Land Cover Along an Urban-Rural Gradient: Implications for Water Quality

August 1998 619FRESHWATER SYSTEMS

619

Ecological Applications, 8(3), 1998, pp. 619–630q 1998 by the Ecological Society of America

LAND COVER ALONG AN URBAN–RURAL GRADIENT: IMPLICATIONSFOR WATER QUALITY

DAVID N. WEAR,1 MONICA G. TURNER,2 AND ROBERT J. NAIMAN3

1USDA Forest Service, Southern Research Station, P.O. Box 12254, Research Triangle Park, North Carolina 27709 USA2Department of Zoology, Birge Hall, 430 Lincoln Drive, University of Wisconsin, Madison, Wisconsin 53706 USA

3School of Fisheries, Box 357980, University of Washington, Seattle, Washington 98195 USA

Abstract. Development pressures in rural mountainous areas of the United States holdcrucial implications for water quality. Especially important are changes in the extent andpattern of various land uses. We examine how position along an urban–rural gradient affectslandscape patterns in a southern Appalachian watershed, first by testing for the effect ofdistance from an urban center on land-cover change probabilities and then simulating theimplied development of a landscape at regular distance intervals. By simulating a commonhypothetical landscape we control for variable landscape conditions and define how landdevelopment might proceed in the future. Results indicate that position along the urban–rural gradient has a significant effect on land-cover changes on private lands but not onpublic lands. Furthermore, position along the gradient has a compounding effect on land-cover changes through interactions with other variables such as slope. Simulation resultsindicate that these differences in land-cover changes would give rise to unique ‘‘landscapesignatures’’ along the urban–rural gradient. By examining a development sequence, weidentify patterns of change that may be most significant for water quality. Two locationsalong the urban–rural gradient may hold disproportionate influence over water quality inthe future: (1) at the most remote portion of the landscape and (2) at the outer envelopeof urban expansion. These findings demonstrate how landscape simulation approaches canbe used to identify where and how land use decisions may have critical influence overenvironmental quality, thereby focusing both future research and monitoring efforts andwatershed protection measures.

Key words: land-cover dynamics; landscape ecology; simulation modeling; southern Appalachians;water quality.

INTRODUCTION

Water connects and focuses activities throughout awatershed, defining the cumulative impacts of humanuse of land and resources. The use and condition ofland in particular have a profound influence on waterquality (Allan and Flecker 1993, Hunsaker and Levine1995). Integrating social and environmental vitalitywithin properly scaled ecosystems requires a differentfocus for science, involving both defining mechanismsthat organize human behavior and then linking humanendeavors and environmental quality. This linkage be-tween people and the environment is bidirectional, in-volving both the effects of people on ecosystems andthe effects of ecosystem structure and environmentalquality on people’s well-being. Because water qualityis so clearly impacted by people’s use of resources, andbecause human well-being is so clearly linked to theavailability of clean water, water provides perhaps the

Manuscript received 6 February 1997; revised 4 December1997; accepted 22 December 1997; final version received 7January 1988. For reprints of this Invited Feature, see foot-note 1, page 569.

strongest demand for social–ecological interdisciplin-ary research (Naiman et al. 1995).

This article examines factors that organize humanbehavior in a developing landscape and examines thepotential consequences for water quality. The over-arching theme is that by understanding patterns of de-velopment and linking them to their implications forwater quality, planning and regulatory efforts then canbe focused on areas critical in determining environ-mental quality. For example, knowledge of this sortcan be used to identify where land use is most likelyto be intensified within a watershed and, conversely,where it will likely remain stable or change in onlytrivial ways. In a world of limited resources, this kindof targeting may be much more effective than broadregulations intended to protect water quality (Wear etal. 1996).

