Stoch Environ Res Risk Assess DOI 10.1007/s00477-012-0671-0

ORIGINAL PAPER

Modeling urban land use conversion of Daqing City, : a comparative analysis of ‘‘top-down’’ and ‘‘bottom-up’’ approaches

Wenliang Li • Changshan Wu • Shuying Zang

Ó Springer-Verlag Berlin Heidelberg 2012

Abstract During the past decades, Daqing City, China Keywords Urbanization Stochastic cellular automata has experienced unprecedented urban expansion due to the CLUE-S model Daqing rapid development of petroleum industry. With rapid urbanization and lack of strategic planning, Daqing is facing many socio-economic and environmental problems, 1 Introduction and it is essential to examine the process of urbanization, and to develop policy recommendations for sustainable During the past 20 years, many metropolitan areas in China development. To address this problem, this paper examined have experienced unprecedented expansion due to popu- the urbanization process of Daqing City through develop- lation growth and migration. Urban built-up areas in China ing two multi-level models: an integrated system dynamic have increased from 10,161 km2 in 1986 to 32,600 km2 in (SD) and CLUE-S model (SD-CLUES), and an integrated 2006, with an increment of 220.83 % (China Statistic SD and stochastic cellular automata model (SD-CA). Yearbook 2003; China Association of Mayors 2007). This Analysis of results suggests that these two models generate high-speed urbanization is associated with the rapid growth significantly different results. With the SD-CLUES model, of urban population. In particular, urban population has new urban developments are clustered in the downtown increased from 302 million in 1990 to 456 million in 2000, area or along major transportation networks, indicating and it is projected that in 2020, *900 million Chinese exogenous driving forces playing an important role in people will reside in urban areas (Song and Ding 2009). shaping urban spatial dynamics. With the SD-CA model, Simultaneously, the percent of urban population to the total on the contrary, the resultant new urban cells are spread population has increased from 26 % in 1990 to 36 % in over the entire study area, and associated with existing 2000, and to 50 % in 2010, and it is projected that urban urban areas. Further, visual comparisons and validations population percentage will reach 65 % in 2050 (Chen et al. indicate that the SD-CA model is a better alternative in 2009; Song and Ding 2009). Rapid urbanization in China is explaining the urbanization mechanism of Daqing City. In associated with high speed industrialization and a phe- addition, analysis of results suggests that the stochastic nomenal economic growth. China has maintained the factor in the CA model has significant impact on the fastest gross domestic production (GDP) growth rate (e.g. modeling accuracy. 9.6 % annually since 1987) in the world since the imple- mentation of the economic reform and open-door policies. As a result of rapid economic development during these & W. Li C. Wu ( ) S. Zang years, China was ranked as one of the four world’s largest Key Laboratory for Remote Sensing Monitoring of Geographic Environment, College of Province, Normal economies with a GDP of $5.88 trillion US dollars in 2010. University, Harbin, Heilongjiang 150025, China While rapid urbanization brought economic benefits and e-mail: [email protected] improved the quality of life, ill-planned urban growth also generated numerous challenging socio-economic issues W. Li C. Wu Department of Geography, University of Wisconsin-Milwaukee, (e.g. social and economic inequity, excess commuting, PO Box 413, Milwaukee, WI 53201, USA congestion) and environmental problems, such as air and 123 Stoch Environ Res Risk Assess water pollution, urban micro-climate alteration, excess of urbanization (Chomitz and Gray 1996; Huang et al. carbon emission, reduction of biodiversity, degradation of 2009; Irwin and Geoghegan 2001; Landis and Zhang 1998; surrounding ecological systems, and resource depletion Luo and Wei 2009; Mohapatra and Wu In Press; Nugroho (Barasa et al. 2011; Cao and Ye 2012; Guttikunda et al. et al. 2011; Wu et al. 2012). In addition to these regression- 2003; Lin and Ho 2003; Newman and Kenworthy 1999; based techniques, systems dynamic (SD) models proposed Pauleit and Duhme 2000; Pielke 2005; Stevens et al. 2007; by Forrester (1961, 1969) have also been applied to Song and Xu 2011, Tu et al. 2012; Van Metre and Mahler examine the driving forces of urban spatial dynamics 2005; Wang et al. 2012; Zang et al. 2011; Zhou et al. 2004). (Verburg et al. 2002; He et al. 2006; Neto de et al. 2006; In particular, due to urban expansion, excess commuting Han et al. 2009; Yu et al. 2011). The SD model has the and congestion have become major issues for many cities in ability to uncover complicated relationships among dif- China. Taking , China as an example, the average ferent driving forces within a system, and it can be commuting time is *60 min during rush hours, and it is employed to simulate a number of urban development estimated that the congestion cost per person is about $53 scenarios under different policy recommendations. Besides per month (Xinlang Auto 2009). Moreover, the conversion the macro-level models, micro-level models have been from rural to urban land uses has modified physical developed to simulate land use changes (e.g. conversion parameters of the earth surface, resulting in reduced bio- from rural to urban land uses) at individual locations. In diversity and degraded natural ecosystem functions (Wang particular, two types of micro-level models, top-down and et al. 2009; Yue et al. 2012; Zang et al. 2011). The process bottom-up techniques, have been successfully applied in of urbanization in a resource-based city is particular inter- modeling urban spatial dynamics. Specially, top-down esting to scholars and urban planners. Unlike other cities, models consider the land use conversion at an individual the development of a resource-based city is mainly driven location is mainly due to exogenous forces, instead of local by the exploration of its resources, as well as the develop- interactions (Verburg and Overmars 2009). Therefore, a ment of resource related industries. For the exploration of top-down model allocates the demands of a particular land high intensity resources (e.g. oil, coal, timer, etc.), there is use category to individual cells according to their relations always a tradeoff between economic benefit and environ- with exogenous forces. A widely used top-down urban land mental cost. In addition, the city always has to face the use change model is the conversion of land use and its