a neural network for modeling multicategorical parcel use change

20 International Journal of Applied Geospatial Research, 2(3), 20-31, July-September 2011

Copyright © 2011, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.Copyright © 2011, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.

Keywords: GIS,LandUse,NeuralNetworks,ParcelUse,SpatialModeling

INTRODUCTION

Land use change has been a major cause of many environmental problems. Changes are often necessary to accommodate population growth and economic development. Degraded environ-ments, however, threaten sustainable develop-ment. As more states and local governments are willing to create and implement smart growth strategies, planners and policy makers need to

know the impacts of future land use change and related regulatory decisions. Predictive model-ing is not only a learning tool to understand causal factors and dynamic changes of a land use system, but also an important mechanism to predict future changes and simulate the po-tential effects under different growth scenarios (Pijanowski et al., 2002).

Despite the emergence of numerous pre-dictive models over the last half-century, our ability to accurately predict land use change remains limited. Recently, 18 scientists jointly conducted an evaluation of the performance of

A Neural Network for Modeling Multicategorical

Parcel Use ChangeKangShouLu,TowsonUniversity,USA

JohnMorganIII,TowsonUniversity,USA

JefferyAllen,ClemsonUniversity,USA

ABSTRACTThispaperpresentsanartificialneuralnetwork(ANN)formodelingmulticategorical landusechanges.Comparedtoconventionalstatisticalmodelsandcellularautomatamodels,ANNshaveboththearchitectureappropriateforaddressingcomplexproblemsandthepowerforspatio-temporalprediction.Themodelconsistsoftwolayerswithmultipleinputandoutputunits.Bayesianregularizationwasusedfornetworktraininginordertoselectanoptimalmodelthatavoidsover-fittingproblem.Whentrainedandappliedtopredictchangesinparceluseinacoastalcountyfrom1990to2008,theANNmodelperformedwellasmeasuredbyhighpredictionaccuracy(82.0-98.5%)andhighKappacoefficient(81.4-97.5%)withonlyslightvariationacrossfivedifferentlandusecategories.ANNalsooutperformedthebenchmarkmultinomiallogisticregressionbyaverage17.5percentagepointsincategoricalaccuracyandby9.2percentagepointsinoverallaccuracy.TheauthorsusedtheANNmodeltopredictfutureparcelusechangefrom2007to2030.

DOI: 10.4018/jagr.2011070102

International Journal of Applied Geospatial Research, 2(3), 20-31, July-September 2011 21


nine land cover/land use models applied in 13 regions (Pontius et al., 2008). These models included conventional logistic regression (Mc-Connell, Sweeney, & Mulley, 2004), cellular automata based SLEUTH (Dietzel & Clarke, 2004; Silva & Clarke, 2002), multiple agent based SAMMBA (Boissau & Castella, 2003; Castella, Trung & Boissau, 2005), multistage model CLUE and its variant CLUE-S (Duan et al., 2004; Verburg et al., 2002; Veldkamp & Fresco, 1996; Verburg & Veldkamp, 2004), and multifunctional LTM (Pijanowski et al., 2005), multiple objective GEOMOD (Pontius, Cornell & Hall, 2001; Pontius & Malanson, 2005; Pontius & Spencer, 2005). It was found that prediction accuracy, as measured in percent of the correctly predicted against the observed, falls within a range of 1-73 percent, falling below 30 percent in seven cases and above 50 percent only in three cases. Some of the assessments were against the sample datasets used for model calibration or training. But judging on these results, how much can we trust the predictions of the models?

Pontius et al. (2008) point out that some-thing must be wrong with the mechanics of these models. We attribute the poor performances more to the oversimplification of complex land use systems in both semantics and syntax. Semantically, most models, particularly cellular automata models, tend to overemphasize “par-simony” and use too few predictive variables to truly represent complex reality. It may not be a coincident that SLEUTH, constructed with only three effective predictive variables, is the worst performer among all the models examined by Pontius et al. (2008). Although there is no suggested threshold for how many predictive variables are deemed appropriate, a good model should include variables that measure key driv-ers of changes in land use such as population, economy, and technology (Turner & Meyer, 1994). More importantly, the model should be able to generate meaningful predictions, and pass the reality test.