Our specific focus is on understanding patterns ofland use change along the urban–rural gradient in theSouthern Appalachian Highlands. Portions of this re-gion have experienced nearly exponential populationgrowth with associated development of land since the1950s. Growth centers have included Hendersonville

620 INVITED FEATURE Ecological ApplicationsVol. 8, No. 3

and Boone, North Carolina, USA. In areas where pop-ulation growth and development are in initial phases,it may be possible to anticipate and thereby redirectsome development that would prove detrimental to en-vironmental and aesthetic qualities. We test various hy-potheses regarding the way that development, mea-sured as extent and patterns of land-cover change, isorganized along the urban–rural gradient and then ex-amine how these changes could be related to planningfor the protection of water quality in a developing land-scape. In particular, we use statistical models of land-cover change to test whether: (1) land-cover changesare generally determined by topographic and locationalfeatures, (2) land-cover regimes are influenced by po-sition along the urban–rural gradient, and (3) land-cov-er changes differ between public and private lands atvarious positions along the urban–rural gradient. Weuse a simulation experiment to test for: (1) differencesin the implied equilibrium land cover at different po-sitions along the urban–rural gradient and (2) Impliedchanges in equilibrium land cover as development pro-ceeds. Findings from these tests are then used to ex-plore the implications that continued developmentmight have for water quality in the southern Appala-chians.

STUDY AREA AND DATA

Our study site was the Little Tennessee River basinin southwestern North Carolina, USA, extending fromthe North Carolina border north to Fontana Dam andincluding all of Macon County and parts of Jackson,Swain, and Clay counties. This 104 000-ha area ismountainous with a diverse complement of forests andhuman residents. The region has grown considerablyin the last two decades, with the population in MaconCounty expanding by ;20% between 1980 and 1990(SAMAB 1996b). Most growth has radiated fromFranklin, the Macon County seat, which also serves asthe dominant central place for most services and com-merce.

Data for the estimation of statistical land-coverchange models were compiled as layers in the Geo-graphic Resources Analysis Support System (GRASS)geographic information system (USACERL 1991, seeTurner et al. 1996 for complete description). For thisanalysis we used the following data layers:

1) Land cover interpreted for 1975 and 1980 fromLandsat Multi-spectral Scanner (MSS) imagery. Threeclasses were identified: (1) forest, (2) grassy, and (3)unvegetated.

2) Slope, aspect, and elevation derived from 7.5 foot(2.286 m) digital elevation model (DEM) data obtainedfrom the U.S. Geological Survey.

3) Land ownership digitized from U.S. Forest Ser-vice and other maps.

4) Primary and secondary road locations and pop-

ulation density from TIGER/Line Census files in ARC/Info format.

All data were converted to a raster format using anapproximate 90-m square grid cell (cell size was lim-ited by the resolution of the MSS imagery).

METHODS

We considered urbanization to be a factor or gradientthat organizes disturbance regimes and ecosystems ingeneral (see McDonnell and Pickett 1990). The urban-ization gradient is best viewed as describing a wholecomplex of processes that have a bearing on ecosystemstructure and function. Two somewhat problematicmeasures of position along an urban–rural gradient arecandidates: human population density and distance tothe closest metropolitan area. For this study we choseto measure position along the gradient as the traveldistance along roads to the closest city (scaling shouldnot be a problem, because we measure distance withrespect to only one city). However, we also includepopulation density as an explanatory variable in theempirical analysis of land-cover changes and accountfor its covariance with distance. Human populationdensity was not used to measure position along thegradient because, while it is measured by the U.S. Cen-sus Bureau for relatively small tracts, the size of thetract increases as population density decreases. Het-erogeneity within tracts therefore increases towards therural end of the gradient.

We applied both empirical and simulation analysisto study the effects of the urban–rural gradient on land-cover patterns. First we extended previous empiricalanalysis of land-cover change in the southern Appa-lachians (Turner et al. 1996) to test for effects of po-sition along the gradient on probabilities of land-coverchanges. We measured land-cover change from satelliteimages and related the observed change (or lack ofchange) to various topographic, social, and locationalvariables, including position on the urban–rural gra-dient. We tested for the effect of position on the urban–rural gradient after accounting for the influence of othervariables through regression analyses. Comparing dif-ferent forms of the regression models defines tests forthe extent and form of the influence that position onthe gradient has had on land-cover changes. Estimatedregression models also provide useful insights into theprobability that specific changes may occur. However,it is the long-term interaction of these changes thateventually shape a landscape.

