+ All Categories
Home > Documents > Using the land transformation model to forecast vacant land

Using the land transformation model to forecast vacant land

Date post: 27-Nov-2023
Category:
Upload: tamu
View: 0 times
Download: 0 times
Share this document with a friend
27
Full Terms & Conditions of access and use can be found at http://www.tandfonline.com/action/journalInformation?journalCode=tlus20 Download by: [Texas A&M University Libraries] Date: 28 March 2016, At: 11:20 Journal of Land Use Science ISSN: 1747-423X (Print) 1747-4248 (Online) Journal homepage: http://www.tandfonline.com/loi/tlus20 Using the land transformation model to forecast vacant land Galen Newman, Jaekyung Lee & Phil Berke To cite this article: Galen Newman, Jaekyung Lee & Phil Berke (2016): Using the land transformation model to forecast vacant land, Journal of Land Use Science, DOI: 10.1080/1747423X.2016.1162861 To link to this article: http://dx.doi.org/10.1080/1747423X.2016.1162861 Published online: 28 Mar 2016. Submit your article to this journal View related articles View Crossmark data
Transcript

Full Terms & Conditions of access and use can be found athttp://www.tandfonline.com/action/journalInformation?journalCode=tlus20

Download by: [Texas A&M University Libraries] Date: 28 March 2016, At: 11:20

Journal of Land Use Science

ISSN: 1747-423X (Print) 1747-4248 (Online) Journal homepage: http://www.tandfonline.com/loi/tlus20

Using the land transformation model to forecastvacant land

Galen Newman, Jaekyung Lee & Phil Berke

To cite this article: Galen Newman, Jaekyung Lee & Phil Berke (2016): Using the landtransformation model to forecast vacant land, Journal of Land Use Science, DOI:10.1080/1747423X.2016.1162861

To link to this article: http://dx.doi.org/10.1080/1747423X.2016.1162861

Published online: 28 Mar 2016.

Submit your article to this journal

View related articles

View Crossmark data

Using the land transformation model to forecast vacant landGalen Newman , Jaekyung Lee and Phil Berke

Department of Landscape Architecture and Urban Planning, Texas A&M University, College Station, TX, USA

ABSTRACTGrowing or shrinking cities can experience increases in vacant land. Asurban populations and boundaries fluctuate, holes can open in oncetight urban areas. Many cities chase growth-oriented approaches todealing with vacancies. It is critical to understand land-use alteration toaccurately predict transformations of physical change in order to makebetter informed decisions about this phenomenon. This research utilizesthe land transformation model (LTM), an artificial neural networkingmechanism in Geographic Information Systems, to forecast vacant land.Variable influence on vacant land prediction and accuracy of the LTM isassessed by comparing input factors and patterns, using time-series datafrom 1990 to 2010 in Fort Worth, Texas, USA. Results indicate that theLTM can be useful in simulating vacant land-use changes but moreprecise mechanisms are necessary to increase accuracy. This will allowfor more proactive decisions to better regulate the process of urbandecline and regeneration.

ARTICLE HISTORYReceived 8 January 2016Accepted 1 March 2016

KEYWORDSLand transformation model;geographic informationsystems; urban regeneration;land use/land coverprediction models; vacantland

1. Introduction

Urbanization is a global trend (Kourtit, Nijkamp, & Scholten, 2014). The World Bank/InternationalMonetary Fund (2013) suggests that many large cities may grow into future megacities (Kourtit,Nijkamp, & Reid, 2014). The United Nations (2011) projects that between 2011 and 2050, theworld’s population will increase by 2.3 billion, with urban populations rising from 3.6 billion to 6.3billion, nearly 67% of the world’s population. This increase will not be evenly distributed, as not allcities are growing in size/population (Shetty & Reid, 2014). Relatedly, the distribution of vacanturban land will also not be evenly distributed.

Vacant land (VL) is often only associated only with shrinking, or massively depopulating, cities.Interestingly, Bowman and Pagano (2004) found that that expanding cities (boundaries increase insize) with population increases tended to report higher levels of VL than did stable cities (bound-aries decrease or stay the same in size), which reported higher levels of structural abandonment.These conditions make it imperative that we better understand patterns of urban vacancy andimprove methods of assessing the urban condition in growing as well as shrinking cities (Brophy &Vey, 2002; Kozloff, 2007). In both growing and shrinking cities, there has been a restructuring ofurban economies away from manufacturing (Lester, Kaza, & Kirk, 2013). Many US Midwesternlegacy cities, older industrial cities experiencing sustained job and population loss over the pastfew decades (Mallach & Brachman, 2013), have borne the consequences of large-scale regionalshifts in population, primarily to the South and Southwest. Simultaneously, new technologies andservice-sector jobs in cities have increased urban-living demands but can drive-up housing prices.In growing cities, this can increase the pressure to convert urban-land uses to residential,

CONTACT Galen Newman [email protected] Department of Landscape Architecture and Urban Planning,Texas A&M University, 3137 TAMU, Langford Architecture Center A334, College Station, TX 77843, USA

JOURNAL OF LAND USE SCIENCE, 2016http://dx.doi.org/10.1080/1747423X.2016.1162861

© 2016 Informa UK Limited, trading as Taylor & Francis Group

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

sometimes limiting the ability to attract new industries. Most growing cities adopt urbanismstrategies based on peripheral or infill development in vacant lots, while a small portion ofmassively depopulating, or shrinking, cities have begun to simply decline smartly (Hollander &Németh, 2011). With smart decline, the goal is not to chase developmental incentives, but to createa quality-of-life standard based on the maintenance of vacant lots. The current urban challenge is,therefore, not just to focus on encouraging and managing urban growth, but also to find ways tomanage land-use changes and their resultant VL patterns.

Park and Rabenau (2015) suggest that spatial solutions, such as more accurate and proactiveland-use planning mechanisms, may be more effective than reactive policies in dealing withvacancies. Relatedly, land-use/land cover (LUCC) prediction models have emerged as means forparameterizing urban land-use changes, forecasting future situations, and simulating policy sce-narios (Landis, 2011). Current LUCC models are typically used to predict large scale land-usechanges, most seeking to predict urban growth. There is an urgent need for new urban modelsthat explore current problems such as the vacancy issue (Batty, 1992; Mallach, 2012b) but no singlemodel has been employed which specifically targets VL. An effective tool for forecasting VL couldhelp provide the basis for future infill and regeneration efforts in growing cities and/or targetedreprogramming for smart decline in shrinking/legacy cities.

This research tests existing VL use forecasting variables and their effectiveness in helping predictvacant urban land using the land transformation model (LTM), a geographic information systems(GIS)-based LUCC prediction tool designed to simulate urban land-use change based on thecomputation of land-use change probabilities (Pijanowski, Brown, Shellito, & Manik, 2002). TheLTM is analyzed according to input variable influence and prediction output accuracy throughcomparisons of the model’s predicted vacancies to actual vacancy occurrences, using time-seriesdata from 1990 to 2010 in Fort Worth, Texas, USA. Fort Worth is a growing city with expandingboundaries situated within the Dallas–Fort Worth–Arlington, TX Metropolitan Statistical Area, thelargest metropolitan area in the South. While the overall amount of VL has declined in the city since1990, rapid land-use changes, population migrations to the urban edge, and an urban corecharacterized by increased and fragmented VL parcels make the city a representative case fortesting the LTM. If the selected variables and the LTM prove reliable, similar methods could be usedto help forecast future VL in other growing cities or adapted to assist in predicting vacant patternsin shrinking/legacy cities.

2. Literature review

2.1. Effects and reuse of VL

It is difficult to disentangle the effects of urban VL from its causes as its effects typically becomecausal factors for its increase. Population decline is one of the most telling indicators of increasedVL; as people leave urban areas, the land declines and becomes underutilized (Goldstein, Jensen, &Reiskin, 2001). The effects of these vacant properties are various. Vacant properties can alienateportions of the local community from one another and create safety threats, thus lowering thequality of life (Kivell, 1993) and generating unattractive urban spaces which can then amplifycriminal activity (e.g. drug trafficking, gangs, arson, etc.). This, then, blights surrounding areas,deters future development, and decreases the ability to sustain economic growth (Kivell, 1993).These negative impacts can eventually contribute to decreases in property values and potentialdecline (Setterfield, 1997). As a neighborhood declines, rents drop and properties generate lessincome, making building maintenance difficult. In turn, more lots become unkempt, and morebuildings fall into disrepair, driving down property values further and making the lot a tax burdento its owners (Goldstein et al., 2001).

Both urban growth and decline are interrelated; as populations decline in one city, they mayincrease in another, even within the same region (Van den Berg, Drewett, Klaasen, Rossi, &

2 G. NEWMAN ET AL.

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

Vijverberg, 1982). While the effects of urban population changes are felt across multiple scales,these changes are felt most intensely at the local level (Shetty & Reid, 2014). As such, municipalshrinkage/growth should not be confused with regional shrinkage/growth. Shrinking cities can belocated in growing regions and vice versa. For example, between 1950 and 2010, Detroit’spopulation dropped from 1.8 million to 700,000, but the population of the Detroit–Warren–Livonia Metropolitan Statistical Area rose from 1.2 million to 3.8 million (Shetty & Reid, 2014).Even at the local level, population changes are not uniformly distributed as growing and decliningneighborhoods can exist simultaneously in different parts of the same city (Longoria & Rogers,2013). Relatedly, while suburbanization and sprawl are strongly linked to economic growth, theycan also be seen as contributing factors to urban depopulation and decline (Couch, Karecha, Nuissl,& Rink, 2005). Most solutions to the vacancy dilemma have tended to focus on razing andrebuilding the properties or raising taxes (Newman, 2013; Newman & Saginor, 2014). Contrary toexpectation, these attempts have only had a marginal effect. Many cities still have difficulty findingsolutions and have issues recovering their economies due to uncertainties in return on investment,ongoing costs for maintenance, low property values, and inaccurate analyses and predictions aboutfuture financial and environmental situations (Schilling & Logan, 2008).

The inability to retain viable developments in vacant areas has necessitated a pending shiftfrom traditional urban growth to the allowance of right sizing or smart decline (Hollander &Németh, 2011; Schilling & Logan, 2008) in which cities allow themselves to strategicallyrepurpose vacant lots rather than force new developments into decaying areas. Other, morerecent approaches such as tactical urbanism, guerilla urbanism, user-generated urbanism,insurgent urbanism, and pop-up urbanism are grassroots attempts to claim and repurposevacant lots through implementing temporary functions, urban agriculture, ecological functions,park space, or other community needs. Pagano (2014) refers to these approaches as Do-It-Yourself Urbanism, a label applied to the re-appropriation of unused urban space for commoncommunity needs. Urban-planning concepts such as sustainable development and SmartGrowth have also been widely discussed as solutions to the vacancy issue (Berke, Godschalk,& Kaiser, 2006; Kaiser & Chapin, 1995). VL represents a significant opportunity for Smart Growthprinciples such as infill, brownfield, and grayfield redevelopment and the creation of urbangreen space (Németh & Langhorst, 2014). Retrofit developments, reclamation projects, andrestoration/adaptive reuse techniques are all examples of how development managementprograms have attempted to reintegrate vacant properties into productive societal space(Fulton, 2006). Successful implementation of these approaches, however, is heavily reliantupon knowledge about urban land-use change.

2.2. Urban land-use modeling

Approaches to repurposing VL are gaining momentum; broader land-use planning mechanisms arealso being scientifically integrated and technologically driven to aid in this purpose (Foley et al.,2005; Gutman et al., 2004; Rindfuss, Walsh, Turner, Fox, & Mishra, 2004; Verburg, 2006; Verburg,Schot, Dijst, & Veldkamp, 2004). Digital models of LUCC have become powerful tools to helpunderstand and analyze important connections between socioeconomic processes and urbanprocesses (Turner & Meyer, 1991). Current research uses these models to support analyses ofboth causes and consequences of land-use dynamics (Verburg et al., 2004).