From the syntax perspective, most existing models, particularly conventional statistical models, lack the multilayered, hierarchical,

interconnected structure needed for modeling complex land use systems as Batty and Torrens (2005) have suggested. Multiple layers mimic the processes of land use changes that involve land transaction, speculation, and development; whereas interconnections through these layers capture the interrelationships and interactions between dependent and independent variables, whether they are observable, quantifiable or comprehensible. The CLUE (-S) model is to some degree structured in this way and thus performed relatively well (Pontius et al., 2008). Batty and Torrens (2005) did not provide spe-cific suggestions about what algorithms to use and how to derive them in order to describe such interrelationships and implement the framework. The structure of the conceptual model, however, implicitly points to an artificial neural network (ANN) approach which does offer such capabilities.

This paper presents an artificial neural network (ANN) for modeling multi-categorical parcel use changes. We intended to improve the model by: 1) incorporating more effective predictor variables for a better semantic repre-sentation; 2) expanding the model structure for a multiclass land use system; 3) using a multilayer neural network to enhance predictive power; and 4) applying a Bayesian regularization (BR) training algorithm to avoid the over-fit problem. We applied the model in Beaufort County, South Carolina, to predict and simulate future land use changes from 2000 to 2030 under different growth scenarios.

BACKGROUND

ANNs have been around since the early 1940s, but it was not until the mid 1990s that they were introduced into the geosciences for spatial in-teraction, interregional telecommunication, re-source management, suitability assessment, and image classification (Fischer & Gopal, 1994; Openshaw, 1993). Openshaw and Openshaw (1997) identified four major benefits of neural networks for modeling: better performance, greater representational flexibility and freedom



from current model design constrains, the op-portunity to handle explicitly noisy data, and the incorporation of spatial dependency in the net representation (which is currently ignored).

Researchers in land use modeling have been a half-decade slower in realizing such benefits. Pioneering studies mainly focus the use of ANNs for simulating the change in urbanization patterns (Li & Yeh, 2001; Pijanowski, Gage, & Long, 2000; Pijanowski et al., 2002; Yeh & Li, 2002). Pijanowski et al. (2005) later incorpo-rated an ANN as one of the three components of the Land Transformation Model (LTM), and applied it in several other areas. Allen and Lu (2004, 2006) found that ANN models not only attained fair to high prediction accuracy but also outperformed the logistic regression in several applications in coastal regions. They also found that ANN models were capable of discerning isolated, sparsely distributed features or events that other models often fail to predict. Land use systems modeled in these studies are all binary. It remains to be tested whether ANNs can achieve similar success in predicting multi-category land use change.

Figure 1 provides both a graphic illustration and mathematic notations (MatLab) of a two-layer feedforward neural network appropriate for land use prediction. The network has R input units, S1 hidden units and S2 output units. The connectors between the input units and the hidden units are input weights, collectively denoted by IW1,1, which is a Rx S1 matrix. The superscripts represent the destination and source of a connector. The connectors between hidden units and output units (second layer) are layer weights, denoted by LW2,1, which is a S1x S2 matrix. Each unit in the hidden layer or output layer has a bias term, b1 or b2. They together form a bias vector, b1 or b2. The input vector is p, whereas a1 and a2 are the output vectors for the hidden layer and output layer, respectively. The vector a2 is also the output of the whole network; f1and f2 are transfer functions for the hidden layer and output layer, respectively. A logistic function is often used as transfer func-tion for classification problems.

The prediction starts from left to right. First, each hidden unit sums with its own bias b1 the product of the input value and correspond-

Figure1.Atwo-layerfeedforwardneuralnetworkformodelingmulti-categoricallanduses



ing weight lw1,1 from every input unit p. Then, a transfer function, f1, is applied to the sum n1 before its output a1 is sent to every unit of the output layer. Each unit of the output layer repeats the process, but uses the output of every hidden unit as the input. The result is the output, a2, or prediction of the network. This process is also called feedforward prediction. For this ap-proach, the determination of the weights and biases for multilayer networks used to be a challenge until recently.