To examine the potential cumulative impacts of allland-cover changes taken together, we constructed asimulation model that applies the empirical land-coverchange models to a specific landscape for a long periodof time. To isolate the effect of position along the ur-ban–rural gradient on landscape development, we con-structed a hypothetical landscape with topographic andsocial conditions in the range of those observed in the


study area. We then simulated the development of thishypothetical landscape at different positions along theurban–rural gradient and measured various features oflandscape structure at each position.

Statistical models of land-cover change

Observed land-cover changes are the outcomes ofdecisions by landowners regarding the uses of theirlands. Following Turner et al. (1996), we posited thatland use choices reflect the highest valued utilizationof the site, and changes in land use reflect changes inthe ordering of alternative use values. While utilityderived from various uses is not observed, the out-comes of, and inputs to, utility comparisons are ob-served. We assumed that the utility derived from chang-ing land uses is a function of variables that describethe topography and locational characteristics of a site(e.g., slope, distance to road, and distance to a marketcenter), and therefore influence both the benefits andcosts associated with each land use. Because we ex-amine land-cover change for a single time period in arelatively small area, we assume that the values of pro-duced goods and services are constant across sites. Ac-cordingly, differences in land use between sites de-pends on site-specific features that influence the pro-duction of goods and services and the costs related tothe operability and accessibility of the site. The netbenefits of a site dedicated to land use i (NBi) can bedescribed as:

NBi 5 B(pr, X) 2 C(X) (1)

where B(pr, X) defines benefits that depend on pricesof produced goods and services, pr, and site charac-teristics, X, and C(X) defines costs that are dependentonly on the structural and locational characteristics ofthe site. Pr are fixed across the study area, so spatialdifferences in use-values within the study area are com-pletely described by the attributes of the site (X).

The probability of land-cover change can be esti-mated using a multinomial Logit model (Turner et al.1996). The probability of land cover for a grid cellchanging from an initial class i to class j (Pij) is definedas:

(X9b )jeP 5 j 5 1, . . . , k (2)ij k

(X9b )seOs51

where k is the number of land-cover states. For eachinitial land-cover class i, k land-cover change equationsare defined (including the null case of no change).Among explanatory variables contained in X is the dis-tance from the grid cell along roads to the closest citywhere products could be marketed or services pro-cured. We use this variable to define the position ofthe cell along the urban–rural gradient.

Equation 2 defines the simple case where distancehas an additive relationship to other variables in the

Logit model. To test for more complex interactionsbetween position along the urban–rural gradient andland-cover probabilities, we defined a varying-param-eters model that allows the effect (coefficient) of eachX variable to vary with position along the gradientthrough a set of interaction terms. A quadratic varying-parameters model is implemented by introducing theset of interaction terms into the Logit model as follows:

(X9a )jeP 5 j 5 1, . . . , k (3)ij k

(X9a )jeOs51

where ai 5 aj0 1 Daj1 1 D2aj2, D is the distance variableand X now includes the intercept and all structural andlocation variables describing the site except distance.

The models (Eq. sets 2 and 3) were estimated usinga sample of observations on land-cover change between1975–1980 and measures of the following variables(elements of X): (1) slope, (2) elevation, (3) populationdensity, (4) distance to the closest road, and (5) distancealong road to the city of Franklin, North Carolina. Sep-arate equation sets were estimated for the three initialcover types and for public and private lands. Modelswere estimated using a maximum likelihood procedurein the software package LIMDEP (Greene 1995).

The significance of models was tested against thenull of no explanatory power using log-likelihood ratiotests. In addition, because equation set (2) is a con-strained version of (3), we tested the significance ofthe varying parameters model against the null of thelinear model using log-likelihood ratio tests (see Judgeet al. 1985). The effect of position along the urban–rural gradient on land-cover changes was similarly test-ed by estimating constrained models with all coeffi-cients related to the distance variable set to zero andthen comparing likelihood values with the appropriateunconstrained model (standard tests on individual co-efficients were not applicable because of multiple co-efficients related to each variable). Other explanatoryvariables were tested in the same manner.