Econometric models using multiple regression analysis are a common statistical techniquefor land-use forecasting (Briassoulis, 2004; Chapin & Weiss, 1968). There are several otherapproaches involving interactions between land-use patterns and environmental and socio-economic elements such as spatial interaction models and spatial input–output models (Batty,2009). These are primarily digital models which can be less rigorous, but more user friendly fordecision makers. As computing systems continue to develop, data sources have become moreorganized, accessible, and easier to calibrate. As a result, computer simulation models such as

JOURNAL OF LAND USE SCIENCE 3

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

UrbanSim (Waddell, 2002), Cellular Automata (Batty, Xie, & Sun, 1999; Torrens, 2003), andSLEUTH (Clarke, Hoppen, & Gaydos, 1997) have grown remarkably. These models can bemore effective for land-use planners because they provide graphical outputs and rely notonly on economic theory, but also real situations and urban trends. The ability to generateroad maps for local policy makers, developers, and residents who are not familiar with urbantheories and statistics is much easier.

2.3. The LTM

Regression-based LUCC models seek to establish relationships between spatial predictor variablesand locations of predicted change (Theobald & Hobbs, 1998). These include logistic regression(Landis, 1994), hedonic price (Alig, 1986; Geoghegan, Wainger, & Bockstael, 1997), and artificialneural network (ANN’s) models (Pijanowski et al., 2000). More recently, spatial analytic programssuch as GIS have been used as vehicles to house LUCC models. A GIS-based model known as theLTM has risen in popularity as one of the more accurate ANN-based LUCC models to forecast land-use changes.

The LTM is a prediction tool to analyze spatial and temporal land-use dynamics, estimate theimpacts of changes, and forecast future land uses (Pijanowski et al., 2002). It was originallydeveloped to simulate local-scaled LUCC patterns (Pijanowski et al., 2000; Pijanowski et al., 2002;Pijanowski et al., 2014; Pijanowski, Long, Gage, & Cooper, 1997) by the Human–EnvironmentModeling and Analysis Laboratory by Pijanowski et al. (1995) from Purdue University. The LTMfollows four sequential steps: (1) data processing – input data are stored and managed within GIS,(2) spatial rule application – predictor variables are related to land-use transitions for each locationare examined, (3) grid integration – all cell sizes are set to a fixed base layer and the relativelikelihood of change for each cell is derived based on the ANN, and (4) temporal scaling ofprediction output – the amount of land expected to transition is determined (Pijanowski et al.,2002) (see Figure 1).

Figure 1. Conceptual diagram of an artificial neural network (ANN) showing input drivers and the process of the ANN’s typicaloutput.

4 G. NEWMAN ET AL.

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

2.4. ANNs

ANNs are self-programming networks that find and resolve complex interactions between variableinputs and predicted outputs (Vafeidis, Koukoulas, Gatsis, & Gkoltsiou, 2007). They allow the use ofdifferent types of variable/driver input data which can range from cultural, social, economic, and/orenvironmental factors (Almeida, Gleriani, Castejon, & Soares-Filho, 2008; Guan, Wang, & Clarke,2005; Li & Yeh, 2002; Moore, 2000; Yeh & Li, 2003) and have been shown as appropriate formodeling land-use change, if appropriate spatial data exists (Clarke et al., 1997; Fischer & Abrahart,2000; Openshaw, 1998). ANNs consist of input layers, output layers, and optionally hidden orexclusionary layers (Bishop, 1995). The input layers are driven by the logic of the modeler andconsist of the variables which drive the connecting neurons. These connections decipher solutionsbetween inputs (i.e. drivers of change) and outputs (e.g. locations of change occurring betweentwo time periods) (Pijanowski, Pithadia, Shellito, & Alexandridis, 2005). The ANN learns patternsusing historical land-use data from at least two different time periods, calculates the changebetween these periods, and uses this change as an influencing raster data set alongside theinput drivers (Tayyebi et al., 2013).

Neural nets train themselves through numerous cycles. Short cycles of training have beenshown to produce relatively poor results (Reed & Marks, 1998). Conversely, an excessive numberof cycles for a neural net model can result in an over fitting of data (Pijanowski et al., 2005). Moretraining does not necessarily produce the best goodness-of-fit (Pijanowski, Alexandridis, & Müller,2006). Model accuracy is calculated using a Kappa statistic which calculates the relative percentagesuccess of a model (Sousa & Kaymak, 2002). Typically, models are run with 10 or more input driversand those which have a 40% or more accuracy rate are said to perform with relatively highpredictive ability with more than 60% being highly acceptable predictability (Almeida et al.,2008; Pijanowski et al., 2006, 2002; Tayyebi et al., 2013).

The benefits of using ANN-based models identified by Pijanowski et al. (2002) include: the abilityto generalize well across datasets as well as theoretically across space and time, a relatively lowsensitivity to errors when training data (a typical limit when using land-use data), and a capabilityto handle nonlinear patterns in data. As a further verification, Pijanowski et al. (2005) tested LTMpredictions for land-use change within regions and compared these predictions against actualobserved change. Cells predicted to transition to urban were compared with cells that actuallytransitioned and a Kappa statistic for each model was calculated. The study found that neural net-based models performed well for predicting LUCC, the number of training cycles built into themodel produced differing outputs, and that the Kappa statistic was a relatively reliable test foraccuracy in areas undergoing rapid land-use change (Pijanowski et al., 2005).

Only a small amount of research has tested the accuracy of LUCC models (Conway, 2009).Herold, Couclelis, and Clarke (2005) used Santa Barbara, California, USA to develop a framework formodeling land-use change combining remote sensing and urban growth which would eventuallyamplify research on evaluating the effectiveness of LUCC models. It developed spatial metricsaimed at improving model analysis. Prior to this, tests on applications of LUCC models at differingscales were the primary means of assessment. Pijanowski et al. (2002) presented a version of theLTM using Michigan’s Grand Traverse Bay Watershed to explore how landscape characteristicsinfluenced urbanization patterns, finding that the predictive ability of the LTM improved whenapplied at coarser scales and using higher multitudes of predictor variables. However, the indivi-dual contribution of each predictor variable was shown to vary across spatial scales. Conway (2009)attempted to predict individual land uses and assess the potential for accurate predictive changeby studying the impact of changing the number of land use classes on a model’s predictions in theBarnegat Bay Watershed in New Jersey, USA. The research found that changing the level of classresolution impacted the predicted outcomes, but did not necessarily improve model accuracy.

It has become customary in remote sensing research to report both the Kappa index ofagreement and the proportion of observations classified correctly for the purposes of accuracy

JOURNAL OF LAND USE SCIENCE 5

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

assessment (Pontius & Millones, 2011). Kappa analysis is considered a required component of mostimage analysis software packages which include accuracy assessment procedures (Congalton,Oderwald, & Mead, 1983). While the Kappa statistic is one of the typical means of accuracyassessment, it does have its limitations. On the surface, the Kappa index seems to be a fairlyappropriate approach to assess prediction accuracy. The use of only Kappa can be somewhatlimited when judged on its ability to accurately evaluate both quantity and location accuracy ingrid cells on a map (Pontius, 2000). Pontius (2000) is somewhat critical of only using the Kappascore to assess accuracy, describing it as a one-dimensional index which does not specify reasonsfor the disagreement between the utilized maps. He states that the Kappa index can sometimesmuddle information about the quantity of each category in the maps with information about thelocation of each category in the maps. For these reasons, quantity disagreement and locationdisagreement for general map comparison can also sometimes provide additional insight (Pontius& Millones, 2008). ROC/AUC (receiver operating characteristic/area under curve) analysis can also beused to quantify the accuracy and validity in binary classification (henceforth, these binary classesare referred to as vacant and nonvacant) by comparing actual maps and predicted maps (Manel,Williams, & Ormerod, 2001; Osborne, Alonso, & Bryant, 2001). This research utilizes all four measuresto assess prediction accuracy.

2.5. Applications of the LTM

The LTM was developed 15 years ago and has now been utilized in a variety of places aroundthe world (Pijanowski et al., 2014). The model has been applied in the United States and incentral Europe (Pijanowski et al., 2006), East Africa (Olson et al., 2008; Washington-Ottombreet al., 2010), and Asia (Pijanowski, Tayyebi, Delavar, & Yazdanpanah, 2009). Brown, Pijanowski,and Duh (2000) used the model to predict regional forest-cover changes in the Upper Midwest,USA and showed that land use and land cover had a linear functional relationship. Tang, Engel,Pijanowski, and Lim (2005) used the LTM to investigate urbanization patterns on a watershedscale, predicting change by the years 2020 and 2040. The research attempted to link LUCCmodels with environmental impact assessment models, using the LTM to generate informationabout future urbanization patterns and potential environmental impacts. Similarly, Ray andPijanowski (2010) ran the LTM backward in an effort to examine the impacts of land-usemorphology on environmental processes. The LTM has also recently been coupled with mesos-cale drivers to project urban growth using multiple city-scaled projections combined andassessed on a national scale. Tayyebi et al. (2013) found that LTM models performed relativelywell using this method and that the introduction of small scaled data into large-scale LTMsimulations significantly increased model accuracy. More current applications of the LTM arebeing explored to forecast civic boundary expansion to help predict and regulate urban growth(Tayyebi, Perry, & Tayyebi, 2014). Given the differing objectives of using LUCC modeling, the lackof one overarching theory of land-use science and the multitude of disciplines applying thesemodels, it is impossible to identify the model which has the highest potential in meeting thechallenges presented by urban growth (Verburg, 2006). Attempts to compare outputs ofdifferent LUCC models are rare, but do exist. Pontius et al. (2008) used multiple resolutionmap comparative methods to assess 13 applications using 9 different peer-reviewed LUCCmodels. The research sought to determine how well each model performed and concludedthat the same model could have different outputs in differing settings, as 12 of the 13 LUCCmodels contained more erroneous pixels than pixels of correctly predicted land change (Pontiuset al., 2008). Many of these errors were shown to occur due to the massive (e.g. regional,national, or global) scales in which LUCC’s are generally applied. Applications of the LTMexamining small-scaled (e.g. municipal or neighborhood) urban land-use changes still meritnecessary research (Pijanowski et al., 2014) and there is an awareness of scale dependencies andup-scaling problems in existing models (Verburg et al., 2004).

6 G. NEWMAN ET AL.

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

3. Literature gaps and research objectives

More applications of the LTM and investigations on its accuracy are needed. Simultaneously,predictions for individual land uses have not been thoroughly conducted. These predictionsmust also undergo a thorough analysis of the drivers used to predict each land use for thepurposes of validation. The new generation of land-use modelers must better address the multi-scale characteristics of the land-use system (Verburg et al., 2004). In fact, a recent report by theGeographical Sciences Committee’s (2014), Needs and Research Steering-Committee on advancingLUCC modeling suggests improvements which need to be made to advance LUCC modeling; aprimary opportunity suggested for advancing modeling efforts is listed as new data on landchange and the actors involved with these alterations. Further, the aforementioned disagreementson prediction assessment methods can sometimes be problematic and not much research hasexamined multiple assessment methods for prediction accuracy simultaneously. Municipal-scaledpredictions and accuracy assessments using LUCC models are also rare. No models have beencreated which specifically target VL uses. Land uses are constantly in flux; at any given point, avacant parcel can become activated and vice versa, making them difficult to study. Ironically, VL is aubiquitous urban land use, making it the most common variable for assessing urban growth. Thisresearch tests the accuracy and reliability of the LTM on VL prediction and examines the potentialinfluence of the variables used for predicting. The focus is to use the LTM to initialize and test abaseline set of drivers for VL prediction. This intention will, hopefully, initialize a process of testingdifferent prediction models and developing a reliable model for forecasting future urban vacancy/abandonment. The objectives of this research are twofold: (1) to test the LTM as a feasible optionfor VL prediction and (2) to determine each variable’s effect on prediction output in order to makewell-informed decisions on the primary influences of urban vacancy.