A backpropagation training algorithm (Rumelhart et al., 1986) makes it possible to automatically derive the weights and bias for a neural network by learning from a sample. It starts with random weights and bias and proceeds with the feedforward prediction. The prediction error, the difference between the output value and the observed value (often called the target value), is then calculated. The error is redistributed backward to every hidden unit based on a gradient descent rule. The information is then used for updating the corresponding weight and bias for every unit at a certain correction rate before starting another cycle for the next input-target pair. The entire process that goes through every pair in a sample dataset is an epoch. It usually takes many epochs before a network achieves a desirable goal and stops training. The backpropagation method has revitalized neural network sciences and marks the second generation of neural computation (Penny et al., 1999).

Many new algorithms have been developed for increasing the training speed, maximizing approximation capability, or avoiding over-fitting problems. One such innovative algorithm is the Bayesian Regularization (BR) developed by MacKay (1992). Instead of minimizing just sum of squared error (SSE) as the training objective, BR incorporates sum of squared weights (SSW) into the performance function by using two additional parameters (δ and γ) for regularizing SSE and SSW, respectively. Built upon Bayesian probability framework, BR is capable of 1) determining all parameters based on the posterior probabilities conveyed in the sample data, 2) automatically selecting an

optimal model that avoids over-fitting problems for better generalization, and 3) retaining great approximation power once sufficiently trained. It is no longer necessary to split a sample into three subsets for training, testing, and validation as an essential technique to avoid over-fitting problems.

METHODS

Study Area

Beaufort County, South Carolina, was selected as the study area. It is a well-known coastal tourist destination. The county covers 923 mi2 in area, with 63 percent land and 37 percent water. Driven by tourism development, new immigra-tion, service sector expansion, and military demand, Beaufort County has seen tremendous growth in population, urban area, and impervi-ous surface over the past two decades. From 1990 to 2000, its population increased by 39.93 percent, from 86,425 to 120,935. This growth rate triples the national average (13 percent) and leads all counties in South Carolina. From 1990 to 2008, the total number of built parcels increased by 136.4 percent from 23,446 to 56,929. Impervious surface almost quadrupled (380 percent) in the fast growing Bluffton area. The land conversion rate has been high, with an average of 2,575 parcels (1,500 acres) per year over the last decade. As one of the top seven retiree communities in the US, the county is expected to continue growing at a rapid pace. These changes impose tremendous pressure on natural environments, local infrastructure, and public services. There is a tremendous demand for predictions and impact assessments needed for smart growth decisions.

Neural Network Specification

We used a two-layer, feedforward neural net-work with 18 x 18 x 5 architecture to model a five-class parcel use system. The five output units included commercial, urban residential, rural residential, recreation (golf course), and undeveloped, all of which are binary (1 or 0).



Each parcel is assigned to one type of use, and each parcel is encoded by a vector of five ele-ments. For example, if a parcel was developed for commercial use, it was coded as (1, 0, 0, 0, 0). Input units represent 18 independent vari-ables listed in Table 1. The number of hidden units was initially set to 18, but may change, depending on the BR training result.

We used a logistic function, f = 1 / (1+ exp (- x)), as the transfer function for all layers, and a softmax function, yj = exp (- xj) / ∑ exp (- xk), for land transition probability and output clas-sification. The former allows us to simulate parcel changes under different scenarios. The BR algorithm is available in the Neural Network Toolbox of Simulink 8 (MatLabtm). We wrote scripts to enable data input, classification, and error assessment. Several ArcGIS-based sub-modules were developed for sample extraction, data processing, and spatial prediction.

Performance of the ANN was assessed in measures of: 1) mean squared error (MSE), sum of squared error (SSE), and sum of squared weights (SSW); 2) prediction accuracy; and 3) the Kappa coefficient. Multinomial logistic regression was used as a benchmark model for

comparisons. We have also assessed the effects of predictor variables.