Simulation experiment

To examine the effects of position along the urban–rural gradient on landscape structure and water quality,we conducted a landscape simulation experiment. Thesimulation approach was applied because physicallysimilar areas could not be identified at regular intervalsalong the gradient. Instead we simulated developmentof a common hypothetical landscape using the statis-tical estimates of land-cover change models describedin the preceding section. Comparisons were construct-ed by simulating development at various distances fromFranklin. Moving from urban to rural positions alongthe gradient also defined a development sequence forthe landscape. Comparisons between adjacent land-scape positions indicate how land would develop as the


TABLE 1. Range of variables in data set used for statisticalanalyses.

Name UnitsMini-mum

Maxi-mum

SlopeDistance to roadDistance to FranklinPopulation

Elevation

degreescellscellsindividuals/square

milemeters

118

15

521

3644

6181415

1521

FIG. 1. (A) Elevation profile (meters) and road location for the hypothetical landscape and (B) positioning of the hy-pothetical landscape for simulation scenarios (scenario labels indicate the distance in cell lengths from Franklin, NorthCarolina, to the lower left corner of the landscape).

edge of the city moved outward, and where changeswould be the most severe.

The hypothetical landscape.—A landscape charac-teristic of the study area in terms of the independentvariables (X) was constructed by using observed vari-able ranges (Table 1) as a general guide. The landscapewas defined as a 50 3 50 one-hectare cell matrix fora total size of 2500 ha. Topography was defined asshown in Fig. 1 and slope was calculated using thegreatest absolute difference in elevation between ad-jacent cells. A road was placed along the diagonal ofthe landscape with the lower left corner being closestto Franklin. Distance to the closest road was calculatedin cell lengths. Position along the urban–rural gradientwas calculated differently for each scenario.

The density of population varies along the urban–rural gradient, with highest concentration in Franklinand declines with distance. To assign population den-sity to individual cells, we estimated a regression modelthat defined population density (PD) as a function of

distance from town and the slope of the site. We usedan inverse logistic functional form,

1PD 5 K 1 2 (4)

a1b DTOT1b SLOPE1 21 21 1 e

and the model was estimated using ordinary leastsquares applied to a log-odds form:

Zlog 5 a 1 b DTOT 1 b SLOPE (5)1 21 21 2 Z

where Z 5 PD/K and K is the maximum populationdensity for the rural areas of the study site and DTOTis a measure of total distance (distance to Franklinalong roads plus two times the distance to the closestroad). K was set at 176 individuals/km2 (only 5% ofthe area had a density .176 individuals/km2). Becausescenarios varied according to the distance variable,population density was calculated differently for eachscenario. For scenarios involving U.S. National For-ests, the population density was set at a constant 9.54individuals/km2.

Simulation structure.—Each simulation started witha completely forested landscape. Estimated land-coverchange equations were then used to estimate the tran-sition probabilities for each of the 2500 cells and land-cover changes were implemented using comparisonswith random number draws. This process was com-pleted for 50 five-year time steps (by 50 time stepseach simulated landscape had reached a stable structurein terms of amount of cover by type) and a map of theresulting landscape was stored for subsequent analysis


TABLE 2. Significance tests for Multinomial Logit modelsof land-cover change. Log-likelihood ratios are reportedfor testing overall significance of the referenced model.

Ownership

Initial land cover

National forest

Linear

Varyingparam-

eter

Private

Linear

Varyingparam-

eter

ForestGrassyUnvegetatedDegrees of freedom

31.15*NA

8.5110

53.65*NA

19.6628

205.80*36.42*37.13*10

282.33*74.14*58.11*28

Note: Critical chi-squared values are 41.34 (28 df) and18.31 (10 df).

* Indicates significance (P 5 0.05).