challenge of economic transformation when the resources effects (CLUE) model developed by Verburg and his col- are depleted. As a typical resource-based city, Daqing in leagues (Verburg et al. 2002, 2006). Further, Verburg and Heilongjiang Province, China was established in 1959 fol- Overmars (2009) developed a revised model, the Dyna- lowing the discovery of oil wells. Since then, urban infra- CLUE model, and argued that a top-down modeling structures have been constructed around the explored oil approach is more appropriate for examining urban land use wells for continuous oil exploration, and the economy of conversion. Comparatively, bottom-up approaches assume Daqing is highly dependent on petroleum production. that complicated urban spatial dynamics are the results of Daqing has the largest oil field of China, which is also one local interactions, instead of exogenous forces. While some of the largest ones in the world. During the past decades, bottom-up approaches have incorporated regional envi- Daqing has experienced unprecedented urban expansion ronmental constraints and land use zoning policies, due to the rapid development of petroleum industry, and neighborhood effects have played a much important role in new urban development mainly located around the explored shaping urban spatial dynamics. Typical bottom-up urban oil wells. Because of rapid urbanization and generally lack dynamic simulation models include the cellular automata of strategic planning, Daqing is facing many socio-eco- (CA) (Clarke et al. 1997; Kamusoko et al. 2009; Li and nomic and environmental problems, and it is essential to Yeh 2002; Wu et al. 2010; Wu and Chan 2011; Zhang et al. examine the process of urbanization, and to develop policy 2011) and agent-based models (Evans and Kelley 2004; recommendations for sustainable development. Mena et al. 2011; Parker et al. 2003). For examining the process of urbanization, a number of Although both ‘‘top-down’’ and ‘‘bottom-up’’ approa- models have been developed in the literature, and they can ches have been widely applied in analyzing urban spatial be divided into two broad categories: macro-scale and dynamics at the micro level, so far no research has been micro-scale models (Irwin et al. 2009). The major objec- conducted for an empirical comparative analysis of these tives of macro-scale models are to examine exogenous two types of models for a specific urban development drivers (e.g. socio-economic, political, or biophysical fac- process, especially for a resource-based city like Daqing. tors) of urbanization, and to predict the amount of land use These models, moreover, are significantly different in changes in aggregated geographic regions. Regression terms of the forces of urban land use conversion at the techniques, such as econometric models and panel data micro level for different urban development process, and analysis, have been applied to examine the driving forces consequently the resultant urban spatial dynamics should 123 Stoch Environ Res Risk Assess be highly different. Therefore, this study attempted to take maps as references. The overall Kappa values for all the Daqing, China, as an example to explore which model, the classification results are over 0.85, indicating that these ‘‘top-down’’ or ‘‘bottom-up’’ approach, is suitable to land use maps have satisfactory accuracies for further explain the urbanization mechanism of an oil resource- analysis. based city. This paper is aimed to conduct an empirical Besides the remote sensing imagery, digital elevation comparison of top-down and bottom-up models in Daqing, data with a 90 m resolution was obtained from the Global China. Then these models were validated using visual Topography Database of the Consultative Group on Inter- comparisons and computer-based validation techniques. national Agricultural Research (CGIAR) Consortium for Further, with the CA model, the impact of stochasticity on Spatial Information (http://srtm.csi.cgiar.org). Soil maps model performance was examined, and a suitable random were acquired from the Institute of Soil Science, Chinese effect was recommended. Academy of Sciences (ISSCAS). Moreover, transportation network data (e.g. railway, provincial and rural roads) were obtained from the Surveying and Mapping Department of 2 Study area and data Heilongjiang Province, China. Further, related socio-eco- nomic data such as population estimates and GDP values Daqing City, located in Heilongjiang Province, China, was for the study area were obtained from the Daqing Statistics chosen as the study area. Daqing is situated between the Yearbooks of 2001, 2006, and 2011. latitudes of 45°460–46°550 N and longitudes of 124°190– 125°120 E (Fig. 1). It has a geographic area of 5,144 square kilometers, and is comprised of five sub-districts including 3 Methods Ranghulu, Saertu, Longfeng, Honggang, and . The population of Daqing is *1.31 million and the average To examine the urban spatial dynamics of the Daqing City, GDP was about $15,934 per person in 2010, ranked the first an integrated macro-scale and micro-scale approach was in the and the eighth in the mainland employed. At the macro scale, we developed a system China (Ministry of Housing and Urban-Rural Development dynamic model to explore the role of social, economic, of People’s Republic of China, and Gohighfund 2011). political, and biophysical factors on urban dynamics. Fur- For analyzing the process of urban dynamics in Daqing, ther, two micro-scale models, a CLUE-S model and a CA Landsat Thematic Mapper (TM) images acquired in 2000 model, were developed as the top-down and bottom-up and 2005, and SPOT images taken in 2010 were obtained approaches respectively. from the China Remote Sensing Satellite Ground Station and the Second National Land Use Survey Office of China. 3.1 Macro-scale analysis: SD model These images were geo-rectified and mosaicked with high- resolution (with a scale of 1:5) topographic maps as ref- In order to examine exogenous drivers of urbanization and erences. Six land use classes, agricultural land, forest, predict the demand of urban lands at aggregated geo- grassland, built-up, water, and barren land, were derived graphic scales, we developed an SD model using the using the unsupervised classification function provided by Vensim PLE 5.10 program, a commercial software pack- the ERDAS Imagine 9.3 software. In order to further age developed by the Ventana Systems, Inc. To implement improve the classification accuracy, manual interpretation this model, we identified two groups of driving forces, land and digitization were conducted with historical land use use management policies and socio-economic factors, as