Data Preparation

Parcel use in 2008 was derived from the parcel shape file (ESRI, Inc., Redlands, California) and Tax Assessor’s Table provided by Beaufort County. We grouped parcel uses into five general categories: 1) commercial, including all com-mercial, industrial, institutional, and service/utility uses; 2) urban residential, including all developed accommodation use < 5 ac; 3) rural residential, with parcel size falling between 5-10 ac; 4) recreational (only golf courses); and 5) undeveloped, including wetlands, open space, forestlands, croplands, parks and reserves, and other preserved and protected lands. Of a total 99,781 parcels, 33,250 (33.32 percent) were developed or built between 1990 and 2008. We call this ‘new development’ or ‘net change’ for model validation. Because of rich temporal information, we were able to derive changes on an annual or decadal basis for dif-ferent land uses, periods, and regions. We used the growth rates for three different lengths of

Table1.Predictorvariables

Group Independent Variables

Physical properties • land area • wetland ratio • slope • distance to water front • forest wetland

Road accessibility • cost distance to major road • distance to local street

Proximity to utilities • distance to waterline • and distance to sewer line

Neighborhood effects • cost distance to central business district • developed parcels • distance to built parcels • agricultural use

Demographic features • block population • block neighborhood population

Land value • parcel land value • total property values

Land ownership • private ownership



periods to simulate future changes: 1980-2008, 1990-2008, and 2000-2008.

We prepared two sets of data for all predic-tive variables, 1990 for model training, and 2008 for future prediction. National Wetland Inven-tory data from 1989 were used to extract forested wetland, land area, and wetland ratio. Slope was derived from the 30 m DEM downloaded from USGS’s National Map Seamless Server. Block population data were from Census Data 1990 and 2000. Beaufort County provided the updated data for roads and streets from which proximity and coast distance variables were derived. Waterline and sewerline layers were from South Carolina Department of Commerce. Protected lands were downloaded from the Natural Resources GIS Data Clearinghouse of the South Carolina Department of and updated by using the 2008 parcel data. The tax assessor file was the source for building counts, built year, parcel size, land price, land value, and land ownership at the finest spatial units. Since the coastal area is highly fragmented by water bodies and wetlands, we constructed two cost distance variables to measure the influences of major roads and central business district over a cost surface created by weighting these natural barriers more as frictional factors.

All data were aggregated in parcel units and the parcel shape file was joined with the tax assessor’s table. Most distances were measured from the centroid of a parcel, but distance to waterfront is the zonal average of the grid’s Euclidian distances to a parcel. Data for each independent variable were scaled to a range between 0 and 1 by using the same set of min-max values for both training and prediction sets.

The training dataset was extracted by using a pseudo-stratified systematic random sampling method. The strata are five land use classes, but subsample size for each stratum is not purely proportional to the total number of parcels, but proximately proportional to the total area of the stratum in concern. The sample size (S) for the entire training set was estimated based on the formula S=Nw / Et where Nw is the total number of weights including biases, and Et is the targeted error threshold. The final training

sample includes 7,812 parcels, accounting for nearly 8 percent in number and 17 percent in area of all parcels in the county. We also pre-pared four more datasets for model validation and testing. They included two sample sets, a net change set, and the full dataset. The first three sets accounted for 20 percent, 10 percent, and 33 percent of the total number of parcels in the county respectively.

RESULTS

Model Performance

We performed multinomial regression analysis in a preliminary study to screen predictive variables. Eighteen of the variables included in the preliminary study were found statistically significant and thus used to construct the ANN model. Based on the 4:2:2 ratios, the training sample was split into three subsets for training, validation, and testing, respectively, during the training process. The training process stopped within 100 epochs in several runs. MSE is 0.018 for the training subset and 0.038 for both testing and validation subsets. Results of BR suggest that an optimal model with 415 effective parameters (18 x 17 x 5) would be adequate.

The ANN model performed well against the training dataset (Table 2). The prediction accuracy, overall or categorical, is quite impres-sive as measured in terms of correct percent (82.0-98.5 percent) and Kappa coefficients (0.81-0.98), particularly for a change of 33.3 percent over newly two decades. Kappa coef-ficient is considered a conservative measure of agreement with chance corrected, and a value above 0.80 suggests an excellent prediction. Although both accuracy and Kappa tend to increase with the total number of parcels within the category, their variations across different land uses are relatively small.

The benchmark MLR was also statisti-cally significant based on the likelihood ratio tests (Chi-square = 11,100, df = 76, p < 0.001). It was, however, outperformed by the ANN model in all measurements (Tables 2 and 3). The ANN model improved prediction accuracy



by an average of 17.5 percentage points across the five uses and by 9.2 percentage points in overall. The two numbers are even more impres-sive for Kappa coefficients (0.19 and 0.14). The greatest improvement occurred where MLR failed to generate satisfactory results, by 33.6 and 32.2 percentage points for commercial and rural residential respectively.