(each scenario was replicated five times). To compareland-cover patterns over the urban–rural gradient, thissimulation experiment was completed at six differentdistances from Franklin for both public and privatelands. The hypothetical landscape was placed first atthe edge of Franklin and then was moved away fromthe city at 100 cell-length intervals (10 km; Fig. 1) tocover the range of the distance variable in the test dataset (see Table 1). We label these scenarios by theirdistance from Franklin; i.e, ‘‘0,’’ ‘‘100,’’ and so on.

Analysis of simulated landscapes.—Each simulatedlandscape was analyzed using a landscape analysissoftware package (SPAN, Turner 1990). The amountof cover, number of patches, and area-weighted patchsize were computed for each of the three cover types.Patches were defined as groups of contiguous adjacent(horizontally or vertically) cells of the same cover type.Total edge between all cover types was calculated alongwith two overall landscape indices. Dominance, D,measures the deviation from maximum possible land-scape diversity and ranges from zero to one, with highervalues indicating a landscape dominated by one or fewcover types (O’Neil et al. 1988). Contagion, C, mea-sures the aggregation of land-cover types and also rang-es from zero to one (Li and Reynolds 1993). Highvalues of C indicate a clumped distribution of landcover, whereas low values indicate a dispersed arrange-ment.

These landscape metrics can be generally linked towater quality. Research by Swank and Bolstad (1994)in the Coweeta Creek watershed, a part of the LittleTennessee River Basin (Coweeta Creek is 4350 ha orroughly twice the size of our hypothetical watershed),indicates significant correlations between landscapemetrics and chemical, physical, and bacterial measuresof baseflow water quality. Of the landscape metricsthey considered, the percent of non-forest land use andthe density of buildings defined significant relation-ships with the largest number of water quality vari-ables. In addition, findings of a recent assessment ofthe southern Appalachian region ascribe two-thirds of

all water quality impacts with activities associated witheither mining or non-forest land uses (SAMAB 1996a).Hunsaker and Levine (1995) similarly found that per-centage of non-forest cover has a strong influence onwater quality in rivers in Illinois and Texas. They alsofound contagion and dominance of land cover to berelated to various measures of water quality.

These findings suggest using the share of non-forestland within the watershed as an indicator of overalllandscape effects on water quality. To further capturethe effects of spatial pattern on water quality, we con-structed another measure that weighted non-forest cov-er by proximity to the major stream in our hypotheticallandscape (Fig. 1). The resulting non-forest index (NFI)is:

n n Yi,jNFI 5 Y (6)O O max1 2 @) )(d 1 1)i51 j51 i,j

where Yij is a binary variable equal to 1 when cell i,jis in non-forest cover (zero otherwise), dij is the dis-tance from cell i,j to the stream in number of cells(because a cell containing a stream cell has distanceequal to zero, we add one cell length to each distancein the denominator), and Ymax is the maximum possiblevalue of the sum in brackets (i.e., if all cover was non-forest). NFI therefore ranges from 0 for no non-forestcover effects, to 1 for maximum effects, and waterquality should be inversely related to the non-forestindex.

Post-simulation analysis.—Landscape metrics werecompared between simulated landscapes. ANOVA wasused to do pairwise comparisons of means of the land-scape metrics for adjacent scenarios along the urban–rural gradient (the GLM procedure in SAS was used;SAS Institute 1990). In addition, public and privatelands were compared at each location along the gra-dient.

RESULTS

Land-cover change models

Based on likelihood ratio tests, estimated land-coverchange models (Eq. 2 and 3) are different for publicand private lands (Table 2). For private land, the simplelinear formulation (Eq. 2) proved significant in ex-plaining land-cover transitions for all three initial covertypes (forest, grassy, and unvegetated). For publiclands, only the model for transitions from forest provedsignificant using Eq. 2. We could not reject the nullhypothesis of no explanatory power for the linear mod-el applied to the unvegetated cover type (there werenot enough observations to estimate the models forgrassy cover on public lands). This is consistent withthe observation that almost all cells with non-forestcover types in 1975 became forest in 1980 on U.S.National Forest land.