Fig. 1 Daqing City, Heilongjiang Province, China

123 Stoch Environ Res Risk Assess major exogenous variables affecting urban dynamics. 3.2 Micro-scale analysis These driving forces were identified following the recom- mendations from the International Geosphere and Bio- For the micro-scale urban dynamic simulation, we devel- sphere Program (IGBP) and the International Human oped a top-down (CLUE-S) and a bottom-up (CA) model. Dimensions Program (IHDP) (Turner et al. 1995; Nunes These two models were coupled with the macro-scale SD and Auge 1996; Vellinge 1998). With all the identified method to form two multi-level models: SD-CLUES and socio-economic and land use related driving forces, we SD-CA. divided the SD model into two subsystems, land use sub- system and socio-economic subsystem. The internal 3.2.1 CLUE-S model structures of these two sub-systems are shown in Fig. 2. Within the land use subsystem, the major components are At the micro scale, the CLUE-S model was first established to the six land use types (e.g. agricultural land, forest, simulate urban spatial dynamics. As a top-down model, grassland, water, built-up, and barren land), and the chan- CLUE-S allocates the land use demands estimated from the ges of each land use demand are dependent on other land SD approach into specific locations based on conversion use types and are also affected by the driving forces in the rules. In particular, the CLUE-S model decides whether to socio-economic subsystem. Further, the socio-economic assign a land use type to a specific location based on three subsystem models the impact of socio-economic factors, criteria, including (1) global land use probability, (2) land use such as GDP growth, population growth, and technology change elasticity (ELAS) value, and (3) neighborhood effects. progress, etc., on land use changes. In addition, a feedback For calculating the global land use probability, fourteen loop structure was constructed to examine the relationship variables, including elevation, slope, soil types (e.g. black among each component in the system (see Fig. 2). For soil, black and calcium soil, sandy soil, meadow soil, example, the quantity of built-up lands is highly influenced swamp soil, and saline-alkali soil), distance to rivers and by population and economy related variables. With high ponds, distance to railway, distance to provincial roads, speed population and economy growth, much more resi- distance to rural roads, distance to the nearest town, and dential and commercial land uses are needed to accom- distance to county boundary were chosen as the driving modate the growing population and the developments of factors to examine the probability that a cell belonging to a secondary and tertiary industries. As a result, the quantity specific land use type. A logistic regression analysis was of built-up lands will increase to meet the demands. In the conducted to examine the relation between land use prob- model, therefore, their relationships were set as: Population ability and a number of driving forces (see Eq. 1). ? ? ? growth rate ? Population ? Built-up demand ? - ? ? ? Pk Built-up and GDP growth rate ? GDP ? Secondary Log ¼ a0 þ a1X1;k þ a2X2;k þþanXn;k ð1Þ 1 P and tertiary industries??Built-up?. With the socio-eco- k nomic and land use information as inputs, we constructed where Pk indicates the probability of a cell being assigned to a and calibrated the SD model. As a result of the SD model, land use category k (k = 1..6), Xj,k indicates the jth driving aggregated land use demand for each land use category was factors of land use type k,andn represents the total number of estimated. factors. In addition to land use probability information, land