4. Methods

4.1. Study area and data

The City of Fort Worth in Tarrant County, Texas, USA is utilized as the study area. Tarrant County(1809,034) is the third most populous county in Texas (U.S. Census, 2010). Since 1990, there was anapproximate 54.6% increase in population (U.S. Census Bureau, 2015). GIS data for the site wereobtained from the City of Fort Worth (annexation data and land-use plans), the North CentralCouncil of Governments (time series land-use and VL inventories) the Texas Natural ResourceInformation System (land-cover data), and the U.S. Census Bureau (demographics and economicdata) was used to make predictions.

The City of Fort Worth (2014) defines VL as three land uses, brownfields, vacant structures/housing units, and vacant agricultural. Vacant brownfields are described as underutilized, obsolete,or structurally deteriorated industrial or commercial properties where improvements are hinderedby real or perceived contamination. Vacant structures/housing units include VL which contains ahouse, apartment, mobile home or other unit, vacant but intended for occupancy as separate livingquarters. Vacant agricultural describes areas with one residential unit per structure on a one ormore acre lot with no city water or sewer service; or land with no existing buildings, except forthose related to mining, crops, or grazing.

From 1983 to 2012, Fort Worth annexed over 40,000 acres of land; from 1990 to 2010, the totalamount of VL in Fort Worth dropped nearly 50% to about 12.3%, with a large majority of currentvacant parcels residing in the city’s core. Figure 2 shows historical land-pattern changes between1990 and 2010 in 10-year increments. A spike in VL amount in 1995 (50%) was largely due to large-scale annexation during this period but it quickly decreased as new development on the peripheryrapidly replaced newly annexed vacant parcels. Overall, the city decreased in VL by 37.6% from1990 to 2010 while increasing in size by 1.2% (from 182,387 acres to 222,912 acres). The total

JOURNAL OF LAND USE SCIENCE 7

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

amount of vacant parcels, however, actually increased from 2772 parcels in 1990 to 25,442 in 2012but has relocated from the civic edge to the urban core. Parcel changes, however, are highlyrelated to platting efforts by the city, as the overall parcel count increased from 43,446 in 2005 to254,442 in 2010. These continual alterations make Fort Worth a good candidate for studying theLTM because LUCC models that are applied to landscapes which have larger amounts of netchange have been shown to have higher rates of predictive accuracy (Pontius et al., 2008).

4.2. LTM process

The LTM uses two different types of input drivers to predict land uses: raster-based causal variableslinked to a spatial location (henceforth referred to as input factors) and raster-based historical landuse maps from which land-use change rasters are generated (henceforth, referred to as inputpatterns). Assessment of the LTM’s accuracy in this research is based on a four-tiered process: (1)assessing the influence of each input factor on model performance through alternative versions ofthe model, (2) comparisons of actual vacancy rates to predicted vacancy rates using 5-year inputpatterns but differing amounts of input factors, (3) the examination of LTM-based output statistics(Kappa and percent correct match) as well as quantity and allocation agreements and ROC/AUC,and (4) assessment of output similarities for predicted VL for the year 2020 using differing inputpatterns but identical input factors.

The LTM model is constructed based on VL pattern changes using GIS data from 1990, 1995,2000, 2005, and 2010. Input factors were entered using 14–16 variables (see Table 1), which havebeen shown to contribute to VL increase. Three exclusion areas were omitted from the analysis dueto their specialized functionality (e.g. military bases, airports, and existing vacant areas). Using therasterized variables as input factors, the input patterns (historic VL inventories) and the threeexclusionary layers, we forecasted future VL patterns (Figure 3). Initially, five differing input patternsfrom 1990 to 2010 (in 5-year increments) were exported at a resolution of 30 × 30 m (roughly100 × 100 ft).

4.3. Variable selection

Variable selection can greatly alter the outcome of predictions; each input can sway the resultshigher or lower. There are several studies that identify the principal causal mechanisms contribut-ing to vacant urban land but there is a lack of quantification of the exact influence on VL accretion.These causes vary and can be more influential based on the type of city (e.g. shrinking or growing).Primary causes can be classified into four differing, but overlapping categories which are primarilya product of population and demographic makeup/alteration: (1) deindustrialization or shifts froman industrial to service economy (Buhnik, 2010; Lindsey, 2007; Németh & Langhorst, 2014; Rieniets,2009), (2) weak market conditions and downturns (Bontje, 2005; Johnson, Hollander, & Hallulli,

Figure 2. Vacant land patterns and ratios of vacant land to city area in Fort Worth, TX, USA from 1990 to 2010 in 5-yearintervals.(Data sources: City of Fort Worth, GIS Program).

8 G. NEWMAN ET AL.

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

Table1.

Yearsexam

ined,level

ofdata,d

escriptio

n,andprevious

research

oninpu

tfactorsutilizedas

driversforvacant

land

predictio

n.

Inpu

tfactors

Inpu

tpatterns

Data

level

U.S.C

ensusdefinitio

nReferences

forinpu

tfactors

90–

0090–

1000–

1090–

9590–

0500–

05

Land

value

XParcel

Land

valuepersquare

meter

ofeach

parcel

inthebo

ttom

30%

Glaeser

andGyourko

(2001),C

apozza

andHelsley

(1989),D

ong(2013),A

ryeetey-Attoh,

Costa,

Morrow-Jon

es,M

onroe,andSommers(1998)

Marketvalue

XParcel

Land

value+Improvem

entvaluepersquare

meter

ofeach

parcel

inthe

bottom

30%

Glaeser

andGyourko

(2001),G

laeser,G

yourko,and

Saks

(2006),A

nas(1978),P

ondandYeates

(1993)

Hou

sevalue

XX

XX

XCensus

tract

Hou

seValueforallo

wner-occupied

housingun

its:lessthan

$50,000

Glaeser

andGyourko

(2001),G

laeser

etal.(2006),

Anas

(1978),A

ryeetey-Attohet

al.(1998)

Hou

seho

ldincome

XX

XX

XX

Census

tract

Hou

seho

ldIncomein

thebo

ttom

50%

Glaeser

(2013),Fee

andHartley(2011),R

yan(2012),

Aryeetey-Attoh

etal.(1998)

Highw

ayX

XX

XX

X–

Proximity

tohigh

ways

Rapp

aport(2003),Bou

rne(1996),D

ong(2013),Lester

etal(2013)

Railroad

XX

XX

XX

–Proximity

torailroads

Rapp

aport(2003),B

ourne(1996),Lesteret

al.(2013)

Parcel

size

XX

XX

XX

Parcel

Parcel

size

smallerthan

465square

meters

Colwelland

Mun

neke

(1997)

Carrion-Flores

andIrw

in(2004),Pon

dandYeates

(1993),Lesteret

al.(2013)

Hom

eow

nership

XX

XX

XX

Census

tract

Owneroccupied:lessthan

40%

Bradford

(1979),P

ondandYeates

(1993),A

ryeetey-

Attohet

al.(1998),Park

andCiorici(2013)

%below

povertyrate

XX

XX

XX

Census

tract

Individu

alpo

vertyrate:o

ver20%

(individu

alsbelow

poverty=‘und

er0.50’+

‘0.50–0.74’+

‘0.75–0.99’)

Glaeser

(2013),Fee

andHartley(2011),R

yan(2012),

Park

andCiorici(2013)

%Civilian

unem

ployment

XX

XX

XX

Census

tract

Unemploymentrate:o

ver10%

(based

ontotalp

opulation16+.C

ivilian

person

sun

employed

dividedby

totalcivilian

popu

latio

n)FeeandHartley(2011),A

ryeetey-Attohet

al.(1998)

Vacancyrate

XX

XX

XX

Census

tract

Vacancyrate

(occup

ancy

status):over

10%

Don

g(2013),M

allach

(2012a)

Rate

ofservice

indu

stry

XX

XX

XX

Census

tract

Serviceindu

stry

rate:Lessthan

30%

(Inform

ation|Finance,insurance,real

estate

andrentalandleasing|P

rofessional,scientific,andtechnical

services

|Edu

catio

nal,health

andsocial

services

|Arts,entertainm

ent,

recreatio

n,accommod

ation,

andfood

services)

Glaeser

(2013),Fee

andHartley(2011),M

allach

(2012a),Glaeser

andKahn

(2004),Lesteret

al.

(2013)

Rate

ofsecond

ary

indu

stry

XX

XX

XX

Census

tract

Second

aryindu

stry

rate:o

ver35%

(mining,

constructio

n,manufacturin

g)Glaeser

(2013),Fee

andHartley(2011),M

allach

(2012a),Glaeser

andKahn

(2004),W

egener

(1982),

Don

g(2013)

Ageof

buildings

XX

XX

XX

City

Built

before

1950

(exceptbu

ildings

inhistorical

preservatio

ndistricts)

Wegener

(1982),N

ewman

(2015)

Ethn

icity

XX

XX

XX

Census

tract

Non

-white

popu

latio

n:over

50%

Ryan

(2012),Fee

andHartley(2011)

Education

XX

XX

XX

Census

tract

Percentof

person

s25

yearsof

ageandolder,with

less

than

orequalto

high

-schoolg

radu

ate(in

clud

esequivalency)

Glaeser

(2013),Fee

andHartley(2011),M

allach

(2012a),Park

andCiorici(2013)

Popu

latio

nchange

XX

XX

Census

tract

Zero

ornegativepo

pulatio

nchange

betweeneach

perio

dWegener

(1982),P

ondandYeates

(1993),D

ong

(2013)

Numberof

variables

1515

1514

1416

––

JOURNAL OF LAND USE SCIENCE 9

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

2014; Ryan, 2012), (3) decreasing personal wealth (Audirac, 2007; Cunningham-Sabot & Fol, 2009;Rybczynski & Linneman, 1999), and (4) odd physical characteristics/bad location (Cunningham-Sabot & Fol, 2009; Henry, Schmitt, & Piguet, 2001; Németh & Langhorst, 2014). These factors canresult in an oversupply of VL which can then depress land prices, property values, and tax revenues,increase abandonment, decrease employment rates, sales, investments, and vitality and result inlosses of residential, commercial, and business activities (Schilling & Logan, 2008). With this in mind,the next step was to collect appropriate input factors which could potentially contribute to VL (seeTable 1). Each model was composed of up to 16 input factors, depending on data availability.

Depopulation is typically noted as the primary cause of VL. In this study, areas with zero ornegative population change between time periods were assumed to increase in VL. Remainingpopulations require higher education to stay in the workforce or retain viable long-term housinglots (Fee & Hartley, 2011). Many studies have shown that the level of educational attainment ofoccupying populations can affect urban decline (Fee & Hartley, 2011; Glaeser, 2013; Wachter &Zeuli, 2013). For example, Mallach (2012a) pointed out the percentage of adults who hold abachelor or higher degree is one of the most critical factors contributing to urban shrinkage.Therefore, it was assumed that as the percent of persons 25 years of age and older with less than orequal to a high-school graduate education level increased, VL would also tend to increase.

When businesses leave an urban area, so do their tax revenues and the secondary economicactivity generated by these businesses. Those that remain in urban areas devoid of thiseconomic activity can be low-income and/or minority populations that face declining municipalservices funded from a smaller tax base and lower diversity of retail offerings; this results inhigher prices, fewer job opportunities, and a lower quality of life (Goldstein et al., 2001). Theconcentration of minority populations and enclaves of isolated ethnicities can be a negativecomponent contributing to VL (Ryan, 2012). Over time widespread racial discrimination seen inhiring trends, can systematically limit relocation options for minorities (Massey & Denton, 1993).When a neighborhood loses jobs, minorities can also have fewer housing choices, furtherincreasing racial concentrations (Hollander, 2010). Based on this premise, the model assumedthat areas which were over 50% minority in make-up would be affected by increases in VL inthe future.

