The effect of disaggregating land use categories in cellular automata during model calibration and forecasting

Spatial models of urban growth have the ability to play an important role in the planning process; if not in aiding in policy decisions, then in processes such as visioning, storytelling, and scenario evaluation. One question that has not adequately been addressed is to what degree does disaggregating land use types from urban/non-urban categories add to these simulations? This paper aims to answer this question by modeling urbanization in San Joaquin County (CA) using the SLEUTH urban growth model with two equal, but different datasets; one with urban/non-urban data, and the other with the same data, but the non-urban data disaggregated in nine land uses. The results show that there is an explicit link between the likelihood of urbanization, and the type of land use that will be converted to urban, and suggest that future exercises using spatial models should not ignore the impact of aggregating individual land use categories into urban–non-urban classes. 2005 Elsevier Ltd. All rights reserved. 0198-9715/$ see front matter 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.compenvurbsys.2005.04.001 * Corresponding author. E-mail addresses: [email protected] (C. Dietzel), [email protected] (K. Clarke). C. Dietzel, K. Clarke / Comput., Environ. and Urban Systems 30 (2006) 78–101 79 1. Urban simulation and land use dynamics Urban growth and land use change are dynamic spatio-temporal processes of great interest to planners, conservationists, ecologists, economists, and resource managers. Over the past decades, research in these disciplines has sought to develop models of these processes for forecasting future development, evaluating future plans, and identifying endangered natural areas. Despite past failures in urban modeling (Lee, 1973, 1994), there has been a renaissance of spatial modeling in the last two decades due to increased computing power, improved availability of spatial data, and the need for innovative planning tools to aid in decision support (Brail & Klosterman, 2001; Geertman & Stillwell, 2002). These models include the development of new computational methods, including micro-simulation, agent-based and cellular automata (CA), which show potential in representing and simulating the complexity of the dynamic processes involved in urban growth and land use change. Complexity and scaling approaches have provided an additional level of knowledge and understanding of the spatial and temporal patterns of land use change (Batty & Longley, 1994). Furthermore, the models have been used to anticipate and forecast future changes or trends of development, to describe and assess impacts of future development, and to explore the potential impacts of different policies (Jantz, Goetz, & Shelley, 2004; Landis & Zhang, 1998). As previously implied, there are several different types of urban and land use models (EPA, 2000). Since planning is to some degree a management of an economic market, models have been developed to incorporate the economics of land use change (Alberti & Waddell, 2000; Irwin & Geoghegan, 2001). Others have suggested the uses of agents (Parker, Manson, Janssen, Hoffmann, & Deadman, 2003), simulating the decisions of individuals within a system. Cellular automata (CA) are yet another method for simulating urban-land use dynamics, where a set of rules and spatial constraints govern interaction among land uses and their transitions (Batty, Xie, & Sun, 1999; de Almedia et al., 2003; White, Engelen, & Uljee, 1997). We contend that there are currently two schools in CA modeling; both use the same basic foundation (CA), but have different approaches when it comes to incorporating increased details on the dynamics within a system by disaggregating data into multiple land use classes. The first approach is that of Ward, Murray, and Phinn (2000, 2003), Wu (2002), Yeh and Li (2001, 2002) and Li and Yeh (2001). The models that these researchers have developed, while elegant, treat the urban system as a basic entity, comprised of urban and non-urban components. These non-urban components may be referred to as rural or agriculture, but no matter what the nomenclature, the system is decomposed into two classes (we disregard the inclusion of a water category, as this is largely stationary, and lacks the dynamic characteristics that other land uses possess)—urban and non-urban. The second school of CA modeling takes the approach that the landscape is comprised of multiple land uses, be they at the broad landscape level or within the city itself, and that the feedback and dynamics among these classes is essential in modeling. This school includes the works of de Almedia et al. (2003), Xie and Batty (2005) and White and Engelen (2000). Even more recent efforts in CA modeling have begun to integrate classical economic 80 C. Dietzel, K. Clarke / Comput., Environ. and Urban Systems 30 (2006) 78–101 theory into the rules that govern land use transition. Caruso, Rounsevell, and