Fig. 2 Framework of the SD model 123 Stoch Environ Res Risk Assess use change ELAS for a particular land use type was also The resultant global urban expansion probability is as calculated from the historical land use conversion rate. The same as the global land use probability for the urban land resultant ELAS value has a range from 0 to 1, indicating the use type employed in the CLUE-S model. In addition to the degree of difficulty of converting from a land use type to global probability, the neighborhood effect Ni, the most other types. That is, an ELAS of 1 indicates that it is almost important factor in the CA model, was calculated through impossible to convert this type oflandusestoothers.Further, dividing the number of urban cells within the neighborhood the neighborhood effects were considered and incorporated in by the total number of cells in a neighborhood (e.g. 3 by 3 the CLUE-S model. To examine the impact of neighborhood cells). The resultant Ni, therefore, represents the impact of effects on land use conversion, the enrichment factor for each neighboring land uses on the land conversion probability of land use type was calculated as follows (Verburg et al. 2004). a particular cell. Furthermore, two constraint factors (Ci), slope and water body, were incorporated in the CA model. nk=n F ¼ ð2Þ In particular, a cell with a slope of 22.5° and higher, or k N =N k identified as a water body was excluded to be developed where Fk is the enrichment factor of land use type k in a into urban lands. Finally, a random factor (Ri) was neighborhood (e.g. 3 by 3 cells), nk represents the number incorporated into the CA model due to the stochastic of cells with land use type k within the neighborhood, n is characteristics of the urban spatial dynamics. The random the total number of cells within the neighborhood, Nk is the factor was calculated as follows (see Eq. 5). total number of cells with land use type k in the study area, a Ri ¼ 1 þðln cÞ ð5Þ and N is the total number of cells in the study area. With the enrichment factor for each land use type and each cell, where c is a random number between 0 and 1, and a is the we applied a logistic regression analysis method to control parameter ranging from 1 to 10. With all the cal- examine the effect of neighborhood on the probability of culated factors, including the global probability Pi, land use conversion (see Eq. 3). Neighborhood effect Ni, constraint factors Ci, and random factor Ri, the probability of a cell being converted to urban Qk Log ¼ b0 þ b1F1 þ b2F2 þþbnFn ð3Þ lands was derived as the product of these factors. 1 Qk where Qk is the neighborhood-based probability that a cell 3.3 Modeling result validation being devoted to land use type k, Fi is the enrichment factor for land use type i(i = 1..6), and bi is the regression With the simulated results from the coupled SD-CLUES coefficient. and SD-CA models and the 2005 and 2010 land use maps With the global land use probability, ELAS values, and derived from classifying the Landsat TM and SPOT images neighborhood effects, we applied the CLUE-S model to as ground truth data, we applied multiple validation allocate land use demands to each cell. Then the allocated approaches to assess the modeling accuracy, including (1) areas of each land use type were compared with the demands pixel matching, (2) spatial and feature pattern recognition estimated using the SD model, and adjusted iteratively until (Torrens 2011). In particular, pixel matching evaluates the the allocated geographic areas for each land use type equals to proportion of agreement between the simulated and the demand estimated by the SD model. observed results through a pixel by pixel comparison. Although it is considered the simplest and most straight- 3.2.2 Cellular automata (CA) model forward approach, the pixel matching method ignores the overall spatial pattern, and the results can be affected sig- The other micro-scale urban spatial dynamics simulation nificantly by locational errors. To address this problem, a model is the CA model, a widely accepted ‘‘bottom-up’’ multi-resolution procedure proposed by Costanza (1989) approach. The general structure of a CA model can be was utilized in this study through comparing the observed expressed as follows. and simulated maps over different resolutions. The good- ness of fit at different sampling window sizes can be cal- U ¼ f ðP ; N ; C ; R Þð4Þ i i i i i culated using the following formula: P where Ui is the probability of cell i being converted to P p ai;jb tw i¼1 j i;jj urban land uses, Pi is the global urban expansion j¼1 1 2w2 probability, Ni is the neighborhood effect, Ci is the Fw ¼ ð6Þ tw constraint factor, and Ri is the random factor. To determine the global urban expansion probability Pi,we where Fw is the goodness of fit at the sampling window employed the logistic regression analysis method (see size w (w is the linear dimension of a sampling window),

Eq. 1) to be consistent with the CLUE-S model. ai,j is the total number of cells with land use type i in the 123 Stoch Environ Res Risk Assess sampling window j (with a size of w) of the simulated map, 4 Results bi,j is the total number of cells with land use type i in the sampling window j of the reference map, p is the total 4.1 Macro-scale model number of different land use types, and tw is the total number of sampling windows with a size of w in the maps. With the land use information and socio-economic vari- With the goodness of fit Fw at each window size, the ables acquired in 2000 and 2005, an SD model was con- overall fit between the simulated map and reference map structed and calibrated. Then this model was applied to can be calculated with a weighted average of the fit over all predict the demand of each land use type in 2010. Results window sizes (see Eq. 7). (see Table 2) indicate that the SD model performs rea- P cðw1Þ sonably well, with the relative errors for all land use types Pw Fwe Ft ¼ ð7Þ \8 %. In particular, for built-up lands, the estimation error ecðw1Þ w is relatively low (-4.08 %), indicating that this model has where Ft is the overall fit and c is a constant that deter- a satisfactory accuracy, and can serve as the macro-scale mines the weight given to sampling window size w.If model for estimating the aggregated urban land use c equals to 0, all sampling window sizes are given the same demand. Table 2 also indicates that from 2005 to 2010, weight; whereas if c equals to 1, only the goodness of fits Daqing has experienced a rapid urban expansion, with a with small window sizes are considered important. In this significant increase of built-up lands at the cost of agri- study, c is set to be 0.1 following the guidance of Costanza cultural land, grassland, and water body. Specially, the (1989). geographic area of the built-up land has increased from Further, in order to evaluate the compactness and 32,380 ha in 2005 to 50,224 ha in 2010, with an increment complexity of urban land uses, landscape metrics mea- of 55.11 %. Simultaneously, rural lands including agri- surements were employed to examine the spatial pattern. In culture land, grassland, and water body have decreased particular, five indices, number of patches (NP), aggrega- about 2.16, 14.07, and 27.84 % respectively. As the tion index (AI), area-weighted mean fractal dimension objective of this study is to model urban dynamics, rural index (AWMPFD), landscape shape index (LSI), and patch land use types were merged into a single land use type, and edge density (ED), were calculated using the Fragstat 3.3, only two land use classes, urban and rural lands, were an open source software package (see Table 1). employed for further analyses.