Manufacturing employment is down by about 40% from the 1970s, when it was at its peak,while service employment has nearly doubled (Glaeser, 2013). Only 9% of the workforce is currentlyin manufacturing, down from 22% in 1979, while the service share jumped from 54% to 69%(Glaeser, 2013). Many urban regions experiencing depopulation held a large share of urbanmanufacturing industries at one time. This shift has created a tremendous population and eco-nomic decline in some cities where manufacturing was historically concentrated (Fee & Hartley,2011). This condition makes industry makeup, services offered, and unemployment variables toconsider. Unemployment rates were calculated based on the number of civilian persons unem-ployed divided by total civilian population; higher unemployment rates (over 10%) were assumedto increase VL. Areas with low amounts or decreasing service industries (less than 30%) and areaswith higher secondary industry rates (over 35%) were also assumed to increase in VL in the future.

Market conditions change across cities and scales and are a factor which can contribute to VL; allcities contain widely varying housing market conditions and these conditions include both areasexperiencing abandonment and well-maintained neighborhoods (Heckert & Mennis, 2012). From2000 to 2010, the total number of vacant housing units in America increased by over 4.5 million, anescalation of 44% (Mallach, 2012c). However, different situations are reflected across differingmunicipalities. It should also be noted that economic situations change across time which canresult in housing market anomalies. For example, the 2008 U.S. housing crash had a direct negativeimpact not only on housing values, but mortgage markets, home construction and real-estatemarkets. Many cities continuing to register population losses currently are doing so despite majorchanges in urban market conditions nationally; these cities have not benefited from national/regional changes and may be unlikely to do so in the future (Mallach, 2012a). While revitalization

10 G. NEWMAN ET AL.

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

efforts are underway in many of these cities, they often have an immense oversupply of housingand other buildings relative to demand. Areas with 10% or higher vacancy rates were assumed tocontribute to future increases in VL. Vacancy rates can increase when housing demands ultimatelysink to a certain threshold, owners tend to abandon their structures, and these spaces becomevacant over time (Keenan, Lowe, & Spencer, 1999). Simultaneously, owner occupancy rates arehighly related with vacancy. Areas which are occupied by renters rather than owners can becomesuperannuated and affluent home-owners of these area tend to move out (Farris, 2001). Medianbuilt year of structures and rate of home-ownership are, then, interrelated factors contributing toVL (Temkin & Rohe, 1996). Any structure built before 1950 which was outside of a historic districtwas assumed to become a probable future vacant site while areas with 40% or more not owneroccupied were also assumed to contribute future VL.

Factors associated with individual personal wealth are highly associated with job availability.Jobs follow people and people follow jobs; as employers and residents move out of urban areas,property values and income can decrease and poverty can become concentrated in specific areas(Park & Ciorici, 2013). Personal wealth, in this research, captures the linkages between wealth andpoverty as measured by four different variables at two different levels of data: property values (landvalue and market value) at the parcel level, and median household (income and poverty rate) atthe block group level. It was assumed that a lack of wealth and lower land and market valueswould result in increased VL.

Parcel size is associated with the ability to regenerate a vacant parcel. Irregularly shaped/smallparcels can be difficult to develop and are sometimes referred to as leftover/remnant properties(Lester et al., 2013). These lots are typically re-used as temporary purposes. The area of each parcelwas calculated and parcels less than 465 m (5000 ft2) were selected as remnant parcels. The smallerthe parcel size, the more prone it was assumed to remain vacant.

Many people who have left urban areas commute suburb-to-suburb rather than into the city.Proximity to highways and railroads represent the accessibility and connectivity (especially to thesuburbs), and refer to the enlarged size of the geographic area in which people can both live andwork (Rappaport, 2003). Transportation services can either improve the access of the site to largerurban areas (e.g. county roads and highways) or increase the amenity value of a site (e.g.

Figure 3. LTM process diagram showing input factors, input patterns, and output comparison maps.

JOURNAL OF LAND USE SCIENCE 11

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

shorelines, inland lakes, rivers). Since access to urban services affects development patterns, it isexpected that sites nearer to existing major transportation lines would be more likely to develop.The distances from the center of highway and railroads were classified in five different categoriesusing 1 km radii.

4.4. LTM application

After assimilating the data, the neural network was trained and simulation cycles were run for eachcombination of time periods. All input factors were normalized to a range from 0 to 1, and outputlayers became binary data depending on cell location change (1 = vacant, 0 = nonvacant). The ANNwas then trained using the drivers for calculating the predicted changes of VL between the timeframes. Usually, over 4000 cycles of training are necessary to stabilize the error level to a minimumvalue (Pijanowski et al., 2002). We ran each training session up to 100,000 cycles. The cycle with thehighest match rate between real change and simulated model was then selected for futureprediction purposes and assessment. The spatial accuracy of the models were automaticallycalculated every 1000 cycles. The LTM produced two automated statistics, percent correct metric(PCM) and Kappa values, detailing how well the real change and predicted change between thetime frames matched one another. The LTM calculates the PCM as follows:

PCM ¼ Cells correctly predicted to change� 100ð ÞCells actually transitionedð Þ

In the field of remote sensing, there has been a considerable debate about accuracy assessmentof observed sample data. Generally, the Kappa coefficient is thought to be the most standardmethod, but it is not always the most appropriate (Foody, 2002). As noted, Pontius and Millones(2011) recommends using two measures: quantity disagreement and allocation disagreement. Forthe purposes of this research, we utilize both kappa coefficients, PCM, and agreement/disagree-ment measures. The calculation process provides quantity disagreement, allocation disagreement,and overall agreement. As the result of neural network training, actual, and probable transitionstoward vacancy between two periods can be compared visually and mathematically through usingthe Raster Calculator tool in ArcGIS. By juxtaposing real and predicted maps, one can assess fourdifferent conditions: 0 = no real and no predicted transition (true negative), 1 = no real butpredicted (false negative), 2 = real but no predicted (false positive), and 3 = both real and predictedtransition (true positive). Based on the scores, the total number of pixels of quantity disagreement,allocation disagreement, and overall agreement can be calculated. Figure 4 indicates a possibleagreement between an actual (A) and predicted map (P). As shown in the figure, both (A) and (P)have same quantity of transitioned pixels, meaning no quantity disagreement. Four of the ninepixels are in differing locations, suggesting the allocation disagreement output of 0.44. Theremaining pixels overlap, resulting in an overall agreement of 0.56. The difference between thePCM and overall agreement is that the PCM only calculates the agreement of transitioned pixels,while the overall agreement shows more a general agreement in entire study areas.

Based on the concept, the values of quantity disagreement, allocation disagreement, and overallagreement are formularized as follows:

Quantity disagreement ¼ 1þ 3ð Þ � 2þ 3ð Þf gj j0þ 1þ 2þ 3ð Þ

Allocation disagreement ¼ 1þ 2ð Þ0þ 1þ 2þ 3ð Þ

Overall agreement ¼ 0þ 3ð Þ0þ 1þ 2þ 3ð Þ

12 G. NEWMAN ET AL.

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

Similarly, ROC (or AUC) analysis can also be used as measures for comparing actual andprobable transition maps. It uses the true positive rate (TPR), the proportion of positives that arecorrectly identified, and false positive rate (FPR), the proportion of negatives that are correctlyidentified to determine accuracy. In this research, TPR measures the percentage of vacant area thatis correctly identified as becoming vacant, while FPR as such the percentage of nonvacant area thatis correctly identified as not becoming vacant. Based on the TPR and FPR, a ROC plot is depicted onthe x-axis at a ratio of 1-specificity, showing the FPR and on the y-axis sensitivity and representingTPR. TPR and FPR are calculated using the formula below, based on the overall agreement cellscore outputs:

TPR ¼ 33þ 1ð Þ

FPR ¼ 22þ 0ð Þ

5. Results

5.1. Variable influence

Predictive analysis can become difficult when multiple qualitative, or noisy, input variables are utilized(Lee, Hanna, & Loh, 2004). ANN’s are powerful tools which seek to model complicated problems, butthey are essentially clean slates which provide little insight as to the importance of the each factorutilized. The modeler chooses the factors to input and most statistical models can have difficultiesdealing with highly correlated data (Lee et al., 2004). To determine the influence of each input factor,PCM and Kappa statistics are compared using different prediction models. Using the influence testapproach developed by Pijanowski, Shellito, Bauer, & Sawaya (2001), the effect of each factor onmodelperformance was quantified by creating 16 alternative versions of the model using input patterns from2000 to 2010. Each model was created by dropping one variable out, and repeating the ANN trainingprocess described earlier. New PCM and Kappa statistics were generated for each dropped variablemodel to determine if the results would have higher scores than the overall model.

Overall, the input factors proved to be sound drivers of VL increase. When compared to the fullmodel (all input factors), only one dropped variable model showed higher output statistics. Table 2displays each excluded variable and compares these PCM and Kappa outputs to the full modelresults and ranks each variable by its influence of change on the overall model from high to low(1 = lowest and 16 = highest). The number of probability trainings varied from 3500 to 90,000. Of allthe input factors utilized, only when dropping the rate of secondary industry (proportion ofsecondary industries to all industries) did the model produce a higher PCM and Kappa than theoverall model. This suggests that this factor may not be as influential as literature suggests whenpredicting for VL. This may be partially due to the fact that many existing manufacturing industriesmay not be deindustrializing in a growing city such as Fort Worth. This factor may be more influentialin legacy or shrinking cities, considering their higher rates of deindustrialization. Poverty rates provedto be not as influential as previously assumed, as the PCM, was equal to the full model output whenthe factor was excluded, but showed a slightly lower Kappa coefficient (0.01 less) than rate ofsecondary industry. The lower Kappa suggests that the factor is influential, but only marginally.

Not surprisingly, market condition and economic variables such as market value and land valueseemed to have a stronger influence on the model than most other factors. PCM and Kappastatistics decreased immensely when these variables were removed. Personal wealth variables suchas income and employment also showed a strong influence on increasing prediction accuracy.Surprisingly, however, it was expected that population change would be the primary demographicfactor in predicting VL but, while a necessary input factor, another demographic variable, ethnicity,was actually more influential on improving the model’s accuracy. Physical and locational

JOURNAL OF LAND USE SCIENCE 13

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

characteristic variables had weak but positive influences. Not surprisingly, proximity to highwaysproved to be a stronger influence than proximity to railways.

5.2. Identical input patterns with differentiation of input factors

Comparisons of actual vacancy rates were made to predicted vacancy rates using 5-year inputpatterns but differing amounts of input factors. Table 3 shows the total number of variables utilizedin each scenario to predict vacancy, the highest probability or each training cycle and the PCM andKappa coefficient for each time increment examined. For example, during the first 5-year period(between 1990 and 1995) 14 variables were utilized as input factors for each increment and 16,531cells changed between each increment at a resolution of 100 ft. The 45,000th training had thehighest training probability and the PCM rate at this cycle was approximately 49.3%. Generally,60–80% accuracy is exceptional and 40–60% of PCM is considered a highly acceptable model(Almeida et al., 2008; Pijanowski et al., 2006, 2002; Tayyebi et al., 2013). The same process andmethods were adopted for all other time periods examined.

Table 4 indicates the results of the statistics of quantity disagreement, allocation disagreement andoverall agreement of the model. Since the LTM controls the number of transitioned pixels, there couldbe no quantity disagreement. Typically, having values higher than 85% for the overall accuracy andover 40% for the PCM are considered an acceptable model (Almeida et al., 2008; Foody, 2002;Pijanowski et al., 2006, 2002; Tayyebi et al., 2013). Results for all comparisons yielded high enoughPCM and overall agreements to be considered highly acceptable. The AUC values for each modelranged from 0.71 to 0.77, indicating a consistent model accuracy. Overall, 0.70–0.80 of AUC is generallyconsidered a fair or good model (Osborne et al., 2001; Rutherford, Bebi, Edwards, & Zimmermann,2008). There were also consistencies in the overall agreement and PCM measures for each time framepredicted. Other than the 2000–2010model, a lower PCM tended to also have lower overall agreement.

LTM prediction maps were modeled for the years 2000, 2005, and 2010. Prediction output mapswere then compared to the actual historic VL inventories (see Table 5). Results revealed a few

Figure 4. Diagram showing agreement calculation process.