Results of model validation against four different datasets are summarized in Table 4. Overall accuracy is high in all cases (> 94.9 percent) with only little variation (< 1.3 percent-age points) across different cases. For all but net change validation, prediction accuracy dropped only by average 2.4 percentage points in all 19 measurements with a range of 0.8-2.6 percentage points. The net change case is an exception because we observed three extreme

discrepancies: the largest jump (+12.2), the smallest drop (-1.2), and the largest drop (-11.1) in prediction accuracy from those for the train-ing set. ANN’s predictive ability did not show a consistent decay pattern which was not ex-pected.

Predicted Changes

The predicted long-term trajectories for parcel development are shown in Figure 2. They reflect quantitative demands at three different growth rates. We used 2008 data as the new input of the trained ANN model to predict land transi-tion probabilities for five parcel uses. These two types of information were used to map future change through 2030. Figure 3 shows the distribution of predicted parcel use changes

Table2.PredictionaccuracyoftheANNmodelagainstthetrainingsample

ObservedPredicted

Row Total Accuracy (%) Kappa

I II III IV V

I 511 24 24 9 6 574 89.0 0.88

II 12 2815 30 6 14 2877 97.8 0.97

III 10 22 866 7 38 943 91.8 0.91

IV 4 5 5 218 34 266 82.0 0.81

V 16 11 18 2 3105 3152 98.5 0.98

Column total 553 2877 943 242 3197 7812 96.2a 0.94b

Notes: I=Commercial, II=Urban residential, III=Rural residential, IV=Golf course, and V=Other use.a. Overall prediction accuracy, b. Overall kappa coefficient.

Table3.PredictionaccuracyoftheMLRmodelagainstthetrainingsample

ObservedPredicted

Row Total Accuracy (%) Kappa

I II III IV V

I 318 197 12 12 35 574 55.4 0.53

II 19 2659 50 3 146 2877 92.4 0.88

III 74 147 562 4 156 943 59.6 0.56

IV 13 24 5 177 47 266 66.5 0.66

V 5 52 8 4 3083 3152 97.8 0.96

Column total 429 3079 637 200 3467 7812 87.0a 0.80b

Notes: I=Commercial, II=Urban residential, III=Rural residential, IV=Golf course, and V=Other use.a. Overall prediction accuracy, b. Overall kappa coefficient.



for selected classes and areas under selected growth rates and constraints. Table 5 provides a quantitative summary of simulated regional development with and without a growth bound-ary constraint.

DISCUSSION AND CONCLUSION

A neural network was applied to predict parcel use change in Beaufort County, South Carolina. The model is not only theoretically appropriate

for addressing complex problems, but also has demonstrated improved predictive capability. The accuracy ranges from fairly good to excel-lent in all 29 measurements, regardless of varia-tions in land use category, sample set, parcel size, and time span. The ANN outperformed both benchmark MLR and models evaluated by Pontius et al. (2008). A critique of ANNs is the over-fitting problem. But this problem, if pres-ent, might have been kept to minimal in our case, due to the use of BR and large training sample. Cross-validations do not show any substantial

Table4.Predictionaccuracy(%)oftheANNmodelinfourvalidations

Parcel UseCategory Training Set

Validation Sets

Use-based Area-based Net Change a Full Set

I 89.0 86.6 83.8 79.2 87.1

II 97.8 96.5 96.5 96.6 96.2

III 91.8 88.5 84.9 80.7 87.4

IV 82.0 79.8 77.6 94.2 80.6

V 98.5 97.7 97.6 97.7

Overall 96.2 96.4 96.1 94.9 96.2

Notes: I=Commercial, II=Urban residential, III=Rural residential, IV=Golf course, and V=Other use.a Category V or Other use is treated as the background category and is thus not applicable in this case.

Figure2.Predicted trajectoriesofparceldevelopment inBeaufortCountyat threedifferentgrowthrates



drops in prediction accuracy nor suggest any pattern of variations across different sample sets and land uses. It was found again that ANNs enable us to predict rare, sparsely distributed events with the level of accuracy other models often fail to reach (Allen & Lu, 2006).

