separate them from each other. And the land use classification was implemented by the maximum likelihood algorithm in ENVI 5.2 [32]. The results are raster maps with a resolution of 30 m. These grid maps of area proportion of each land use were then reclassified into 1 kmˆ1 km resolution to compare with the spatial patterns of PM 10 concentrations simulated by the LUR model.

Landscape Pattern
Landscape metrics are algorithms that quantify a specific spatial configuration of various land use/covers [33,34]. Considering the diversity and heterogeneity of landscape [35,36], five representative landscape-level metrics characterizing the urban sprawl were calculated by Fragstats 4.0 based on the resampled land use maps with a resolution of 1 kmˆ1 km [37]. The metrics include Aggregation index (AI), Contagion index (CONTAG), Landscape shape index (LSI), Perimeter-area fractal dimension (PAFRAC), and Shannon's diversity index (SHDI). They can be expressed by Equations (1) to (5): where g ii and g ik are the number of like adjacencies (joins) between pixels of patch type (class) i obtained by the single-count and double-count method, respectively, and max-g ii is the maximum g ii . P i is the proportion of patch type (class) i. m and n are the number of patch types and classes present in the landscape. E* refers to the total edge length (meter) of the landscape, while A is the total area (m 2 ). a ij and p ij are area (m 2 ) and perimeter (meter) of patch ij. N is the total number of patches in the landscape. Among the above five landscape-level metrics, AI is an index measuring the amount of the maximum possible number of like adjacencies given any landscape composition. Higher AI values indicate more aggregative patches of the same type. A landscape with interspersed patch types will have relatively low CONTAG values. While a single patch type occupies a very large percentage of the landscape, the CONTAG is high. The greater the value of LSI, the more dispersed the patch types. If patches have simple geometric shapes, the PAFRAC will be relatively low. The greater the value of SHDI, the richer the land use types.

Predictor Variables
Similar to previous reported LUR modeling researches [24], correlation analysis is needed to identify possible predictors. The predictor variables identified in this study include area proportions of each land use, road length, and the distance to a nearest major road, as well as the population density. They are generated at buffers with different radius (50-5000 m) of each monitoring site, so the selection of buffer size plays an important role in determining the performance of the LUR model [24]. Moreover, the annual averages of Aerosol Optical Depth (AOD) measurements of the 23 monitoring sites were extracted from the MODIS Terra data (MOD04 Level 2 Collection 5, at a 10ˆ10 km spatial resolution) obtained by the US-based National Aeronautics and Space Administration (NASA) [39]. Other potential variables include elevation, urban fractal dimension, contagion, and surface meteorological elements (i.e., relative humidity, temperature, precipitation, and wind speed). Population and elevation statistical data were made available by the Statistical Bureau of Hunan Province and USGS [40], respectively. Meteorological data were released by the Hunan Meteorology Bureau [41]. All variables ( Table 1) were extracted by ArcGIS10.0.

Land Use Regression Modeling and Validation
A general LUR model can be defined as follows: PM 10,s " a 0`a1 X 1,s`a2 X 2,s`. . .`a n X n,s`µ (6) where PM 10,s is the estimation of the annual average PM 10 concentration, and is regarded as the dependent variables of site s, X i,s (i = 1, 2, . . . n) are independent variables, a k (k = 0, 1, 2, 3) are the regression coefficients estimated, and µ is the random error under the condition that the value of Σ s i"1 pPM 10,1 -P M 10, i q is minimized over the observations. In this study, LUR models were conducted using a supervised stepwise regression. In this process, we first separately entered each potential predictor derived from 2013 to identify which one could explain the largest variance. Second, we evaluated whether other added variables will lead to an increase larger than 1% in the adjusted R 2 . This procedure was repeated until no more variables entered the model. We also calculated the collinearity statistics, such as the tolerance and variance inflation factor (VIF). It should be noted that variables can only be entered in the model when the coefficients are statistically significant and there is no multicollinearity between these variables. Additionally, statistics including the model fitting R 2 , root mean square error (RMSE), standard deviation (SD), mean relative tolerance (MRT), as well as the CV R 2 were employed to evaluate the prediction ability and reliability of the LUR model of year 2013 based on the Leave-One-Out-Cross-Validation (LOOCV) method.
For visualization purposes, we created nearly 27,900 lattice points with a resolution of 1 kmˆ1 km, where PM 10 concentrations were then predicted using the LUR model of year 2013 based on PM 10 concentrations at those lattice points. Finally, a smooth surface was interpolated through the inverse distance weighting (IDW) method. It has to be noted that the PM 10 concentrations of year 2006 was also estimated by the LUR model of year 2013 in order to test the transferability of LUR models over time in this study.