Table 1 Landscape metrics Name Description

Number of patches (NP) Number of patches of a particular land use type, indicating the level of fragmentation of a study area Aggregation index (AI) Percentage of like adjacencies (joins with the same land use type) for a pixel, indicating the degree of contagion Area-weighted mean fractal dimension index Calculated as the fractal dimension weighted by the patch area, indicating the complexity of (AWMPFD) shapes Landscape shape index(LSI) A measure of patch disaggregation, calculated as the total length of edges divided by the minimum total length possible Patch edge density (ED) Calculated by dividing the total lengths of all edge segments by the total landscape area, indicating the patch complexity

Table 2 Prediction accuracy of the SD model (assessed using 2010 reference data) Years Agriculture land Forest Grass land Water body Built-up Barren land

Start 2005 326,948 3,616 19,892 55,628 32,380 75,988 Reference data 2010 319,896 4,026 17,094 40,141 50,224 83,071 Simulation results 2010 322,322 4,288 18,399 43,267 48,175 78,001 Relative error (%) 0.76 6.51 7.60 7.79 -4.08 -6.10

123 Stoch Environ Res Risk Assess

4.2 Micro-level model Table 4 Enrichment factors (F) and beta-coefficients derived from the logistic regression 4.2.1 Top-down model: CLUE-S Enrichment factors Urban

To implement the CLUE-S model, a logistic regression Non-urban factor 1.61035 model was constructed to derive the land use probability Urban factor 0.96849 map. Results (see Table 3) indicate that the probability of a Constant -4.88124 cell belonging to urban land is associated with elevation, ROC 0.719 slope, soil types (e.g. black and calcium soil, sandy soil, and meadow soil), distance to railway, distance to pro- vincial roads, distance to rural roads, distance to the nearest neighborhood effects should be considered in modeling town, and distance to county boundary. The resultant ROC urban spatial dynamics. With the land use demands esti- value for the model is 0.862, indicating the selected driving mated by the SD model, the land use probability map, forces can successfully explain the distribution of urban ELAS value, and the neighborhood effects, a CLUE-S lands in the study area. After generating the land use model was implemented. Results (see Fig. 3b) indicate probability map, we also calculated the land use change that, with the CLUE-S model, urban expansion mainly ELAS based on the land use conversion rate from 2000 to happens along transportation networks, such as railway, 2005. The resultant ELAS value equals to 1 for urban roads, etc. Moreover, significant urban expansion also lands, and 0.55 for rural lands. This indicates that rural takes place in or around the downtown of Daqing City lands can be converted to urban lands, while urban lands located in the northeastern part. These results clearly cannot be transformed to rural lands. Based on the ELAS indicate that, with the CLUE-S model, many exogenous values, we constructed the conversion matrix, and assigned driving forces have played important roles in shaping a value of zero to urban cells, and one to rural cells, future urban dynamics. indicating that only rural lands can be converted to urban lands. Further, the neighborhood effects were estimated 4.2.2 Bottom-up model (CA) through the logistic regression analysis with enrichment factors for each land use type as inputs. Results (see With the land use information and driving factors for the Table 4) indicate that a significantly positive association years of 2000–2005, we constructed and calibrated the CA exists among urban cells (ROC equals to 0.719), and the model. Further, this model was applied to simulate the urban growth dynamics in 2010. To construct the CA model, land use demands estimated by the SD model were Table 3 Driving factors and beta-coefficients derived from the employed to control the total areas of urban and rural lands, logistic regression and the rural-to-urban conversion probability was gener- Driving factors Beta-coefficients ated through integrating the global urban expansion prob- ability (see Table 3), the neighborhood effect, the Elevation (m) 0.10736 constraint factors, and the random factor. For this research, Slope (degree) 0.23355 we set the control parameter a as 1 following the results of Soil Black soil – other studies. Results of the CA model (see Fig. 3c) show Black and calcium soil 1.36258 that the simulated urban expansion majorly happens around Sandy soil 1.52491 the existing urban centers, indicating that the neighborhood Meadow soil 1.19111 effects have played an important role in driving urban Swamp soil – growth. Saline-alkali soil – Distance to rivers and ponds (m) – 4.3 Comparison and validation of the SD-CLUES Distance to railway (m) -0.00004 and SD-CA models Distance to provincial roads (m) 0.00004 Distance to rural roads (m) -0.00057 With the simulated urban spatial dynamics from the SD- Distance to the nearest town (m) -0.00005 CLUES and SD-CA models, visual comparison and model Distance to county boundary (m) -0.00019 validation have also been performed to examine the trade-offs Constant -17.86102 between these two models. With the SD-CLUES model (see ROC 0.861 Fig. 3b), it appears that the majorities (over 70 %) of simu- Not significant at the 95 % significant level and values are not lated urban expansions are clustered in or around the down- included town of Daqing City (the northeastern part). Moreover, most 123 Stoch Environ Res Risk Assess

Fig. 3 Comparison of simulated and reference urban land uses a reference map in 2005, b simulated map with the SD-CLUES model in 2010, c simulated map with the SD- CA model in 2010, d reference map in 2010

of the simulated urban cells follow the major transportation city, Daqing was initially established around oil wells for networks (e.g. provincial roads, railways, etc.), indicating that convenient oil exploration and production, and therefore the exogenous driving forces have significant impact on urban generating a unique multi-nuclei city pattern. With the high growth patterns. On the contrary, quite different spatial pat- speed urban expansion, new urban constructions in Daqing terns have been found in the simulated map derived from the are highly clustered around existing urban infrastructures, SD-CA model (see Fig. 3c). In particular, it can be observed which were constructed around explored oil wells several that the simulated urban growth is spread over the study area, decades ago. Like many other resource-based cities in China, and most of the growths are around existing urban infra- Daqing is in the process of transforming from a resource- structures in 2005. This result indicates that, with the SD-CA dependent economy to a diversified economy because of the model, the neighborhood effects dominate the simulation incoming issues of oil resource depletion. Associated with the results. This is likely to be consistent with the characteristics high demands of economic growth and diversified industry of urbanization process of Daqing City. As a resource-based development, rapid urban expansion occurs in the fringe of 123 Stoch Environ Res Risk Assess