The use of parcels as objects to be modeled and as units of analysis has brought both benefits and issues. This approach adds more realism to our modeling practice because of the use of real

world units for land transaction and develop-ment (Landis, 1994). Many parcel attributes are predictive variables with measurements that can be readily incorporated into a model to improve model semantics and prediction accuracy. Also, many property variables are spatially well-defined by parcel boundaries and no additional error is induced by aggregation or disaggregation. For the same reason, it is less difficult to interpret the modeling results.

Figure3.PredictedparcelusechangesinBeaufortCountyfrom2000to2030underdifferentgrowthscenarios



Four issues we have encountered are related to large parcel partitioning, vector data processing, parcel use classification, and data availability. How to model the large parcel partitioning process remains to be a technical challenge.

The predicted trajectories and growth pat-terns have policy implications. The anticipated urban growth ratio is relatively moderate (2:1-3:1) compared to that (6:1) for the Charleston area (Allen & Lu, 2003). Potential impacts of land development can be relatively severe because Beaufort’s physical environment is very fragmented and ecological systems are quite vulnerable. While new development is continuously spreading in the urban fringe areas, infill development is expected to intensify in previously sprawled regions. Over time, devel-opment will change the image of the county as a well vegetated, properly landscaped tourist destination. The simulation results suggest caution in implementation of a growth bound-ary. The policy may be able to preserve natural areas and farm lands in North Beaufort, but it has a potential risk of driving land conversion and environmental change in many peripheral islands in South Beaufort too fast for coastal terrestrial and aquatic life to adapt. Incentives should be in place in order to regulate such developments.

ACKNOWLEDGMENT

We would like to thank South Carolina Sea Grant and NOAA’s National Centers for Coastal Ocean Science for funding this project.

REFERENCES

Allen, J. S., & Lu, K. S. (2006). Predicting trajectories of urban growth in the coastal Southeast. In G. S. Kleppel, M. R. Devoe & M. V. Rawson (Eds.), Chang-inglandusepatternsinthecoastalzone:Managingenvironmentalqualityinrapidlydevelopingregions (pp. 47-67). New York, NY: Springer.

Allen, J. S., & Lu, K. S. (2003). Modeling and predic-tion of future urban growth in the Charleston region of South Carolina: A GIS-based integrated approach. ConservationEcology, 8(2), 1–20.

Allen, J. S., & Lu, K. S. (2004). Artificialneuralnetworkvs.binarylogisticregression:Twoalterna-tivemodelsforpredictingurbangrowthintheMyrtleBeachregion. Paper presented at the NOAA’s Land Use-Coastal Ecosystem Study (LU-CES) Program.

Batty, M., & Torrens, P. M. (2005). Modeling and prediction in a complex world. Futures, 37(7), 745–766. doi:10.1016/j.futures.2004.11.003

Table5.PotentialimpactofthegrowthboundaryonregionaldevelopmentinBeaufortCounty

Region

Growth Boundary Gain (or Loss)

Yes No Yes-No

N Area N Area N% Area%

St Henna Island 0 0 5132 4592 -100 -100

S Beaufort Islands 2726 2038 1664 875 64 133

Hilton Head 5239 3968 4884 3035 7 31

NE Corner 0 0 958 1341 -100 -100

Port Royal-City of Beaufort 10332 8148 9631 5945 7 37

Bluffton 9404 7318 8687 5677 8 29

Entire County 27701 21472 30956 21466 -11a <1a

Notes: Prediction is based on the 1990-2008 growth rate. Area units are acres.a Discrepancies are caused by variable parcel sizes due to the use of area-based demand.



Boissau, S., & Castella, J. (2003). Constructing a com-mon representation of local institutions and land use systems through simulation-gaming and multi-agent modeling in rural areas of Northern Vietnam: The SAMBA-Week methodology. Simulation&Gaming, 34(3), 342–347. doi:10.1177/1046878103255789

Castella, J., Boissau, S., Trung, T. N., & Quang, D. D. (2005). Agrarian transition and lowland-upland interactions in mountain areas in northern Vietnam: Application of a multi-agent simulation model. Ag-riculturalSystems, 86(3), 312–332. doi:10.1016/j.agsy.2004.11.001

Dietzel, C. K., & Clarke, K. C. (2004). Spatial differ-ences in multi-resolution urban automata modeling. TransactionsinGIS, 8, 479–492. doi:10.1111/j.1467-9671.2004.00197.x

Duan, Z., Verburg, P. H., Zhang, F., & Yu, Z. (2004). Construction of a land-use change simulation model and its application in Haidian District, Beijing. ActaGeographicaSinica, 59(6), 1037–1046.