LUCC and PM 10 Variations
To test the assumed link between LUCC and PM 10 variation during the urban sprawl process, we first obtained the area proportion of each LUCC and associated landscape metrics at a resolution of 1 kmˆ1 km for years 2006 and 2013. Second, raster calculators were used to extract variations of PM 10 concentrations and area proportion of each LUCC, as well as that of landscape metrics, which were then employed for quantitative comparison and correlation analyses. Pearson's correlations with P values were calculated using SPSS software (version 19.0) [42].

Land Use/Cover and Landscape Pattern in 2006 and 2013
Results of the land use/cover classification are presented in Figure 2a. The overall classification accuracy is 83.57% for the 2013 map and 80.04% for the 2006 map. As we can see, built-up areas are distributed mainly along the Xiangjiang River. From 2006-2013, their proportion increased by 4.79%, from 2984 km 2 to 3134 km 2 . These increased built-up lands are mostly converted from agricultural land and forest, whose area decreased 15 km 2 and 98 km 2 , respectively. The bare soil increased from 4.16% to 6.49% and green land decreased from 9.27% to 6.85%. The change of water area (from 556 km 2 to 543 km 2 ) is small.

LUCC and PM10 Variations
To test the assumed link between LUCC and PM10 variation during the urban sprawl process, we first obtained the area proportion of each LUCC and associated landscape metrics at a resolution of 1 km×1 km for years 2006 and 2013. Second, raster calculators were used to extract variations of PM10 concentrations and area proportion of each LUCC, as well as that of landscape metrics, which were then employed for quantitative comparison and correlation analyses. Pearson's correlations with P values were calculated using SPSS software (version 19.0) [42].

Land Use/Cover and Landscape Pattern in 2006 and 2013
Results of the land use/cover classification are presented in Figure 2a. The overall classification accuracy is 83.57% for the 2013 map and 80.04% for the 2006 map. As we can see, built-up areas are distributed mainly along the Xiangjiang River. From 2006-2013, their proportion increased by 4.79%, from 2984 km 2 to 3134 km 2 . These increased built-up lands are mostly converted from agricultural land and forest, whose area decreased 15 km 2 and 98 km 2 , respectively. The bare soil increased from 4.16% to 6.49% and green land decreased from 9.27% to 6.85%. The change of water area (from 556 km 2 to 543 km 2 ) is small.

LUR Model Development and Validation
As is shown in Table 2, the proportions of built up area, bare soil, road length, and relative humidity are highly significant predictors (p < 0.05) of PM 10 Table 3 summarizes model fitting and validation results for the LUR model of year 2013 and its transferability application evaluation in year 2006. It can be seen that the linearity between PM 10 concentrations and its explanatory variables are significant (Sig. F < 0.0001). The final model has a RMSE of 9.36 µg/m 3 . Meanwhile, the adjusted overall mean R 2 is 0.62, meaning the model can explain 62% of the PM 10 concentration variations. Comparatively, the mean R 2 of LOOCV is 0.69, and it ranges from 0.63 to 0.75 for each site, while the CV R 2 is also acceptable with a value of 0.56. Model-based SD and CV RMSE are 11.09 µg/m 3 and 10.86 µg/m 3 , respectively. The MRT for validation is 8.10% with the relative error of predictions varying from 0.04% to 22.81%. In addition, validation results based on seven monitoring sites in 2006 in Table 3 also confirm the transferability of the LUR model 2013 with the MRT at 7.24% and RMSE at 7.24 µg/m 3 .       Figure 4 shows the change in PM10 concentrations and area proportions of each land use from 2006 to 2013. In general, most PM10 concentration increase occurs around the inner CZT, where the forest disappeared drastically and construction expanded rapidly (e.g., the southern Changsha County, and northern Yuelu and Wangcheng districts). PM10 concentration increased strikingly with the growing area proportion of bare soil in the southern Changsha County and southwestern Shaoshan County. The northeastern Chaling County experienced a complex change in land use. For most grids, more than 25% of land has changed its original use. These places show a clear increase in PM10 concentrations. Moreover, results from correlation analysis (Table 4) further confirm effect of LUCC on air pollution. PM10 concentration variation has a positive relation with the proportion changes of built-up area, bare soil, and agricultural land, and is negatively correlated with that of forest and green land (p < 0.0001). The sensitivity sequence of these five types of land use from high to low is built-up > green land > bare soil > forest > agricultural land. Meanwhile, the proportion change of water area is found to be the weakest predictor of PM10 concentration variation (r = −0.023; p = 0.297).