Table 5 Validation results of Validation approaches 2010 reference data SD-CLUES SD-CA SD-CLUES model and SD-CA model model model Pixel matching Pixel by pixel comparison 93.18 % 93.89 %

Pattern Ft 0.9941 0.9981 recognition NP 881 483 533 ED 5.0727 3.8013 4.1101 AWMPFD 1.2112 1.1951 1.2024 LSI 29.0756 21.7956 23.5156 AI 94.5577 95.8369 95.5525

Fig. 4 Simulated 2010 urban land maps using CA model with different random parameters (a) the city. Therefore, unlike many other cities in which trans- goodness of fit (Ft) (see Table 5) also indicates that the portation networks drive urbanization, Daqing’s urban infra- SD-CA model has a slightly better fit. Further, it can be structures are more associated with the locations of existing observed that the landscape metrics values, including num- urban infrastructures, as well as oil wells, and the urban ber of patches (NP), area-weighted mean patch fractal distribution pattern of Daqing has also been transformed from dimension (AWMPFD), patch edge density (ED), and scattered spots to a zonal landscape through connecting these landscape shape index (LSI), derived from the SD-CA model individual spots. Therefore, the map generated from SD-CA are more similar to those calculated with the 2010 reference model is more close to the reality. Further, a visual compar- map. All of these results indicate that the SD-CA model has ison indicates that the results from the SD-CA model appear slightly outperformed the SD-CLUES model. tobemoresimilartotheobservedmapin2010(Fig.3d), although more rigorous comparisons are necessary. In addition to visual comparisons, we also assessed the accuracy of the simulation results with the pixel-matching and spatial pattern recognition techniques. Results (see Table 5) indicate that, with the pixel matching measure, the accuracy of the SD-CLUES (93.18 %) is slightly lower than that of the SD-CA model (93.89 %). Moreover, the multi- resolution goodness of fit (Fw) shows that the SD-CA model performs better than SD-CLUES model for all resolutions, and the accuracy increases as the resolution decreases. Fig. 5 Accuracy assessment (pixel-by-pixel matching) of the SD-CA A comparison of the overall accuracy of the multi-resolution model with different random factors (a changes from 0 to 10) 123 Stoch Environ Res Risk Assess

When a random factor was added, the overall accuracy continues to drop dependent on the degree of randomness (see Fig. 5). When the spatial pattern was measured, how- ever, the introduction of random factors improves the mod- eling accuracy when the control factor has a value of 1 or 2 (see Figs. 6, 7). For example, when the control factor has a value of 1, the Ft value is the highest (0.9981). Moreover, the landscape metric measures are also similar to those in the reference map. As a summary, the sensitivity analysis showed that, with the pixel-by-pixel measure, the SD-CA model without a random factor performs the best. However,

Fig. 6 Overall accuracy of the model goodness of fit (Ft) with spatial pattern recognition measures, the stochastic model with appropriate randomness assignment (e.g. control 4.4 Effects of the random factor in the SD-CA model factor equal to one in this study) may generate better results.

Due to the importance of the stochastic factors in the SD-CA model, we also performed a sensitivity analysis to examine 5 Conclusions and discussion the effects of the random factors on the modeling results. In detail, we let the control parameter a changes from 0 (without This paper developed two multi-scale modeling approa- random effect) to 10 (with the highest random effect), and ches, the SD-CLUES and SD-CA models, to examine the evaluate the resultant urban spatial dynamics (see Fig. 4). process of oil resources based urban spatial dynamics of Validation results indicate that, with the pixel-by-pixel Daqing City, China. In particular, the SD-CLUES connects comparison measure, the SD-CA model without a random the SD model (a macro-scale approach) with the CLUE-S factor generated the best results with an accuracy of 94.02 %. model (a top-down micro-level method); and the SD-CA

Fig. 7 Landscape metric indices and overall accuracy of for the reference and simulated urban maps in 2010 (the first point shows the value of the reference map in 2010, and others show the values of the simulated maps with the value of a varying from 0 to 10)