For private lands, the varying parameters models


TABLE 3. Tests of significance of varying parameters vs.linear formulation of land-cover change models (null is thelinear model). Log-likelihood ratios are reported.

Ownership

Initial landcover

National forest Private

Log-likelihood ratios

ForestGrassyUnvegetated

22.5NA

NA

76.52*37.70*20.98*

Note: Critical chi-squared value is 28.87 (df 5 18).* Indicates significance (P 5 0.05).

TABLE 4. Tests for significance of individual variables inspecified land-cover change models. Log-likelihood ratiosare for reported testing models constrained to exclude thereferenced variable (numbers in parentheses are degrees offreedom).

Initial landcover Variable

Ownership

Nationalforest Private

Forest Gradient positionDistance to roadElevationSlopePopulation

0.98 (2)6.60 (2)*4.70 (2)3.92 (2)2.70 (2)

83.44 (20)*14.06 (6)*39.44 (6)*18.57 (6)*

5.41 (6)

Grassy Gradient positionDistance to roadElevationSlopePopulation

51.32 (20)*4.22 (6)5.94 (6)

14.46 (6)*10.76 (6)

Unvegetated Gradient positionDistance to roadElevationSlopePopulation

12.34 (2)*9.32 (2)*

15.98 (2)*10.12 (2)*16.42 (2)*

Note: Critical chi-squared values are 5.99 (2 df), 12.59 (6df), and 31.41 (20 df).

* Indicates significance (P 5 0.05).

(Eq. 3) provided a significant improvement over thelinear formulation (Eq. 2) in the explanation of land-cover changes for forest and grassy cover types (Table3). This indicates that position on the urban–rural gra-dient significantly interacts with the other variables indefining the propensity of landowners to change landcover. The implication for the effect of slope, for ex-ample, is that slope is less of a constraint on clearingforests closer to Franklin than farther away. For un-vegetated private land the varying-parameters modeldid not provide significant improvement over the linearformulation. On public lands, we conducted the test forforest transitions only and found that the varying pa-rameters model provided no significant improvementfor explaining land-cover change on this ownership,indicating an additive rather than multiplicative effectof position on the urban–rural gradient on land-coverchanges.

Tests of the significance of position on the urban–rural gradient and other explanatory variables were ap-plied to the models indicated by the preceding tests.That is, we tested their effects in varying-parametersmodels for grassy and forested land in private own-ership and the linear formulation for unvegetated landon forested land in public ownership (no such testswere constructed for grassy and unvegetated land inpublic ownership). For all models estimated for privatelands, position on the urban–rural gradient had a sig-nificant influence on land-cover change regimes (Table4). Likewise, slope had a significant influence on allchange models. Results for the other variables differedbetween models. Distance to road, elevation, and pop-ulation density were all significant in explainingchanges from unvegetated cover and insignificant inexplaining changes from grassy cover. All variablesexcept population density were significant in explain-ing changes from forest cover on private lands. Look-ing across all models for private lands, of the 15 testsof explanatory variables (five variables, three models),11 indicated significant effects.

The results on public lands were much different.Only one model, the one for forest cover, was signif-icant. Of the five variables examined for this model,only distance to road was significant in explainingchanges from forest cover. For position on the urban–

rural gradient, slope, elevation, and population densitywe could not reject the null hypothesis of no significanteffects.

Constructing the simulation model also required fit-ting the population density model, Eq. 5, using ordinaryleast squares. The coefficient estimates (a 5 20.8821,b1 5 20.0035, and b2 5 20.0158) were all significant(P 5 0.01) and indicated the expected inverse rela-tionships between distance and slope and populationdensity. The overall model was significant (P 5 0.01)according to the standard F test.