Impacts of Landscape Change on PM10 Concentration Variation
Clearly, the increased LSI, PAFRAC, and SHDI and the decreased AI, and CONTAG relate to PM10 concentration increase in the east and south suburban and rural areas of the CZT ( Figure 5). In the newly-developed built-up area of the inner urban areas (i.e., the northern Yuelu district and the area adjacent to Wangcheng district, Yuetang district, and Xiangtan County), PM10 concentration increases with the growing AI and CONTAG and decreases with the increment of LSI and SHDI. Conversely, PM10 concentration shows a decline in eastern Ningxiang County and mid-western Xiangtan County where the AI and CONTAG increased greatly, but LSI and SHDI decreased. Moreover, results from correlation analysis (Table 4) further confirm effect of LUCC on air pollution. PM 10 concentration variation has a positive relation with the proportion changes of built-up area, bare soil, and agricultural land, and is negatively correlated with that of forest and green land (p < 0.0001). The sensitivity sequence of these five types of land use from high to low is built-up > green land > bare soil > forest > agricultural land. Meanwhile, the proportion change of water area is found to be the weakest predictor of PM 10 concentration variation (r =´0.023; p = 0.297).

Impacts of Landscape Change on PM 10 Concentration Variation
Clearly, the increased LSI, PAFRAC, and SHDI and the decreased AI, and CONTAG relate to PM 10 concentration increase in the east and south suburban and rural areas of the CZT ( Figure 5). In the newly-developed built-up area of the inner urban areas (i.e., the northern Yuelu district and the area adjacent to Wangcheng district, Yuetang district, and Xiangtan County), PM 10 concentration increases with the growing AI and CONTAG and decreases with the increment of LSI and SHDI. Conversely, PM 10 concentration shows a decline in eastern Ningxiang County and mid-western Xiangtan County where the AI and CONTAG increased greatly, but LSI and SHDI decreased. PM10 concentration increase in the east and south suburban and rural areas of the CZT ( Figure 5). In the newly-developed built-up area of the inner urban areas (i.e., the northern Yuelu district and the area adjacent to Wangcheng district, Yuetang district, and Xiangtan County), PM10 concentration increases with the growing AI and CONTAG and decreases with the increment of LSI and SHDI. Conversely, PM10 concentration shows a decline in eastern Ningxiang County and mid-western Xiangtan County where the AI and CONTAG increased greatly, but LSI and SHDI decreased.  Table 5 further demonstrates the correlations between the variations of landscape metrics and PM10 concentration changes. On one hand, changes of AI (r = −0.097) and CONTAG (r = −0.114) negatively correlate with the changes of PM10 concentrations (p < 0.0001). On the other hand, positive correlations between the changes of landscape metrics and PM10 concentrations follows a typical declined sequence of SHDI (r = 0.060, p = 0.004) and LSI (r = 0.046, p = 0.024). However, no statistically significant correlation between changes of PAFRAC and PM10 concentration (p = 0.866) was found.

Discussion
Fine-resolution monitoring data for air pollutants are rarely available for most countries in the world [22]. Prefecture-level cities in China, for example, had no continuous automatic air pollution monitoring system until 2013. In order to estimate air pollution in suburban and rural areas distant from monitoring sites in the CZT urban agglomeration, a LUR model, called LUR model of year 2013, was developed and its transferability across time was also validated in this study. The LOOCV R 2 (0.63-0.75) and absolute bias (MRT 0.04%-22.81%) of this developed LUR model demonstrate comparable reliability to previously-reported LUR models with R 2 ranging from 0.22 to 0.72 and absolute bias ranging from 17% to 22% [21,23].
Based on the data of consumption proportions of raw coal, petroleum, natural gas, clean energy (i.e., hydropower, nuclear power, wind power, etc.) and other sources in 2006 (68.51%, 12.87%, 0.66%, 14.87%, 3.09%, respectively) and 2013 (62.23%, 11.66%, 1.55%, 14.04%, 10.52%, respectively) of the CZT are [43] and given the efficient and conservative use of energy has been promoted by the 'Two Oriented Society' policy, we assume the change of energy production and consumption  Table 5 further demonstrates the correlations between the variations of landscape metrics and PM 10 concentration changes. On one hand, changes of AI (r =´0.097) and CONTAG (r =´0.114) negatively correlate with the changes of PM 10 concentrations (p < 0.0001). On the other hand, positive correlations between the changes of landscape metrics and PM 10 concentrations follows a typical declined sequence of SHDI (r = 0.060, p = 0.004) and LSI (r = 0.046, p = 0.024). However, no statistically significant correlation between changes of PAFRAC and PM 10 concentration (p = 0.866) was found.