123 Stoch Environ Res Risk Assess model integrates the results of the SD model with the the trans-boundary river Sio catchment using remote sensing and bottom-up CA modeling approach. Further, the results of GIS. Ann Gis 17(1):73–80 Cao K, Ye X (2012) Coarse-grained parallel genetic algorithm these two models were compared through visual exami- applied to a vector based land use allocation optimization nation and computer-based validation. Analysis of results problem: the case study of Tongzhou Newtown, Beijing, China. suggests several major conclusions. Stoch Environ Res Risk Assess. doi:10.1007/s00477-012-0649-y First, significantly different results have been generated Chen MX, Lu DD, Zhang H (2009) Comprehensive evaluation and the driving factors of China’s urbanization. Acta Geograpica using the SD-CLUES and SD-CA models. As a top-down Sinica 64:387–398 approach, the simulated urban growth from the SD-CLUES China Association of Mayors (CAM) (2007) Urban development model clustered around the downtown of Daqing City and report of China (2006). China City Press, Beijing along the major transportation networks. It indicates a China Statistic Yearbook (2003).China statistic press. Beijing Chomitz KM, Gray DA (1996) Roads, land use, and deforestation: a strong influence of exogenous driving forces on urban spatial model applied to Belize. World Bank Econ Rev 10(3): spatial dynamics. With the SD-CA model, on the contrary, 487–512 simulated urban cells are spread over the entire study area, Clarke KC, Hoppen S, Gaydos L (1997) A self-modifying cellular and closely associated with existing urban infrastructures. automation model of historical urbanization in the San Francisco Bay Area. Environ Planning B 24:247–261 This indicates that the resultant urban dynamics derived Costanza R (1989) Model goodness of fit-a multiple resolution from the SD-CA model are majorly due to the neighbor- procedure. Ecol Model 47(3–4):199–215 hood effects, instead of the exogenous factors. Analysis of de Neto ACL, Legey LFL, Gonza´lez-Araya MC, Jablonski S (2006) results indicates that the map generated by the SD-CA A system dynamic model for the environmental management of the Sepetiba Bay watershed, Brazil. Environ Manag 38:879–888 model is more close to the reality as it represents the Evans TP, Kelley H (2004) Multi-scale analysis of a household level characteristics of the urbanization process in Daqing City, agent-based model of land cover change. J Environ Manag in which urban expansion is clustered around existing 72:57–72 urban infrastructures, as well as oil well locations. Sec- Forrester JW (1961) Industrial dynamics. Pegasus Communications, Waltham ondly, through visual comparisons and computer-based Forrester JW (1969) Urban dynamics. The Massachusetts Institute of validations, we found that the SD-CA model generated Technology Press, Cambridge slightly better results with the pixel-by-pixel comparison Guttikunda SK, Carmichael GR, Calori G, Eck C, Woo JH (2003) The and spatial pattern recognition approaches. Finally, it contribution of megacities to regional sulfur pollution in Asia. Atmos Environ 37:11–22 appears that the stochastic factor in the SD-CA model has Han J, Hayashi Y, Cao X, Imura H (2009) Application of an significant impact on the modeling results, and the choice integrated system dynamics and cellular automata model for of such a factor should be carefully examined. urban growth assessment: a case study of , China. Modeling the process of urbanization has been an Landsc Urban Planning 91:133–141 He CY, Okada N, Zhang Q, Shi P, Zhang J (2006) Modeling urban important research topic in the geography and urban expansion scenarios by coupling cellular automata model and planning literature during the past decades. In particular, system dynamic model in Beijing, China. Appl Geogr 26: with the advances of computer-based modeling capabili- 323–345 ties, bottom-up simulation approaches have emerged, and Huang B, Xie C, Tay R, Wu B (2009) Land-use-change modeling using unbalanced support-vector machines. Environ Planning B consequently debates on whether these approaches are 36(3):398–416 appropriate for urban growth analysis have been carried Irwin EG, Geoghegan J (2001) Theory, data, methods: developing out. Although this paper sheds some lights on the trade-offs spatially-explicit economic models of land use change. J Agric between top-down and bottom-up models, systematic Ecosyst Environ 85(1–3):7–24 Irwin EG, Jayaprakash C, Munroe DK (2009) Toward a comprehen- examinations of these two modeling approaches are nec- sive framework for modeling urban spatial dynamics. Landsc essary as future research. Further, more studies are neces- Ecol 24:1223–1236 sary on evaluating the impacts of exogenous driving forces Kamusoko C, Aniya M, Adi B, Manjoro M (2009) Rural sustain- and local effects on urban spatial dynamics. ability under threat in Zimbabwe-Simulation of future land use/ cover changes in the Bindura based on the Markov- cellular automata model. Appl Geogr 29:435–447 Acknowledgments This research was supported by the National Landis JD, Zhang M (1998) The second generation of the California Natural Science Foundation of China (Nos. 41030743, 41171322). urban features model: part 2: specification and calibration results We would like to acknowledge the anonymous reviewers for their of the land-use change submodel. Environ Planning B constructive and valuable suggestions on the earlier drafts of this 25(6):795–824 manuscript. Li X, Yeh AG (2002) Neural-network-based cellular automata for simulating multiple land use changes using GIS. Int J Geogr Inf Sci 16(4):323–343 Lin GC, Ho SP (2003) China’s land resources and land-use change: References insights from the 1996 land survey. Land Use Policy 20:87–107 Luo J, Wei DYH (2009) Modeling spatial variation of urban growth Barasa B, Majaliwa JGM, Lwasa S, Obando J, Bamutaze Y (2011) patterns in Chinese cities, the case of . Landsc Urban Magnitude and transition potential of land-use/cover changes in Planning 91:51–64