Cojocaru (in press) and Caruso (2003), are the basis for land use allocation in some planning support systems (RIKS, 2004). Skeptics may dismiss the use of modeling in planning due to the uncertainty of the future and black-box nature of many spatial models, but models, no matter what type, or which school they are from, undoubtedly have a role in the planning practice, providing insight into possible outcomes of policy implementation and providing a link with other techniques for evaluating future plans and decisions, namely visioning (Helling, 1998), storytelling (Guhathakurta, 2002), and scenarios (Xiang & Clarke, 2003). While both schools of modeling focus on the dynamics of urban growth, there is a disparity between how land use is treated; is it just urban/non-urban, or is it disaggregated into detailed categories? More recent research has begun to focus on the essential issues regarding the role of land use categorization in spatial data and models. The work of Pontius, Agrawal, and Huffaker (2003) in developing a technique to estimate the uncertainty in extrapolations of spatially-explicit land-change simulation models has been key in linking spatially explicit models of land use change with non-spatial models, while the work of Brown and Duh (2004) has demonstrated the ability to translate between land use and land cover, a critical advance. Most recently the role of aggregation has come to the forefront as both Pontius and Malizia (2004) and Dietzel and Clarke (2004) both examined the role of categorical aggregation in land use/land cover change and modeling. Despite these advances, it is still necessary to research the role of land use in spatial models. Land use of various types, can, in some cases, be quite robust to change, and this needs to be recognized and accounted for in models, and exercises using them. But to what degree does the interaction between various land uses differ from the simple urban/non-urban interaction? This has not yet been explored and may have a drastic impact on model projections, hence the evaluation of any sort of policy or scenario within a spatial model. This paper explores the impact of disaggregating simple urban/non-urban datasets into detailed multiple land use categories during both the calibration and forecasting phase of modeling exercises, and is a litmus test between which school of thought in CA modeling is more appropriate, and why. Using a cellular automata model, the SLEUTH urban growth model (Clarke, Hoppen, & Gaydos, 1997), urban growth for a rapidly growing county in California was modeled using historical data from 1988 to 1996 to calibrate, and then forecast until 2030. Results from calibrating and forecasting the model using the urban/non-urban data were compared within a GIS and through tabular data comparison to the results disaggregating the data into multiple categories. While the SLEUTH model is used, it (the model) is not the focus of this research, but a vehicle for examining the impact that disaggregation has during model calibration and forecasting, and to determine which of the two approaches to CA modeling is the more appropriate. Of all the models available, SLEUTH may be the most appropriate because it is a hybrid of the two schools in CA modeling—it has the ability to model only urban growth as Ward et al. (2000, 2003), Wu (2002), Yeh and Li (2001, 2002) and Li and Yeh (2001) have done, and incorporate detailed land use data as de Almedia et al. (2003), Xie and Batty (2005) and White and EngeC. Dietzel, K. Clarke / Comput., Environ. and Urban Systems 30 (2006) 78–101 81 len (2000) have. Still a detailed introduction to SLEUTH is necessary to provide the reader with an understanding of the methodology, and how demonstrate how the model is a hybrid between the two schools of thought in CA modeling. The paper concludes with recommendations to future modelers and planners who plan to integrate land use change data into future modeling efforts. While the results of the calibration and forecasting are from one cellular automaton model, with a specific rule set, the findings are presented in a general manner, addressing the question of what the impact of disaggregating land use classes from non-urban data has on calibration and forecasting in CA modeling. 