14 G. NEWMAN ET AL.

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

salient points. First, the LTM tended to consistently over-predict. When averaging all predictionoutputs, the mean was around a 16.8% over prediction of area. Second, the use of newer data setsfor VL inventories (within at least 5 years of predicted year) proved to have a higher accuracy inpredicted area with an 18.5% over prediction compared to older data sets which over predictedarea an average of 21.6%. Third, the time frame of input patterns utilized did not seem to affectprediction accuracy considerably. The 5-year input patterns used to predict over-predicted by amean of 23.2% (4 instances), the 10-year intervals by 19.5% (2 instances), and the 15-year intervalby 29.5% (1 instance). Fourth, drastic changes in land uses may result in large discrepanciesbetween predicted vs. actual area outcomes. While over-predictions of VL area for 2000 (7.5%)and 2005 (a mean of 9.1%) were relatively small, by 2010, the LTM over predicted by a mean of33.7%. This is probably because while Fort Worth lost 37.6% of its VL area from 1990 to 2010, it lost26.9% of this between 2005 and 2010. While VL area loss was gradual, unforeseen spikes in rapidgrowth may decrease the reliability of historical change maps for prediction purposes.

5.3. LTM output statistics

Overall results from the data produced from the LTM show a few varying and scattered outcomes(see Table 3). First, as more data were available as input factors for prediction, PCM and Kappastatistics did not seem to increase or decrease. Overall, PCM outputs ranged from 41.6% to 54.7%with a mean of 47.21% per training session. The output AUC values follow similar patterns of PCMand Kappa statistics with every model meeting the appropriate accuracy criteria and the 2000–2010 model showing the highest prediction accuracy. Logically, more input data would lead to ahigher accuracy when predicting, but this was not the case. In fact, the use of 15 input factorsactually showed a decrease in the PCM and Kappa statistics compared to the predictions using 14variables. Training with 16 drivers showed higher PCM and Kappa statistics than predictions madewith 15, but lower than when using 14 drivers. This may be due to the fact that rate of secondaryindustry, when removed, actually created an increase in PCM. It should be noted that the inter-related nature of the variables utilized to predict for VL may lead to the eventual presence of noisyinput factors in the input data set resulting in anomalous behavior. Many of the input factorsutilized are qualitative interrelated variables. For example, the 16 factor training utilized market andland value per parcel as economic drivers of vacancy rather than the value of the house residing inthe parcel utilized in the 14 and 15 factor training cycles, due to limited data availability. Theinterrelated nature of these two factors could also have diluted the variable pool, leading to a

Table 2. Variable influence outputs by dropping one variable per model using 2000–2010 input patterns.

Excluded input factors Highest training probability PCM Kappa coefficient Model influence

Rate of secondary industry* 10,000th 55.1 0.50 1% Below poverty rate 8000th 54.7 0.49 2Age of buildings 80,000th 54.5 0.49 3Rate of service industry 80,000th 54.4 0.49 4Parcel size 30,000th 54.3 0.49 5Railroad 25,000th 53.5 0.48 6Home ownership 60,000th 53.5 0.48 7Vacancy rate 70,000th 53.5 0.48 8Education 80,000th 53.5 0.48 9Highway 3500th 53.1 0.48 10Population change 50,000th 52.7 0.47 11Household income 70,000th 52.6 0.47 12Ethnicity 35,000th 52.6 0.47 13% Civilian unemployment 25,000th 52.2 0.47 14Market value 25,000th 47.8 0.42 15Land value 90,000th 47.7 0.42 16Full model 90,000th 54.7 0.50

JOURNAL OF LAND USE SCIENCE 15

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

heavier weighting unwittingly placed on parcel value which may or may not be the mostsignificant variable. These types of issues are typical in ANN-based models.

The number of cycles run showed no clear pattern for increasing or decreasing Kappa statisticsor PCM. This finding is consistent with other research. Probability peaks ranged from 7000 cycles to90,000 for each prediction scenario. There were also no clear parallels between the number ofinput drivers and the highest training probability. As the number of drivers increased, the numberof training cycles necessary for the highest probability of prediction accuracy did not necessarilydecrease. The amount of over or under estimation when comparing actual cell change to predictedcell change (prediction fit) did show a pattern with the number of cycles. Estimated change whenusing the LTM tended to consistently be higher than actual change. There was an overall averageof 1.45% higher predicted cell change than actual cell change. The LTM tended to over-predict cellchanges when running 45,000 or less cycles and under-predict as more cycles were run (higherthan 45,000). Training using 45,000 or less cycles (5 instances) showed around a 2.12% increase inpredicted change than actual change and training using higher cycles (2 instances) had an under-prediction average of −0.6%. Over predictions tended to be higher than under predictions.Simultaneously, under-predictions tended to have a narrower scope of error in reference to cellchange amounts than the over-predictions. Also, of the seven different models ran for the 5-yearinterval comparisons, the use of lower amounts of input factors tended to result in more over-predictions, with the exception of one instance, when using the highest number of drivers topredict vacancy change.

5.4. Input pattern differentiation with identical input factors

The LTM statistics did not afford the opportunity to gauge the prediction variability when usingidentical input factors but differing input patterns. Prediction variability, in this research, is used torefer to the LTM’s output differentiation when using the same area for an identical target date,despite the use of different input patterns. To assess prediction variability, the LTM was used toforecast VL for the year 2020 using differing combinations, or scenarios, of VL inventories, all withidentical input drivers. Change maps were produced in the LTM for each scenario and wereprocessed using 30 × 30 m (roughly 100 × 100 ft) cells and population changes between thetime periods. Projected population changes for 2020 were provided from the North Central TexasCouncil of Governments (2013). After each prediction map was made for 2020, all four scenarioswere amalgamated to create one composite score scenario map.

The research assumed current contributing factors of urban decline would remain consistent.Population changes between each time period were reflected in forecasting future VL pattern withthe change of vacancy cells. For example, the population change between 1990 and 2000 was276,116 and the number of vacant cells that transitioned at that time was 43,338. Since the

Table 3. LTM statistical output for all models and scenarios trained.

Inputpatterns

Pop.change

Yearpredicted

No. ofinputfactors

Highest trainingprobability (inthousands)

PCM(%)

Kappacoefficient AUC

Vacantcell

change

Predictedvacant cellchange

Predictionfit

1990–1995 105,322 2000 14 45th 49.3 0.47 0.713 16,531 43,338 +2.8%1990–1995 105,322 2005 14 45th 49.3 0.47 0.713 16,531 66,803 +1.2%1990–1995 105,322 2010 14 45th 49.3 0.47 0.713 16,531 100,285 +2.1%1990–2000 276,116 2005 15 70th 44.5 0.43 0.749 15,504 23,898 −0.4%1990–2000 276,116 2010 15 70th 44.5 0.43 0.749 15,504 35,876 −0.8%1990–2000 276,116 2020 16 70th 44.5 0.43 0.749 15,504 49,269 –1990–2005 425,612 2010 14 40th 41.6 0.34 0.744 53,488 80,296 +1.7%1990–2010 638,931 2020 16 7th 44.3 0.37 0.741 46,695 64,127 –2000–2005 149,496 2010 16 15th 48.9 0.43 0.745 59,777 145,074 +2.8%2000–2010 601,334 2020 16 90th 54.7 0.50 0.767 52,192 240,448 –

16 G. NEWMAN ET AL.

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

estimated population change between 1990 and 2020, forecasts to be 877,450 (indicating 3.18times more populace than between 1990 and 2000), cell change was also estimated to increase atthe same rate (3.18). In the same way, possible scenarios of VL in 2020 were predicted. Figure 5shows the predicted VL change in 2020 based on the three different time periods, while Table 3shows the summary of population changes per year based on actual and projected data as well asthe number of vacant cells (30 × 30 m) from 1990 to 2020.

As shown in Table 5, our approach used three different scenarios: Scenario A (input patternsfrom 1990 and 2000), Scenario B (input patterns from 1900 and 2010), and Scenario C (inputpatterns from 2000 and 2010) (see Table 6). Figure 5 displays the output map from each scenario,predicting VL by the year 2020. Scenario A shows most of the VL accumulating on the city’speriphery while Scenarios B and C show most of the VL being generated in the urban core area. Todetermine which of these prediction outputs was more reliable, a pattern density analysis of VLuses and a hot spot analysis was run using current data to determine where VL was statisticallyclustering.

Pattern density analysis shows where higher VL uses concentrate spatially. The output from thisanalysis shows that the highest concentrations resided near the city center, not the periphery (seeFigure 6). The hot-spot tool in GIS identifies statistically significant spatial clusters of high values(hot spots) and low values (cold spots) based on a z-score and p-value for each feature examinedwhich are used as measures of statistical significance indicating whether the observed spatialclustering is more pronounced than one would expect in a random distribution (Grubesic & Murray,2001). This type of spatial analysis has been used to study geographic changes such as soilpollution locations (Li, Lee, Wong, Shi, & Thornton, 2004), low income housing locations (Wang &Varady, 2005), biodiversity impacts (Cincotta, Wisnewski, & Engelman, 2000), crime activity (Anselin,Griffiths, & Tita, 2008), and broad land-use changes (Jusuf, Wong, Hagen, Anggoro, & Hong, 2007).Results show that Scenarios B and C were a more likely prediction than Scenario A consideringmost of the VL tended to be significantly clustering closer to the urban core area (see Figure 7).There was a slight downtown decline between 2000 and 2010 and many small vacant parcelssprouted up within the urban core in that period. Scenarios B and C primarily reflected this change,while Scenario A could not.

Despite this difference, all three scenarios had high enough PCM’s to merit acceptability ofprediction. The number of training cycles still varied in relationship to increase in PCM or Kappastatistic. Scenario C showed the highest PCM and Kappa, followed by Scenario B, then C.Discrepancies in the ratios of predicted vacant area to city size showed that Scenario A may be a

Table 5. Actual versus predicted ratio differences when using identical input factors but differing input patterns.

Input patterns Predicted ratio of vacant land to nonvacant land

2000 2005 2010

1990 + 1995 51.40% 52.10% 53.20%1990 + 2000 44.70% 45.00%1990 + 2005 41.80%2000 + 2005 43.90%Actual ratio of vacant land to nonvacant land 43.90% 39.30% 12.30%

Table 4. Quantity disagreement, allocation disagreement, and overall agreement outputs.

Parameter 1990–1995 1990–2000 1990–2005 1990–2010 2000–2005 2000–2010

PCM 49.3 44.5 41.6 44.3 48.9 54.7QD* 0.0 0.0 0.0 0.0 0.0 0.0AD** 3.9 4.0 14.8 12.3 4.0 9.6OA*** 96.1 96.0 85.4 87.7 96.0 90.4AUC**** 0.713 0.749 0.744 0.741 0.745 0.767

*QD: quantity disagreement; **AD: allocation disagreement; ***OA: overall agreement; ****AUC: area under ROC curve.

JOURNAL OF LAND USE SCIENCE 17

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

less reliable output considering current trends in VL in the city (see Table 6). As of 2010, VL usesoccupied around 12.3% of Fort Worth with a decrease in VL area of −9.4% per 5 years on average.Scenario A predicted around a 45% ratio of VL to city area by 2020 which is not consistent withcurrent patterns while Scenarios B and C predicted a more consistent 16–17%. The use of olderdata from Scenario A predicted a huge over prediction, while the use of more current dataprovided a closer estimation. The use of the current data, however, still predicted around a3–4% increase in VL above what currently exists.