Discussion
Fine-resolution monitoring data for air pollutants are rarely available for most countries in the world [22]. Prefecture-level cities in China, for example, had no continuous automatic air pollution monitoring system until 2013. In order to estimate air pollution in suburban and rural areas distant from monitoring sites in the CZT urban agglomeration, a LUR model, called LUR model of year 2013, was developed and its transferability across time was also validated in this study. The LOOCV R 2 (0.63-0.75) and absolute bias (MRT 0.04%-22.81%) of this developed LUR model demonstrate comparable reliability to previously-reported LUR models with R 2 ranging from 0.22 to 0.72 and absolute bias ranging from 17% to 22% [21,23].
Based on the data of consumption proportions of raw coal, petroleum, natural gas, clean energy (i.e., hydropower, nuclear power, wind power, etc.) and other sources in 2006 (68.51%, 12.87%, 0.66%, 14.87%, 3.09%, respectively) and 2013 (62.23%, 11.66%, 1.55%, 14.04%, 10.52%, respectively) of the CZT are [43] and given the efficient and conservative use of energy has been promoted by the 'Two Oriented Society' policy, we assume the change of energy production and consumption structure in the study region has been stable. This could be the reason why the estimated PM 10 concentrations in 2006 can be accurately predicted by the LUR model 2013 in this study, and also indicates the potential transferability of the LUR model 2013 across time in the CZT area.
Temporal variation in PM 10 concentrations suggests that the "Two Oriented Society" policy has positive effect on reducing PM 10 concentration. The overall mean PM 10 concentration in the CZT has decreased since it became a national comprehensive supporting trial area in the implementation of this policies. In this process, the local governments encourage building greening systems along the urban streets, afforesting waste hills or unclaimed lands, and converting low farmlands into lakes. Moreover, high-tech industrial and science parks are gradually replacing original disordered productive land use (i.e., industrial, warehouse, mining). These actions not only cut down the emissions of possible LUCC related pollution sources, but also blocked the pollutants dispersion. However, as a consequence of the three cities (Changsha, Zhuzhou, and Xiangtan City) joining together and the accelerated industrial development, air pollution in the central CZT is still very challenging. Efforts for emission reduction and air pollution prevention and treatment should especially be put on the inner urban agglomeration, and these actions could be implemented by the incorporations in land use, economic development, industrial layout, and traffic pattern among three cities.
The association between LUCC and PM 10 concentration variation clearly illustrates the potential contribution of land planning in reducing air pollution. The change of forest area shows a negative influence on PM 10 concentration that is strongly positively correlated with the increasing built-up area. This result confirms findings of Weng et al. and Stone [4,16], suggesting that land use strategies including creating urban growth boundaries (i.e., restricting peripheral spread of urban zones and the resulting vehicle increase) and protecting essential ecological sites would be effective in limiting PM 10 concentration growth [13]. Reduction of green space relates to PM 10 increase, corroborating its important role in mitigating air pollution. The positive relationship between PM 10 variation and bare soil proportion change may be induced by the ground dust. The tidal wave of migrant workers leaves a broad agricultural land out of cultivation, which may increase the PM 10 concentration. Additionally, straw burning is one of the major sources of PM 10 in the autumn.
For landscape, the changes of LSI, SHDI, and PAFRAC have positive correlation with the PM 10 concentration variation in suburban and rural areas. The increase of these landscape metrics could be significant signals of human disturbance to predominant, continuous patches of natural landscape [18]. Growing human activities are bound to increase energy and transport demands, which have positive effect on PM 10 concentration. On the other hand, the dominant landscapes in these areas covering over two-thirds of the study area are forest and green land, with high AI and CONTAG, and little pollutant source, hence the PM 10 concentration is rather low. This reconfirmed the results of Liu and Shen [44]. However, the newly developed built-up area is experiencing a decrease of LSI and SHDI whilst an increase of AI, CONTAG, and PM 10 concentration. A possible reason of this could be the concentration of buildings in a small area with simple geometric shapes and high traffic emissions. Moreover, the resultant air pollutant may be entrained in a re-circulatory system due to the "street canyons effect" [45]. All of these indicate that the landscape impact on PM 10 concentration in the study area is closely related to the development stage it is in.

Conclusions
This study developed a LUR model that coupled land use and meteorology factors to evaluate the effect of urban LUCC on air pollution variation during the urban sprawl of the CZT from 2006-2013. While the proposed LUR model of year 2013 was validated in terms of transferability in the CZT with relatively stable energy production and consumption structure across time, this study disclosed that the PM 10 concentrations in the CZT have generally decreased after the implementation of the "Two-Orientated society" policy since 2006. The urban LUCC, including land cover and landscape, influences PM 10 variations. Increases of built-up areas and decreases of forest areas during the period from 2006-2013 have considerable adverse impacts on PM 10 variation, while the effects of landscape pattern variation are only moderate, and these effects vary in newly-developed built-up areas and others. Compared to either the land use or landscape, the combination of them might be more useful in indicating PM 10 concentration levels. And these results imply that more serious consideration of reasonable land use configurations could be a promising way to improve the air quality in urban sprawl, and China's "Two Oriented Society" policy is a sustainable urban development strategy for developing countries to effectively control urban air quality.