Simulation results

We used the land-cover change models indicated bythese tests to construct simulations of the 12 distancescenarios (six positions on the urban–rural gradient forthe two ownerships). For private lands we applied thevarying-parameters models of cover changes using Eq.3. For public lands we applied a binary Logit modelto explain changes from forest cover to non-forest cov-er (that is, we combined grassy and unvegetated coverinto a non-forest cover class) which we categorize as‘‘unvegetated.’’ All unvegetated cover was then con-verted back to forest cover in the subsequent period.This is consistent with observed land-cover changes onNational Forests in the study area, and warranted byour statistical findings (i.e., insignificant land-coverchange models).

The results of one simulation run for each of the 12scenarios are shown in Fig. 2 and mean values of land-scape measures (five replications of each scenario) areshown for each scenario in Table 5. For private land(Table 5), the amount of forest cover increases sub-


→

FIG. 2. Landscapes resulting from simulating land-coverchange over 50 five-year time steps for each distance scenarioapplied to private and public land. Scenarios are defined bydistance from Franklin, North Carolina (0 cells [top] to 500cells [bottom] in increments of 100). Gray cells representforest cover while white and black represent unvegetated andgrassy cover, respectively.

stantially between the 0 and 100 distance scenarios,remains fairly constant from 100 through 300, and thenincreases again to distance scenarios 400 and 500.While the amount of non-forest cover decreases withdistance from the city (though in an irregular pattern),it becomes less concentrated along this gradient. Thisis reflected in the patchiness of the landscape (see totalpatches in Table 5), which increases substantially be-tween 0 and 100, peaks between the 200 and 300 dis-tance scenarios, and then falls off in the 500 distancescenario. Total edge similarly is relatively low (600cell lengths) close to town and increases to a peak at300 and then declines to ,600 cell lengths at the 500distance scenario. According to this metric then thelandscape is most complex in the middle of the gra-dient.

The non-forest index also differs substantially byposition along the urban–rural gradient. On privateland, its highest value is at the urban end of the gradientwith a value of 0.74 (recall that the non-forest indexis scaled between 0 and 1). It ranges between 0.38 and0.52 for scenarios 100 through 400, and declines sub-stantially to 0.20 in distance scenario 500. For publicland, NFI shows a much smaller range, from 0.04 indistance scenario 0 to 0.12 in scenario 500.

For private lands, a majority of landscape attributeswere significantly different between adjacent distancescenarios (e.g., 0 vs. 100, 100 vs. 200, etc. . . . , ac-cording to ANOVA tests). Of the 21 attributes evalu-ated, 10 were statistically different (P , 0.05) betweenevery adjacent pair of distance scenarios. In contrast,measures of landscape pattern on public land were notstatistically different between any adjacent pairs of sce-narios. Differences in attributes were much smaller be-tween scenarios for public land (Table 5). For publicland, the share of the landscape that is in forest coverfalls from 2446 ha (98%) adjacent to the city to 2331ha (93%) for distance scenario 500. In addition, thetotal number of patches and the amount of edge in-creases from scenario 0 through scenario 500. Whilethe range of effects is narrow, the model suggests thatforest management activities increase somewhat to-wards the rural end of the urban–rural gradient.

At each position along the urban–rural gradient wealso compared the two ownerships using ANOVA. Fordistance scenarios 0 through 400, comparisons of land-scape metrics indicated significant differences betweenthe private and U.S. National Forest ownerships for allmetrics compared. For distance scenario 500, however,


TABLE 5. Measures of landscape pattern for six scenarios applied to private land (A), and six scenarios applied to publicland (B) (scenario labels refer to distance from the edge of the hypothetical landscape to Franklin, North Carolina, in celllengths). Values are means for five simulations of the hypothetical landscape.