Each 2020 scenario was given a score of ‘1’ (vacancy) or ‘0’ (nonvacant); areas where all threescenarios overlapped had a total score of 3, a score of 2 where two scenarios overlapped, and ascore of 1 where only one scenario predicated future vacancy for the parcel (see Figure 7). Spatiallocation alignment among predicated VL uses seemed to drastically change with differing inputpatterns. Outputs from all three scenarios only shared spatial overlap for 11.2% of the predictedvacant area (see Table 6). This accounts for 6.3% of the city’s total land area, meaning that 6.3% ofthe city’s land was predicted to become vacant by all three scenarios in identical locations,suggesting a high probability of future vacancy/abandonment. There was a 13.9% overlap forany two combinations of scenarios, suggesting relatively high probability of future vacancy whilenearly 3/4 of the parcels were only predicted by one of the given scenarios (74.9%), suggesting amoderate threat of future vacancy. Further, Scenarios B and C were the only situation whereoverlaps among predicted vacant parcels were actually larger than nonoverlapping parcels. Sharedprediction locations from the B and C models constituted 67.3% of the 13.9% total double scenariooverlap. These results reinforce the suggestion that the use or more current data with smaller timeframes may have less prediction variance and be more beneficial for more accurate output data.

6. Conclusions

This research sought to determine the influence of input factors on VL and test the accuracy of theLTM in predicting future urban vacancies. As noted, existing VL can lead to more VL throughincreased cost burdens on local governments, decreases in property values, and increases in crime.VL increases can actually impede future population growth. Repurposing these lots to reverseurban fortunes is no easy task. The maintenance or demolition of vacant properties can be a largeexpense. While industrial variables may play a larger role in legacy/shrinking cities, this studyshowed that market conditions seemed to be the most influential variable affecting predictionoutcomes when using the LTM in Fort Worth. Therefore, solutions to the vacancy issue shouldclosely align to neighborhood market conditions. Local housing market health and propertycondition can highly impact the ability to return vacant properties to productive urban land uses.

Figure 5. Possible scenarios of vacancy patterns by 2020 based on 1990–2000 (A), 1990–2010 (B), and 2000–2010 (C).(Data sources: City of Fort Worth, GIS Program; U.S. Census Bureau 1990, 2000, and 2010 Decennial Censuses).

18 G. NEWMAN ET AL.

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

In cities with strong markets, the emphasis should be on creating policy to either prevent ordisallow future vacancies. The stabilization of property values can do much to reverse earlysymptoms of urban decline. Thorough data and inventories on vacant lot conditions can beused to target risk areas and limit neighborhood distress. High unemployment and/or minorityareas, according to this research, may occupy a large portion of the early risk areas. The identifica-tion and targeting of these sites for new owner occupation or foreclosure prevention strategies canhelp reduce vacant housing lots. Stimulating the demand of new occupants in neighborhoods tomatch the supply through new commercial revitalization efforts or marketing must be considered.

Table 6. Ratio differences of 2020 vacant land prediction scenarios when using identical input factors but differing inputpatterns.

Scenarios Description of ratio analyzedRatio of predicted vacant area to

actual city area (%)Ratio of total predicted vacant area topredicted vacant area by scenario (%)

A* Predicted vacant area 44.6 100B** Predicted vacant area 15.6 100C*** Predicted vacant area 16.9 100A + B Either A or B scenario,

separate46.1 87.0

Overlap of A and B scenarios 6.9 13.0A + C Either A or C scenario,

separate47.6 87.5

Overlap of A and C scenarios 6.8 12.5B + C Either B or C scenario,

separate6.4 32.7

Overlap of B and C scenarios 19.4 67.3A + B + C Either A or B or C scenario,

separate42.2 74.9

Overlap of any twocombinations of scenarios

7.8 13.9

Overlap of all threecombinations of scenarios

6.3 11.2

*A: 2020 Prediction from 1990 to 2000.;**B: 2020 Prediction from 1990 to 2010;***C: 2020 Prediction from 2000 to 2010.

Figure 6. Land-use pattern map highlighting patterns of vacant parcel density (left) and hot spot analysis showing statisticallysignificant clusters of vacant land in Fort Worth, TX (right).(Data sources: City of Fort Worth, GIS Program; U.S. Census Bureau 1990, 2000, and 2010 Decennial Censuses).

JOURNAL OF LAND USE SCIENCE 19

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

The LTM has shown to be able to assist in both targeting sites of potential future vacancy andlocating risk areas prior to distress if a clear and thorough inventory of VL conditions is available.Overall, findings indicate that there was an adequate level of accuracy in all LTM outputs. All fourdiffering gauges of accuracy showed satisfactory results. Each model had acceptable Kappa scoresand PCM’s (40% or more) with fair to good AUC outputs (between 0.70 and 0.80). Also, the overallagreement was highly acceptable with a mean of 91.93 across all models. Lower PCM’s tended tohave lower overall agreements. The LTM did, however, tend to consistently over-predict, despitechanges in the number of input factors or input pattern intervals. This may be linked to the findingthat the estimated change when using the LTM tended to consistently be higher than actualchange. It could not be shown that these over-predictions were linked to an increase in amount ofinput drivers. In fact, the use of lower amounts of input factors tended to result in higher over-predictions. As more input factors were utilized, there was no clear pattern of PCM and Kappastatistics increase. This is somewhat inconsistent with previous findings that the predictive ability ofthe LTM improves when using a multitude of predictor variables.

Prior research indicates that LUCC models that are applied to landscapes that have largeramounts of net change tend to have higher rates of predictive accuracy (Pontius et al., 2008). VLpredictions using the LTM, however, may actually prove to be more accurate when a gradualchange is occurring. The rapid changes in VL uses resulted in larger errors in both predictionlocation and amounts. Spatial location agreement of predicted VL showed to be heavily dependentupon input patterns and altered when using differing amounts of input patterns. Findings suggestthat the use or more recent input patterns with smaller time frames may have more accurateoutput. The number of training cycles varied per output and there was no clear pattern that highernumbers of cycles resulted in more accurate predictions. A pattern was found that the LTM tendedto over-predict cell change amounts when running 45,000 or less cycles. Findings were alsoconsistent with previous findings that usually, over 4000 cycles of training are necessary

Figure 7. Overlap of all three scenarios for 2020 vacant land predictions in Fort Worth, TX, USA.(Data sources: City of Fort Worth, GIS Program; U.S. Census Bureau 1990, 2000, and 2010 Decennial Censuses).

20 G. NEWMAN ET AL.

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

(Pijanowski et al., 2002) as all models in this research has the highest PCM and Kappa coefficients atthe 7000th cycle or higher.

In the future, it may also be useful to include a spatial autocorrelation metric within the LTM as,logically, factors leading to vacancy in one area will be similar in the immediately surroundingareas. It has been shown that spatial autocorrelation in land-use patterns are scale dependent(Verburg et al., 2004). For example, at a large scale, residential areas can be clustered with positivespatial autocorrelations; however, at the parcel scale, there may be negative spatial interactionamong developed parcels (Irwin & Geoghegan, 2001). In some cases, a developed parcel canactually repel neighboring development due to negative spatial externalities that may be gener-ated from existing development (e.g. congestion). Therefore, a parcel’s probability of developmentcan sometimes actually decrease as the amount of existing neighboring development increases.There are differing causal processes at each scale. So, while spatial interactions should be studied atmultiple scales, relations found at each particular scale should only be utilized at that scale.

The LTM has shown to be a potentially good resource for researchers and spatial analysis expertsbut may require considerable improvements before it can be widely adopted and used for long-rangeurban planning and policy. First, LTM modeling is extremely time consuming. Getting to a finer grainof detail such as using a higher spatial resolution will provide better accuracy. This, however,significantly increases the duration time of running the model. Second, there is also a large learningcurve for those in the planning field who are not familiar with complicated GIS-based tools andextensions. Third, calibration of the LTM can sometimes be impeded by existing classification andaggregation errors in land-use maps. While this research only examined one city, models performedon multiple cities can be plagued by these inconsistencies. In regards to VL studies, the definition ofVL also needs to remain consistent as typologies and classifications differ by municipality. Finally,future LTM research must direct data providers to improve and continually monitor data, therebyreducing the uncertainties of model outputs. VL uses are a continually changing entity and, as such,merit a continual acutely detailed inventory.

Disclosure statement

No potential conflict of interest was reported by the authors.

ORCIDGalen Newman http://orcid.org/0000-0003-4277-5931

References

Alig, R. J. (1986). Econometric analysis of the factors influencing forest acreage trends in the southeast. Forest Science,32(1), 119–134.

Almeida, C. M., Gleriani, J. M., Castejon, E. F., & Soares-Filho, B. S. (2008). Using neural networks and cellular automatafor modelling intra-urban land-use dynamics. International Journal of Geographical Information Science, 22(9), 943–963. doi:10.1080/13658810701731168

Anas, A. (1978). Dynamics of urban residential growth. Journal of Urban Economics, 5(1), 66–87. doi:10.1016/0094-1190(78)90037-2

Anselin, L., Griffiths, E., & Tita, G. (2008). 6. Crime mapping and hot spot analysis. In R. Wortley & L. Mazerolle (Eds.),Environmental criminology and crime analysis (pp. 97–116). Cullompton: Willan Publishing.

Aryeetey-Attoh, S., Costa, F., Morrow-Jones, H., Monroe, C., & Sommers, G. (1998). Central-city distress in Ohio’s elasticcities: Regional and local policy responses. Urban Geography, 19(8), 735–756. doi:10.2747/0272-3638.19.8.735

Audirac, I. (2007). Urban shrinkage amid fast metropolitan growth (two faces of contemporary urbanism). Online [cit.25. 9. 2009]. Retrieved from http://www.coss.fsu.edu/durp/sites/coss.fsu.edu.durp/files/Audirac2009.pdf

Batty, M. (1992). Urban modeling in computer-graphic and geographic information system environments. Environmentand Planning B: Planning and Design, 19, 663–688. doi:10.1068/b190663

JOURNAL OF LAND USE SCIENCE 21

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

Batty, M. (2009). Urban modeling. In D. Goodman & M. Goodman (Eds.), International encyclopedia of humangeography. Oxford: Elsevier.

Batty, M., Xie, Y., & Sun, Z. (1999). Modeling urban dynamics through GIS-based cellular automata. Computers,Environment and Urban Systems, 23(3), 205–233. doi:10.1016/S0198-9715(99)00015-0

Berke, P. R., Godschalk, D. R., & Kaiser, E. J. (2006). Urban land use planning. Chicago: University of Illinois Press.Bishop, C. M. (1995). Neural networks for pattern recognition. Oxford: Oxford University Press.Bontje, M. (2005). Facing the challenge of shrinking cities in east Germany: The case of Leipzig. GeoJournal, 61(1), 13–

21. doi:10.1007/s10708-005-0843-2Bourne, L. (1996). Reurbanization, uneven urban development, and the debate on new urban forms. Urban Geography,

17(8), 690–713. doi:10.2747/0272-3638.17.8.690Bowman, A., & Pagano, M. (2004). Terra incognita: Vacant land and urban strategies. Washington, DC: Georgetown

University Press.Bradford, C. (1979). Financing home ownership: The federal role in neighborhood decline. Urban Affairs Review, 14(3),

313–335. doi:10.1177/107808747901400303Briassoulis, H. (2004). The institutional complexity of environmental policy and planning problems: The example of

mediterranean desertification. Journal of Environmental Planning and Management, 47(1), 115–135. doi:10.1080/0964056042000189835

Brophy, P. C., & Vey, J. S. (2002). Seizing city assets: Ten steps to urban land reform. Washington, DC: BrookingsInstitution Center on Urban and Metropolitan Policy.

Brown, D. G., Pijanowski, B. C., & Duh, J. D. (2000). Modeling the relationships between land use and land cover onprivate lands in the Upper Midwest, USA. Journal of Environmental Management, 59(4), 247–263. doi:10.1006/jema.2000.0369

Buhnik, S. (2010). From shrinking cities to Toshi no Shukushō: Identifying patterns of urban shrinkage in the Osakametropolitan area. Berkeley Planning Journal, 23(1), 132–155.