123 Stoch Environ Res Risk Assess

Mena CF, Walsh SJ, Frizzelle BG, Yao X, Malanson GP (2011) Land Verburg PH, Overmars KR (2009) Combining top-down and bottom- use change on household farms in the Ecuadorian Amazon: up dynamics in land use modeling: exploring the future of design and implementation of an agent-based model. Appl Geogr abandoned farmlands in Europe with the Dyna-CLUE model. 31(1):210–222 Landsc Ecol 24:1167–1181 Ministry of Housing and Urban-Rural Development of People’s Verburg PH, Soepboer W, Limpiada R, Espaldon MVO, Sharifa M, Republic of China, and Gohighfund (2011) The Chinese private Veldkamp A (2002) Land use change modeling at the regional capital investment report, http://www.gohighfund.com/articles/ scale: the CLUE-S Model. Environ Manag 30(3):391–405 21. Accessed Nov 2012 Verburg PH, Eck JR, Nijs TC, Schot MP (2004) Determinants of Mohapatra R and Wu C Modeling urban growth at a micro level: a land-use change patterns in the Netherlands. Environ Planning B panel data analysis. Int J Appl Geospatial Res (in press) 31:125–150 Newman P, Kenworthy JR (1999) Sustainability and cities: over- Verburg PH, Schulp CJE, Witte Veldkamp NA (2006) Downscaling coming automobile dependence. Island Press, Washington DC of land use change scenarios to assess the dynamics of European Nugroho SB, Fujiwara A, Zhang JY (2011) An empirical analysis of landscapes. Agric Ecosyst Environ 114:39–56 the impact of a bus rapid transit system on the secondary Wang SJ, Li J, Wu DQ, Liu J, Zhang K, Wang RQ (2009) The pollutants in the roadside areas of the Transjakart corridors—a strategic ecological impact assessment of urban development structural equation model and artificial neural network approach. policies: a case study of Rizhao City, China. Stoch Environ Res Stoch Environ Res Risk Assess 25(5):655–669 Risk Assess 23(8):1169–1180 Nunes C, Auge JI (1996) Land use and land cover change (LUCC) Wang QS, Yuan XL, Ma CY, Zhang Z, Zuo J (2012) Research on the implementation strategy. IGBP Report No. 48 and IHDP Report impact assessment of urbanization on air environment with No. 10 urban environmental entropy model: a case study. Stoch Environ Parker DC, Manson SM, Janssen MA, Hoffmann MJ, Deadman P (2003) Res Risk Assess 26(3):443–450 Multi-agent systems for the simulation of land-use and land-cover Wu YT, Chan KY (2011) Optimal design and impact analysis of change: a review. Ann Assoc Am Geogr 93(2):314–337 urban traffic regulations under ambient uncertainty. Stoch Pauleit S, Duhme F (2000) Assessing the environmental performance of Environ Res Risk Assess 25(2):271–286 land cover types for urban planning. Landsc Urban Planning 52:1–20 Wu DQ, Liu J, Wang SJ, Wang RQ (2010) Simulating urban Pielke RA (2005) Land use and climate change. Science 310: expansion by coupling a stochastic cellular automata model and 1625–1626 socioeconomic indicators. Stoch Environ Res Risk Assess Song Y, Ding C (2009) Smart urban growth for China. Lincoln 24(2):235–245 Institute of Land Policy, Cambridge Wu K, Ye X, Fang Z, Zhang H (2012) Impact of land use/ land cover Song HM, Xu LY (2011) A method of urban ecological risk change and socioeconomic development on regional ecosystem assessment: combing the multimedia fugacity model and GIS. services: the case of the fast-growing metropolitan Stoch Environ Res Risk Assess 25(5):713–719 area, China. Cities. doi:10.1016/j.bbr.2011.03.031 Stevens D, Dragicevic S, Rothley K (2007) iCity: GIS-CA modelling Xinlang Auto (2009) Futian index: economic cost of Beijing’s traffic tool for urban planning and decision making. Environ Model congestion is ranked the first, http://auto.sina.com.cn/news/2009- Softw 22:761–773 12-25/1039553292.shtml. Accessed Jan 2012 Torrens PM (2011) Calibrating and validating cellular automata Yu W, Zang S, Wu C, Liu W, Na X (2011) Analyzing and modeling models of urbanization. In: Yang X (ed) Urban remote sensing: land use land cover change (LUCC) in the Daqing City China. monitoring, synthesis and modeling in the urban environment. Appl Geogr 31(2):600–608 Wiley-Blackwell Press, Chichester, pp 335–345 Yue WZ, Liu Y, Fan P, Ye XY, Wu CF (2012) Assessing spatial Tu XJ, Zhang Q, Singh VP, Chen XH, Liu CL, Wang SB (2012) pattern of urban thermal environment in Shanghai, China. Stoch Space-time changes in hydrological processes in response to Environ Res Risk Assess 26(7):899–911 human activities and climatic change in the south China. Stoch Zang S, Wu C, Liu H, Na X (2011) Impact of urbanization on natural Environ Res Risk Assess 26(6):823–834 ecosystem service values: a comparative study. Environ Monit Turner II BL, David S, and Liu Y (1995) Land use and land cover Assess 179:575–588 change science/research plan, IHDP Report No. 07 Zhang Q, Ban Y, Liu J, Hu Y (2011) Simulation and analysis of urban Van Metre PC, Mahler BJ (2005) Trends in hydrophobic organic growth scenarios for the Greater Shanghai Area China. Comput contaminants in urban and Reference Lake sediments across the Environ Urban Syst 35:126–139 United States, 1970–2001. Environ Sci Technol 39:5567–5574 Zhou L, Dickinson R, Tian Y, Fang J, Li Q, Kaufman RK, Tucker CJ, Vellinge P (1998) IHDP Industrial transformation. IHDP-IT Publica- Myneni RB (2004) Evidence for a significant urbanization effect tion No. 12.5 on climate in China. Proc Natl Acad Sci U S A 101:9540–9544

123