2. Modeling using SLEUTH Modeling geographic systems using cellular automata models is a recent advance relative to the history of the geographic sciences (Silva & Clarke, 2002). Applying this method of modeling to urban systems for planning applications was recognized early (Couclelis, 1985), and application of these models has proliferated in the last decade, including in the development of SLEUTH. The SLEUTH urban growth model is a cellular automaton model that has been widely applied for simulation and forecasting of landscape changes (Esnard & Yang, 2002; Jantz et al., 2004; Leão, Bishop, & Evans, 2004; Yang & Lo, 2003). While use of the model has mainly been limited to academic exercises, more recent work has used the model to explore possible location of facilities (Leão et al., 2004), examine the implications of land use policy decision at a regional level (Jantz et al., 2004), and in community planning at a local level (Santa Barbara Region Economic Community Project, http:// www.sbecp.org/documents.htm). A complex calibration process trains the model with historical spatial and temporal urban growth (Silva & Clarke, 2002). Written in C code, the model is freely downloadable at http://www.ncgia.ucsb.edu/project/gig, and can be applied to any geographic area with the proper data. Five reasons attributed to choosing this model to answer the main research question: (1) the shareware availability meant that any researcher could perform a similar application or experiment at no cost given they had the data; (2) the model is portable so that it can be applied to any geographic system at any extent or spatial resolution; (3) the presence of a well-established Internet discussion board and list-serve ([email protected]) to support any problems and provide insight into the model s application; (4) a well documented history in geographic modeling literature that documented both theory and application of the model; and (5) the ability of the model to project urban growth based on historical trends with urban/non-urban data or with detailed categorized land use data, which was essential to answering the main research question. SLEUTH is a moniker for the input data required to use the model: Slope, Land Use, Exclusion, Urban, Transportation, and Hillshade. The slope layer helps to implement topographic constraints on the model, focusing growth on flatter, more suitable areas that are less costly to develop. Inclusion of land use in modeling with SLEUTH is optional; the model does not require land use data, but does have a 82 C. Dietzel, K. Clarke / Comput., Environ. and Urban Systems 30 (2006) 78–101 separate model, termed the deltatron model, to model land use change. Two layers of land use, with any number of classes but the same classification, are required if the modeling of land use is desired. Implementation of scenarios within the framework of this model is largely done through the manipulation of the exclusion layer. This layer allows the user to implement constraints on the model, prohibiting growth in some areas, providing resistances or attractions in others. The simulation of urban growth is the main focus of the model. Recognizing that urban growth is not a linear process, four input data layers of urban extent are required so that a more dynamic model of growth can be presented. If both urban growth and land use change are simulated, then it is necessary for the urban land use class in both of these data to be redundant of one another. The influence of transportation on urbanization is a well-known relationship, so a transportation layer is included in the model. Roads are classified into three classes based on their accessibility; so major interstates and highways would be one class, state routes and major arterial routes would be another, and local collector streets would be a third class. But the user has complete control over the classification of these routes. The final layer is a hillshade, or topographic relief layer. The only purpose of this layer is to add some positional reference to the output maps so that users have a geographic sense of where urbanization is forecast to take place. Upon assembly of these data, a user can calibrate, and then forecast urbanization and land use change. Calibration of SLEUTH produces a set of five parameters (coefficients) that describe the historical growth patterns of the system over time based on a fixed set of transition rules. Five coefficients (with values 0 to 100) control the behavior of the system, and are predetermined by the user at the onset of every model run (Clarke & Gaydos, 1998). These parameters are: 1. Diffusion. Determines the overall dispersiveness nature of the outward distribution. 2. Breed coefficient. The likelihood that a newly generated detached settlement will start on its own growth cycle. 3. Spread coefficient. Controls how much contagion diffusion radiates from existing settlements. 4. Slope resistance factor. Influences the likelihood of development on steep slopes. 5. Road gravity factor. An attraction factor that draws new settlements towards and along roads. These parameters drive the transition rules that simulate four types of urban growth. Spontaneous growth is the urbanization of land that is of suitable slope, yet not adjacent to preexisting urban areas. Diffusive growth occurs when newly established urban areas begin to transform the land around them from other uses into urban land cover. Organic growth takes place at the urban fringe and as infill within areas that may not have fully made the transition from another land use to urban. And road influenced growth takes into account the influence that roads have over urbanization and land use change, as they attract new growth. C. Dietzel, K. Clarke / Comput., Environ. and Urban Systems 30 (2006) 78–101 83 In the calibration of this model, the transition rules are applied to the input data to find the set of parameters that best replicates the spatial patterns of urbanization and land use change. For this model the transition rules are implemented uniformly across space and in two nested temporal loops. The first loop is an outermost loop that attempts to use the parameters to best replicate the transitions between the first year of input data, the seed layer, and the last. The second loop attempts to replicate the growth and transitions between the individual input periods. By running the model in calibration mode, a set of control parameters is refined by three sequential brute-force calibration phases: the coarse, fine and final calibrations (Silva & Clarke, 2002). Typically the calibration of SLEUTH is a three-step process: 1. Coarse. The model attempts to simulate the historical growth patterns for a wide range of parameter values that cover the entire parameter space, and generally spaced in increments of 25. Using the first layer of urban extent data as a seedlayer for simulating growth, and the other three urban extent layers as control points, each parameter set attempts to replicate the historical growth of the system over time. The ability of the parameter to recreate the time series of input data is evaluated using a variety of spatial metrics, but most commonly the Lee–Sallee metric, although others have been used (Jantz et al., 2004; Yang & Lo, 2003), including in this exercise. The Lee–Sallee metric describes the degree of spatial matching between the simulated data and the input historical data, and is a rigorous measure of the ability of a parameter set to replicate historical urban growth patterns. The tested parameter sets are sorted based on their goodness of fit, and the parameter values are narrowed to values around the parameter set that produced the best fit between the historical and simulated data. 2. Fine. The narrowed range of parameters from the previous step are used to simulate the historical growth patterns. Results of these simulations are evaluated using spatial metrics of fit, and the range of parameters is narrowed one last time. 3. Final. The historical data is simulated one last time using the renarrowed set of parameters, and the one that best recreates the urban growth is then used in model forecasting. For this research, a composite metric was used to evaluate the performance of the model, and was the product of the compare statistic (Com), population statistic (Pop), Lee and Sallee statistic, and the F-match statistic. The compare statistic is a ratio of the modeled population of urban pixels in the final year to the actual number of urban pixels for the final year. The population statistic (Pop), a least squares regression score (r) for modeled urbanization compared to actual urbanization for the time series. The Lee and Sallee statistic (Lee & Sallee, 1970) measures the spatial urban fit, and F-match measures the proportion of goodness of fit across land use classes. Use of these metrics allowed for narrowing the parameter range based on spatial fit for each year, over the entire time series, and with the replication of land use change patterns. The set of parameter arrived at in the end of the final calibration were then used in the forecasting of land use change. 84 C. Dietzel, K. Clarke / Comput., Environ. and Urban Systems 30 (2006) 78–101 3. Data and study area San Joaquin County (CA) was chosen as a study area (Fig. 1) for this research for three very important reasons. First there is an abundance of high-resolution land use data (derived from aerial photography and satellite imagery) from multiple time periods with the same land use classification scheme. Finding a study area that was regularly monitored and had consistently categorized data with details below the parcel level made this a prime study area for answering the research question. Second, all the data for this research, including the land use data, were available for free to the public, so in the philosophy of science, the results of this research can be replicated because anyone can duplicate this research; using either the same model, data, or both. Finally, San Joaquin County is a county that is positioned between two major metropolitan areas (the San Francisco Bay and Sacramento Metro), and is undergoing growth pressures in multiple directions. These growth pressures were important because they almost assured that there would be significant land use change and urbanization taking place before the data were even examined. In compiling the data on historical urban extent, urban was defined based on the California Farmland Mapping and Monitoring Program s (CA-FMMP) definition, and used as the minimum mapping unit: Fig. 1. The location of San Joaquin County in relation to places within California having a population greater than 50,000 (lower left), and the urban extent for San Joaquin County in 1988, 1992, and 1996.