Capozza, D., & Helsley, R. (1989). The fundamentals of land prices and urban growth. Journal of Urban Economics, 26(3),295–306. doi:10.1016/0094-1190(89)90003-X

Carrion-Flores, C., & Irwin, E. (2004). Determinants of residential land-use conversion and sprawl at the rural-urbanfringe. American Journal of Agricultural Economics, 86(4), 889–904. doi:10.1111/j.0002-9092.2004.00641.x

Chapin, F. S., & Weiss, S. F. (1968). A probabilistic model for residential growth. Transportation Research, 2(4), 375–390.doi:10.1016/0041-1647(68)90103-2

Cincotta, R. P., Wisnewski, J., & Engelman, R. (2000). Human population in the biodiversity hotspots. Nature, 404(6781),990–992. doi:10.1038/35010105

City of Fort Worth. (2014). 2014 comprehensive plan. Retrieved from Fort Worth http://fortworthtexas.gov/uploadedFiles/Planning_and_Development/Planning_and_Design/Comprehensive_Plan/Glossary.pdf

Clarke, K., Hoppen, S., & Gaydos, L. (1997). A self-modifying cellular automaton model of historical urbanization in theSan Francisco bay area. Environment and Planning B: Planning and Design, 24, 247–261. doi:10.1068/b240247

Colwell, P., & Munneke, H. (1997). The structure of urban land prices. Journal of Urban Economics, 41(3), 321–336.doi:10.1006/juec.1996.2000

Congalton, R., Oderwald, R., & Mead, R. (1983). Assessing landsat classification accuracy using discrete multivariateanalysis statistical techniques. Photogrammetric Engineering and Remote Sensing, 49, 1671–1678.

Conway, T. M. (2009). The impact of class resolution in land use change models. Computers, Environment and UrbanSystems, 33(4), 269–277. doi:10.1016/j.compenvurbsys.2009.02.001

Couch, C., Karecha, J., Nuissl, H., & Rink, D. (2005). Decline and sprawl: An evolving type of urban development–observed in Liverpool and Leipzig. European Planning Studies, 13(1), 117–136. doi:10.1080/0965431042000312433

Cunningham-Sabot, E., & Fol, S. (2009). Shrinking cities in France and Great Britain: A silent process. In K. Pallagst (Ed.),Future of shrinking cities-problems, patterns and strategies of urban transformation in a global context. Berkeley, CA:Institute of Urban and Regional Development.

Dong, H. (2013). Concentration or dispersion? Location choice of commercial developers in the Portland metropolitanarea, 2000–2007. Urban Geography, 34(7), 989–1010. doi:10.1080/02723638.2013.778587

Farris, J. T. (2001). The barriers to using urban infill development to achieve smart growth. Housing Policy Debate, 12(1),1–30.

Fee, K., & Hartley, D. (2011). Economic trends: Growing cities, shrinking cities. Federal Reserve Bank of Cleveland. April14, Cleveland, OH. Retrieved from www.clevelandfed.org/resesarch/trends/2011/0411/01labmar.cfm

Fischer, M. M., & Abrahart, R. J. (2000). Neurocomputing-tools for geographers. In S. Openshaw & R. J. Abrahart (Eds.),GeoComputation (pp. 187–217). New York, NY: Taylor & Francis.

Foley, J. A., DeFries, R., Asner, G. P., Barford, C., Bonan, G., Carpenter, S. R., & Stuart Chapin, F., et al. (2005). Globalconsequences of land use. Science, 309(5734), 570–574. doi:10.1126/science.1111772

Foody, G. M. (2002). Status of land cover classification accuracy assessment. Remote Sensing of Environment, 80(1),185–201. doi:10.1016/S0034-4257(01)00295-4

Fulton, G. (2006). Recycling the city: The use and reuse of urban land, edited by Rosalind Greenstein and Yesim Sungu-Eryilmaz and Drosscape: Wasting land in urban America, by Alan Berger. Landscape Architecture, 96(7), 120–121.

22 G. NEWMAN ET AL.

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

Geoghegan, J., Wainger, L. A., & Bockstael, N. E. (1997). Spatial landscape indices in a hedonic framework: Anecological economics analysis using GIS. Ecological Economics, 23(3), 251–264. doi:10.1016/S0921-8009(97)00583-1

Geographical Sciences Committee. (2014). Advancing land change modeling: Opportunities and research requirements.Washington, DC: National Academies Press.

Glaeser, E. (2013). A Nation of Gamblers: Real Estate Speculation and American History. No. w18825. National Bureau ofEconomic Research, 2013.

Glaeser, E., & Gyourko, J. (2001). Urban decline and durable housing. National Bureau of Economic Research, No. w8598.Glaeser, E., Gyourko, J., & Saks, R. (2006). Urban growth and housing supply. Journal of Economic Geography, 6(1), 71–

89. doi:10.1093/jeg/lbi003Glaeser, E., & Kahn, M. (2004). Sprawl and urban growth. Handbook of Regional and Urban Economics, 4, 2481–2527.Goldstein, J., Jensen, M., & Reiskin, E. (2001). Urban vacant land redevelopment challenges and progress. Cambridge, MA:

Lincoln Institute of Land Policy.Grubesic, T. H., & Murray, A. T. (2001). Detecting hot spots using cluster analysis and GIS. Annual conference of CMRC, Dallas.Guan, Q., Wang, L., & Clarke, K. C. (2005). An artificial-neural-network-based, constrained CA model for simulating

urban growth. Cartography and Geographic Information Science, 32(4), 369–380. doi:10.1559/152304005775194746Gutman, G., Janetos, A. C., Justice, C. O., Moran, E. F., Mustard, J. F., Rindfuss, R. R.,& Cochrane, M. A. (Eds.). (2004). Land

change science: Observing, monitoring and understanding trajectories of change on the earth's surface (Vol. 6).Dordrecht: Kluwer Academic. Springer Science & Business Media.

Heckert, M., & Mennis, J. (2012). The economic impact of greening urban vacant land: A spatial difference-in-differences analysis. Environment and Planning A, 44(12), 3010–3027. doi:10.1068/a4595

Henry, M. S., Schmitt, B., & Piguet, V. (2001). Spatial econometric models for simultaneous systems: Application to ruralcommunity growth in France. International Regional Science Review, 24(2), 171–193. doi:10.1177/016001701761013169

Herold, M., Couclelis, H., & Clarke, K. C. (2005). The role of spatial metrics in the analysis and modeling of urban landuse change. Computers, Environment and Urban Systems, 29(4), 369–399. doi:10.1016/j.compenvurbsys.2003.12.001

Hollander, J. B. (2010). Moving toward a shrinking cities metric: Analyzing land use changes associated withdepopulation in Flint, Michigan. Cityscape, 12(1), 133–151.

Hollander, J. B., & Németh, J. (2011). The bounds of smart decline: A foundational theory for planning shrinking cities.Housing Policy Debate, 21(3), 349–367. doi:10.1080/10511482.2011.585164

Irwin, E., & Geoghegan, J. (2001). Theory, data, methods: Developing spatially explicit economic models of land usechange. Agriculture, Ecosystems & Environment, 85(1–3), 7–24. doi:10.1016/S0167-8809(01)00200-6

Johnson, M. P., Hollander, J., & Hallulli, A. (2014). Maintain, demolish, re-purpose: Policy design for vacant landmanagement using decision models. Cities, 40, 151–162. doi:10.1016/j.cities.2013.05.005

Jusuf, S. K., Wong, N. H., Hagen, E., Anggoro, R., & Hong, Y. (2007). The influence of land use on the urban heat islandin Singapore. Habitat International, 31(2), 232–242. doi:10.1016/j.habitatint.2007.02.006

Kaiser, E. J., & Chapin, F. S. (1995). Urban land use planning. Urbana: University of Illinois Press.Keenan, P., Lowe, S., & Spencer, S. (1999). Housing abandonment in inner cities-the politics of low demand for

housing. Housing Studies, 14(5), 703–716.Kivell, P. (1993). Land and the city: Patterns and processes of urban change. London: Routledge.Kourtit, K., Nijkamp, P., & Reid, N. (2014). The new urban world: Challenges and policy. Applied Geography, 49, 1–3.

doi:10.1016/j.apgeog.2014.01.007Kourtit, K., Nijkamp, P., & Scholten, H. (2014). The future of the new urban world. International Planning Studies, 20

(1–2), 4–20.Kozloff, H. (2007). Drosscape: Wasting land in urban America, by Alan Berger. Urban Land, 66(5), 158.Landis, J. D. (1994). The California urban futures model: A new-generation of metropolitan simulation-models.

Environment and Planning B: Planning and Design, 21, 399–420. doi:10.1068/b210399Landis, J. D. (2011). Urban growth models: State of the art and prospects. In E. L. Birch & S. M. Wachter (Eds.), Global

urbanization (pp. 126–140). Philadelphia: University of Pennsylvania Press.Lee, M.-J., Hanna, A., & Loh, W.-Y. (2004). Decision tree approach to classify and quantify cumulative impact of change

orders on productivity. Journal of Computing in Civil Engineering, 18(2), 132–144. doi:10.1061/(ASCE)0887-3801(2004)18:2(132)

Lester, T., Kaza, N., & Kirk, S. (2013). Making room for manufacturing: Understanding industrial land conversion incities. Journal of the American Planning Association, 79(4), 295–313. doi:10.1080/01944363.2014.915369

Li, X., & Yeh, A. G. O. (2002). Neural-network-based cellular automata for simulating multiple land use changes usingGIS. International Journal of Geographical Information Science, 16(4), 323–343. doi:10.1080/13658810210137004

Li, X., Lee, S.-L., Wong, S.-C., Shi, W., & Thornton, I. (2004). The study of metal contamination in urban soils of HongKong using a GIS-based approach. Environmental Pollution, 129(1), 113–124. doi:10.1016/j.envpol.2003.09.030

Lindsey, C. (2007). Smart decline. Panorama: What’s New in Planning, 15(1), 17–21.Longoria, R., & Rogers, S. (2013). Exodus within an expanding city: The case of Houston’s historic African-American

communities. Urban Design International, 18(1), 24–42. doi:10.1057/udi.2012.28Mallach, A. (Ed.) (2012a). Depopulation, market collapse and property abandonment: Surplus land and buildings in

legacy cities. In Rebuilding America’s legacy cities: New directions for the industrial heartland (pp. 85–110).

JOURNAL OF LAND USE SCIENCE 23

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

Mallach, A. (2012b). Laying the groundwork for change: Demolition, urban strategy, and policy reform. Washington, DC:Brookings Institution.

Mallach, A. (2012c). Rebuilding America’s legacy cities: New directions for the industrial heartland. New York, NY:American Assembly, Columbia University.

Mallach, A., & Brachman, L. (2013). Regenerating America’s legacy cities. Cambridge, MA: Lincoln Institute of LandPolicy.

Manel, S. H., Williams, C., & Ormerod, S. J. (2001). Evaluating presence–absence models in ecology: The need toaccount for prevalence. Journal of Applied Ecology, 38(5), 921–931. doi:10.1046/j.1365-2664.2001.00647.x

Massey, D. S., & Denton, N. A. (1993). American apartheid: Segregation and the making of the underclass. Cambridge,MA: Harvard University Press.

Moore, T. (2000). Geospatial expert systems. In S. Openshaw & R. J. Abrahart (editors), Geocomputation (pp. 127–159).London: Taylor & Francis.

Németh, J., & Langhorst, J. (2014). Rethinking urban transformation: Temporary uses for vacant land. Cities, 40, 143–150. doi:10.1016/j.cities.2013.04.007

Newman, G. (2015). The Eidos of urban form: A framework for heritage-based place making. Journal of Urbanism:International Research on Placemaking and Urban Sustainability, 1–20 (ahead of print). doi:10.1080/17549175.2015.1070367

Newman, G. (2013). A conceptual model for measuring neglect in historic districts. Journal of Preservation, Education,and Research, 6, 41–58.