Fischer, M. M., & Gopal, S. (1994). Artificial neural networks: A new approach to modeling interregional telecommunication flows. JournalofRegionalSci-ence, 34(4), 503–527. doi:10.1111/j.1467-9787.1994.tb00880.x

Landis, J. (1994). The California urban futures model: A new generation of metropolitan simulation models. EnvironmentandPlanning.B,Planning&Design, 21, 399–420. doi:10.1068/b210399

Li, X., & Yeh, A. G. O. (2001). Calibration of cellular automata by using neural networks for the simulation of complex urban systems. Environment&PlanningA, 33(8), 1445–1462. doi:10.1068/a33210

Lu, K. S., & Allen, J. S. (in press). Neural networks for modeling urban growth patterns and potential ecosystem impact in the coastal zone . In Brouwer, F., & Goetz, S. J. (Eds.), Thedynamicsoflanduseandecosysteminteractions:Atransatlantic,multi-disciplinaryandcomparativeapproach. New York, NY: Springer.

MacKay, D. J. C. (1992). Bayesian interpolation. NeuralComputation, 4(3), 415–447. doi:10.1162/neco.1992.4.3.415

Openshaw, S. (1993). Modeling spatial interaction using a neural net . In Fischer, M. M., & Nijkamp, P. (Eds.), GIS spatial modeling and policy (pp. 147–164). Berlin, Germany: Springer-Verlag.

Openshaw, S., & Openshaw, C. (1997). Artificialintelligence in geography. Chichester, UK: John Wiley & Sons.

Penny, W. D., Husmeier, D., & Roberts, S. J. (1999). Introduction . In Lisboa, P. J. G., Ifeachor, E. C., & Szczepaniak, P. S. (Eds.), Artificialneuralnet-worksinbiomedicine (pp. 1–9). Berlin, Germany: Springer-Verlag.

Pijanowski, B. C., Brown, D. G., Manik, G., & Shellito, B. A. (2002). Using neural nets and GIS to forecast land use changes: A land transforma-tion model. Computers, Environment and UrbanSystems, 26(6), 553–575. doi:10.1016/S0198-9715(01)00015-1

Pijanowski, B. C., Pithadia, S., Shellito, B. A., & Alexandrids, K. (2005). Calibrating a neural network-based urban change model for two metropolitan areas of the Upper Midwest of the United States. Interna-tionalJournalofGeographicalInformationScience, 19, 197–215. doi:10.1080/13658810410001713416

Pontius, R. G. Jr, Boersma, W., Castella, J., Clarke, K., de Nijs, T., & Dietzel, C. (2008). Comparing the input, output, and validation maps for several models of land change. TheAnnalsofRegionalSci-ence, 24(1), 11–37. doi:10.1007/s00168-007-0138-2

Pontius, R. G. Jr, Cornell, J., & Hall, C. (2001). Modeling the spatial pattern of land-use change with GEOMOD2: Application and validation for Costa Rica. AgricultureEcosystems&Environment, 85(1-3), 191–203. doi:10.1016/S0167-8809(01)00183-9

Pontius, R. G. Jr, & Spencer, J. (2005). Uncertainty in extrapolations of predictive land change mod-els. Environment and Planning B, 32, 211–230. doi:10.1068/b31152

Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning internal representations by error propagation. In J. L. McClelland & D. E. Rumelhart (Eds.), Paralleldistributedprocessing:Explorationsinthemicrostructureofcognition:Vol.1.Founda-tions (pp. 318-362). Cambridge, MA: MIT Press.

Silva, E., & Clarke, K. C. (2002). Calibration of the SLEUTH urban growth model for Lisbon and Porto, Portugal. Computers, Environment andUrbanSystems, 26, 525–552. doi:10.1016/S0198-9715(01)00014-X

Turner, B. L., & Meyer, W. B. (1994). Global land-se and land-cover change: An overview . In Meyer, W. B., & Turner, B. L. II, (Eds.), Changeinlanduseand land cover:A global perspective (pp. 1–10). Cambridge, UK: Cambridge University Press.



KangShouLuisanassociateprofessorofgeographyatTowsonUniversityinBaltimore,Maryland,USA.HereceivedhisPhDfromClemsonUniversity.HisacademicexperiencesincludeservingasaprogramdirectoratShaanxiNormalUniversity,Xian,China,editor-in-chiefofGeographyforSecondaryEducation,avisitingscholaratCaliforniaStateUniversityatNorthridge,andaresearchscientistattheStromThurmondInstituteofGovernmentandPublicAffairs.KangShouisoneoftheauthorsfortheaward-winingCharlestonUrbanGrowthstudyfundedbytheNationalOceanicandAtmosphericAdministration(NOAA).Hehaspublishedmorethan30peer-reviewedpapersandbookchaptersindiversefields.Hiscurrentteachingandresearchfociareonlandusemodeling,urbangrowthsimulation,geographicinformationsystems,artificialneuralnetworks,andwatershedimpairmentassessment.

JohnMorganIII(“Jay”)isaprofessorintheDepartmentofGeographyandEnvironmentalPlanningatTowsonUniversityinBaltimore,MD.PriortojoiningthefacultyatTowsonin1984,Jayworkedwithstateandlocalgovernmentagencies inMaryland.HisexperienceincludesservingasthefirstsystemsanalystandprogrammerontheMarylandAutomatedGeographicInformationSystems(1974to1977).JayhassponsoredanannualGISconferencesince1987,foundedandservedasthedirectoroftheCenterforGISfor13years,andhasbeentheprincipalinvestigatoronover$13millionofgrantsandcontracts.Jay’steachingandresearchinterestsincludegeographic information systems, remote sensing, homeland security and emergencymanagement,andoutdoorrecreationplanningandmanagement.

JeffAllenistheDirectoroftheSouthCarolinaWaterResourcesCenterattheStromThurmondInstituteofGovernmentandPublicAffairsatClemsonUniversity.HeservesonthefacultyofthePolicyStudiesProgramandisalsoanAdjunctAssistantProfessorintheDepartmentofForestryandNaturalResourcesatClemsonUniversity.JeffreceivedaPhDinPolicyStudiesfromClemsonUniversitywithanemphasisinnaturalresourcespolicy.Hisworkinvolvescoor-dinatingwaterresearchwithanationalnetworkofwaterinstitutesandidentifyingandpursuingcriticalwaterresearchneedsforSouthCarolina.Jeffhasservedastheprincipleinvestigatoronmultiplegrantsfromgovernmentagenciesandprivatefoundations.Hehasorganizedandsponsoredseveralstate-widemeetingsonGISmappingandwaterresources.HisrecentworkincludesmappingstructureswithinSouthCarolina’sbeachsetbackzonetodetermineimpactsofsealevelriseandbeacherosionandbuildingurbangrowthpredictionmodelstodescribepotentialfuturedevelopmentpatternsinSouthCarolina.

Veldkamp, T. A., & Fresco, L. (1996). CLUE-CR: An integrated multi-scale model to simulate land use change scenarios in Costa Rica. Eco-logicalModelling, 91, 231–248. doi:10.1016/0304-3800(95)00158-1

Verburg, P. H., Soepboer, W., Veldkamp, T. A., Lim-piada, R., Espaldon, V., & Mastura, S. A. S. (2002). Modeling the spatial dynamics of regional land use: The CLUE-S model. EnvironmentalManagement, 30(3), 391–405. doi:10.1007/s00267-002-2630-x

Verburg, P. H., & Veldkamp, T. A. (2004). Projecting land use transitions at forest fringes in the Philip-pines at two spatial scales. LandscapeEcology, 19, 77–98. doi:10.1023/B:LAND.0000018370.57457.58

Yeh, A. G. O., & Li, X. (2002). Urban simulation using neural networks and cellular automata for land use planning . In Richardson, D., & van Osterom, P. (Eds.), Advances in spatial data handling (pp. 451–464). Berlin, Germany: Springer-Verlag.

a neural network for modeling multicategorical parcel use change

Documents

electronic forms

land use system

land coverland use models

future changes

population growth

dynamic changes

sammba boissau castella

smart growth strategies