Newman, G., & Saginor, J. (2014). Four imperatives for preventing demolition by neglect. Journal of Urban Design, 19(5), 622–637. doi:10.1080/13574809.2014.943705

North Central Texas Council of Governments. (2013). North Texas to 2030: Extending the trends. Vision North Texas:Understanding our options for growth report. Retrieved from http://www.visionnorthtexas.org/regionalchoices/RegChoices_NorthTexas2030.pdf

Olson, J. M., Alagarswamy, G., Andresen, J. A., Campbell, D. J., Davis, A. Y., Ge, J., & Huebner, M. (2008). Integratingdiverse methods to understand climate–land interactions in east Africa. Geoforum, 39(2), 898–911. doi:10.1016/j.geoforum.2007.03.011

Openshaw, S. (1998). Neural network, genetic, and fuzzy logic models of spatial interaction. Environment and PlanningA, 30(10), 1857–1872. doi:10.1068/a301857

Osborne, P. E., Alonso, J. C., & Bryant, R. G. (2001). Modelling landscape-scale habitat use using GIS and remotesensing: A case study with great bustards. Journal of Applied Ecology, 38(2), 458–471. doi:10.1046/j.1365-2664.2001.00604.x

Pagano, C. (2014). DIY urbanism: Property and process in grassroots city building. Marquette Law Review, 97(2), 335–390.

Park, I. K., & Ciorici, P. (2013). Determinants of vacant lot conversion into community gardens: Evidence fromPhiladelphia. International Journal of Urban Sciences, 17(3), 385–398. doi:10.1080/12265934.2013.818388

Park, I. K., & Rabenau, B. V. (2015). Tax delinquency and abandonment: An expanded model with application toindustrial and commercial properties. Urban Studies, 52(5), 857–875. doi:10.1177/0042098014524610

Pijanowski, B., Shellito, B., Bauer, M., & Sawaya, K. (2001, April 23–27). Using GIS, artificial neural networks and remotesensing to model urban change in the Minneapolis–St. Paul and Detroit Metropolitan areas. Proceedings of theAmerican Society of Photogrammetry and Remote Sensing Annual Conference, St. Louis, MO.

Pijanowski, B. C., Alexandridis, K. T., & Müller, D. (2006). Modelling urbanization patterns in two diverse regions of theworld. Journal of Land Use Science, 1(2–4), 83–108. doi:10.1080/17474230601058310

Pijanowski, B. C., Brown, D. G., Shellito, B. A., & Manik, G. A. (2002). Using neural networks and GIS to forecast land usechanges: A land transformation model. Computers, Environment and Urban Systems, 26(6), 553–575. doi:10.1016/S0198-9715(01)00015-1

Pinjanowski, B. C., Long, D. T., Gage, S. H., & Cooper, W. E. (1997, June 5–6). A land transformation model: Conceptualelements, spatial objects class hierarchies, GIS command syntax and an application for Michigan's Sagninaw Baywatershed. Land Use Modeling Workshop, USGS EROS Data Center, Sioux Falls, SD.

Pijanowski, B. C., Machemer, T., Gage, S., Long, D., Cooper, W., & Edens, T. (1995). A land transformation model:Integration of policy, socioeconomics and ecological succession to examine pollution patterns in watershed. Reportto the Environmental Protection Agency (pp. 72–83). North Carolina: Research Triangle Park.

Pijanowski, B. C., Pithadia, S., Shellito, B. A., & Alexandridis, K. (2005). Calibrating a neural network-based urban changemodel for two metropolitan areas of the upper midwest of the United States. International Journal of GeographicalInformation Science, 19(2), 197–215. doi:10.1080/13658810410001713416

Pijanowski, B. C., Tayyebi, A., Delavar, M. R., & Yazdanpanah, M. J. (2009). Urban expansion simulation using geographicinformation systems and artificial neural networks. International Journal of Environmental Research, 3, 493–502.

Pijanowski, B. C., Tayyebi, A., Doucette, J., Pekin, B. K., Braun, D., & Plourde, J. (2014). A big data urban growth simulation at anational scale: Configuring the GIS and neural network based land transformation model to run in a high performancecomputing (HPC) environment. Environmental Modelling & Software, 51, 250–268. doi:10.1016/j.envsoft.2013.09.015

24 G. NEWMAN ET AL.

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

Pond, B., & Yeates, M. (1993). Rural/urban land conversion I: Estimating the direct and indirect impacts. UrbanGeography, 14(4), 323–347. doi:10.2747/0272-3638.14.4.323

Pontius, R., Jr (2000). Quantification error versus location error in comparison of categorical maps. PhotogrammetricEngineering and Remote Sensing, 66, 1011–1016.

Pontius, Jr, R., & Millones, M. (2008, June 1–6). Problems and solutions for kappa-based indices of agreement. Studying,modeling and sense making of planet earth, Mytilene.

Pontius, R., Jr, & Millones, M. (2011). Death to kappa: Birth of quantity disagreement and allocation disagreement foraccuracy assessment. International Journal of Remote Sensing, 32(15), 4407–4429. doi:10.1080/01431161.2011.552923

Pontius, R. G., Jr, Boersma, W., Castella, J.-C., Clarke, K., De Nijs, T., Dietzel, C. . . . Verburg, P. H. (2008). Comparing theinput, output, and validation maps for several models of land change. The Annals of Regional Science, 42(1), 11–37.doi:10.1007/s00168-007-0138-2

Rappaport, J. (2003). US urban decline and growth: 1950 to 2000. Economic Review-Federal Reserve Bank of Kansas City,88(3), 15–44.

Ray, D. K., & Pijanowski, B. C. (2010). A backcast land use change model to generate past land use maps: Applicationand validation at the Muskegon river watershed of Michigan, USA. Journal of Land Use Science, 5(1), 1–29.doi:10.1080/17474230903150799

Reed, R. D., & Marks, R. J. (1998). Neural smithing: Supervised learning in feedforward artificial neural networks.Cambridge: MIT Press.

Rieniets, T. (2009). Shrinking cities: Causes and effects of urban population losses in the twentieth century. Nature andCulture, 4(3), 231–254. doi:10.3167/nc.2009.040302

Rindfuss, R. R., Walsh, S. J., Turner, B. L., Fox, S., & Mishra, V. (2004). Developing a science of land change: Challengesand methodological issues. Proceedings of the National Academy of Sciences of the United States of America, 101(39),13976–13981. doi:10.1073/pnas.0401545101

Rutherford, G. N., Bebi, P., Edwards, P. J., & Zimmermann, N. E. (2008). Assessing land-use statistics to model land coverchange in a mountainous landscape in the European Alps. Ecological Modelling, 212(3–4), 460–471. doi:10.1016/j.ecolmodel.2007.10.050

Ryan, B. D. (2012). Design after decline: How America rebuilds shrinking cities. Philadelphia: University of Pennsylvania Press.Rybczynski, W., & Linneman, P. D. (1999, Spring). How to save our shrinking cities. The Public Interest, 135, 30–44.Schilling, J., & Logan, J. (2008). Greening the rust belt: A green infrastructure model for right sizing America’s shrinking

cities. Journal of the American Planning Association, 74(4), 451–466. doi:10.1080/01944360802354956Setterfield, M. (1997). Abandoned buildings: Models for legislative & enforcement reform. Trinity Center for

Neighborhoods, Research Project 23, March.Shetty, S., & Reid, N. (2014). Dealing with decline in old industrial cities in Europe and the United States: Problems and

policies. Built Environment, 40(4), 458–474.Sousa, J. M., & Kaymak, U. (2002). Fuzzy decision making in modeling and control. Singapore: World Scientific and

Imperial College.Tang, Z., Engel, B. A., Pijanowski, B. C., & Lim, K. (2005). Forecasting land use change and its environmental impact at a

watershed scale. Journal of Environmental Management, 76(1), 35–45. doi:10.1016/j.jenvman.2005.01.006Tayyebi, A., Pekin, B. K., Pijanowski, B. C., Plourde, J. D., Doucette, J. S., & Braun, D. (2013). Hierarchical modeling of

urban growth across the conterminous USA: Developing meso-scale quantity drivers for the land transformationmodel. Journal of Land Use Science, 8(4), 422–442. doi:10.1080/1747423X.2012.675364

Tayyebi, A., Perry, P. C., & Tayyebi, A. H. (2014). Predicting the expansion of an urban boundary using spatial logisticregression and hybrid raster–vector routines with remote sensing and GIS. International Journal of GeographicalInformation Science, 28(4), 639–659. doi:10.1080/13658816.2013.845892

Temkin, K., & Rohe, W. (1996). Neighborhood change and urban policy. Journal of Planning Education and Research, 15(3), 159–170.

Theobald, D. M., & Hobbs, N. T. (1998). Forecasting rural land-use change: A comparison of regression-and spatialtransition-based models. Geographical and Environmental Modelling, 2, 65–82.

Torrens, P. M. (2003). Cellular automata and multi-agent systems as planning support tools. In S. Geertman & J.Stillwell (Eds.), Planning support systems in practice (pp. 205–222). Aarhus: Springer Berlin Heidelberg.

Turner, B. L., & Meyer, W. B. (1991). Land-use and land cover in global environmental-change-considerations for study.International Social Science Journal, 43(4), 669–679.

United Nations. (2011). World urbanization prospects: The 2011 revision highlights. New York, NY: United NationsDepartment of Economic and Social Affairs Population Division. Retrieved from http://esa.un.org/unup/pdf/WUP2011_Highlights.pdf

United States Census Bureau. (2015). American fact finder – Decennial census. United States Census Bureau. RetrievedJanuary 6, 2015, from http://factfinder2.census.gov

Vafeidis, A. T., Koukoulas, S., Gatsis, I., & Gkoltsiou, K. (2007, July 23–28). Forecasting land-use changes with the use ofneural networks and GIS (pp. 5068–5071). Geoscience and Remote Sensing Symposium, IGARSS 2007, IEEEInternational, Barcelona. IEEE.

JOURNAL OF LAND USE SCIENCE 25

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6

Van den Berg, L., Drewett, R., Klaasen, L., Rossi, A., & Vijverberg, C. (1982). Urban Europe: A study of growth and decline.Oxford: Pergamon.

Verburg, P. H. (2006). Simulating feedbacks in land use and land cover change models. Landscape Ecology, 21(8),1171–1183. doi:10.1007/s10980-006-0029-4

Verburg, P. H., Schot, P. P., Dijst, M. J., & Veldkamp, A. (2004). Land use change modelling: Current practice andresearch priorities. GeoJournal, 61(4), 309–324. doi:10.1007/s10708-004-4946-y

Wachter, S. M., & Zeuli, K. A. (Eds.). (2013). Revitalizing American cities. Philadelphia: University of Pennsylvania Press.Waddell, P. (2002). UrbanSim: Modeling urban development for land use, transportation, and environmental planning.

Journal of the American Planning Association, 68(3), 297–314. doi:10.1080/01944360208976274Wang, X., & Varady, D. P. (2005). Using hot-spot analysis to study the clustering of section 8 housing voucher families.

Housing Studies, 20(1), 29–48. doi:10.1080/0267303042000308714Washington-Ottombre, C., Pijanowski, B. C., Campbell, D. J., Olson, J. M., Maitima, J. M., Musili, A. . . . Mwangi, A. (2010).

Using a role-playing game to inform the development of land-use models for the study of a complex socio-ecological system. Agricultural Systems, 103(3), 117–126. doi:10.1016/j.agsy.2009.10.002

Wegener, M. (1982). Modeling urban decline: A multilevel economic-demographic model for the Dortmund region.International Regional Science Review, 7(2), 217–241. doi:10.1177/016001768200700207

World Bank/IMF. (2013). Global monitoring report 2013: Rural urban dynamics and millennium development goals.Washington, DC: World Bank.

Yeh, A. G. O., & Li, X. (2003). Simulation of development alternatives using neural networks, cellular automata, and GISfor urban planning. Photogrammetric Engineering & Remote Sensing, 69(9), 1043–1052. doi:10.14358/PERS.69.9.1043

26 G. NEWMAN ET AL.

Dow

nloa

ded

by [

Tex

as A

&M

Uni

vers

ity L

ibra

ries

] at

11:

20 2

8 M

arch

201

6


Recommended