Variable

Distance constant

0 100 200 300 400 500

A) Private landAmount of cover (hectares)

forestgrassyunvegetated

1448.244.2

1007.6

1825215.8459.2

1716.2482.2301.6

1691.8303.6504.6

196156.2

482.8

2336.231.8

132

Landscape indicesdominancecontagion

0.31380.728

0.31520.6526

0.24380.5956

0.23220.6112

0.46020.7106

0.75080.7998

Edge (100 m)forest : grassyforest : unvegetatedgrassy : unvegetated

Total

82.846756.2

606

486.2438.8164.6

1089.6

893.6540.6215.4

1649.6

723.4876.2156.6

1756.2

174.21109.2

39.81323.2

101422.8

23546.8

Largest patch (hectares)forest 936.8 1772 1212.6 930.6 1852.8 2332.6grassyunvegetated

8.4946.2

23.2354.6

83.262.6

18.2186.8

353.8

1.46.8

Average patch size(hectares)


43.941.74

23.42

59.261.865.7

21.863.542.4

25.21.983.58

43.741.083.06

739.71.021.36

Weighted average patchsize (hectares)


820.923.82

888.88

1720.845.86

274.92

890.8428.7223.14

706.244.86

91.46

1751.061.22

14.36

23291.041.96

Number of patchesforestgrassyunvegetated

Total

33.826.243.4

103.4

31.8117.2

81.2230.2

79.4137126.0342.4

67.8151.8141.2360.8

4551.4

157.6254.0

3.231.297.0

131.4

Non-forest index 0.74 0.52 0.42 0.51 0.38 0.2

B) Public landAmount of cover (hectares)

forestunvegetated

2445.854.2

243763

2420.879.2

2395.6104.4

2356.8143.2

2330.8169.2

Landscape indicesdominancecontagion

0.9050.9574

0.89280.9506

0.8720.9424

0.8420.9392

0.80020.9208

0.77480.905

Edge (100 m)forest : unvegetated

Total204.2204.2

236.6236.6

293.4293.4

385.8385.8

514.4514.4

582.4582.4

Largest patch (hectares)forestunvegetated

2445.82.2

24372.6

2420.83

2395.43

2356.64.4

2329.86

Average patch size(hectares)

forest 2445.8 2437 2420.8 2395.4 2356.6 2329.8unvegetated 1.12 1.12 1.14 1.16 1.22 1.34

Weighted average patchsize (hectares)

forestunvegetated

2445.81.2

24371.24

2420.81.32

2395.21.32

2356.41.46

2328.81.84

Number of patchesforestunvegetated

Total

149.250.2

156.257.2

169.070.0

1.290.892.0

1.2117.2118.4

2126.2126.4

Non-forest index 0.04 0.05 0.06 0.08 0.10 0.12


the pubic and private landscapes did not have signifi-cantly different land cover, patchiness, or forest edge.Only at this most remote location did private and publiclands have similar attributes.

Evaluated in sequence, these simulations thereforeshow characteristic landscape patterns at each pointalong the urban–rural gradient. Distance scenario 0,which is on the urban edge, is dominated by largeamounts of unvegetated and forest cover. Both covertypes are highly aggregated, leading to low patchinessand high contagion. Distance scenario 0 reflects a ma-ture, urbanized landscape profile, with a large share ofunvegetated cover that is essentially stable and highlyconcentrated. Distance scenario 100 is consistent withthe outer envelope of urbanization. Unvegetated coveris agglomerated along the road closer to town (lowerleft-hand corner), but signs of development taper offwith distance (upper right-hand corner). The share ofunvegetated cover falls by 54% from distance scenario0. The contrasts in landscape structure between dis-tance scenarios 0 and 100 are substantial: the weightedaverage patch size and largest patch for unvegetatedcover fall by 69% and 63%, respectively.

Distance scenarios 200 and 300 are consistent withan agricultural, mixed-use landscape with high quan-tities of grassy and unvegetated cover spread over amuch broader range of landscape conditions. Whilethese two scenarios result in an area of forest coverthat is similar to distance scenario 100, the cover ismuch patchier; numbers of patches and the amount ofedge is greater for all cover types. Dominance indicesindicate that distance scenarios 200 and 300 are theleast dominated and the most diverse in terms of extentof cover.

Distance scenario 400 is consistent with transitionfrom the agricultural profile in scenarios 200 and 300to a remote forest profile in distance scenario 500. Non-forest cover is concentrated in areas close to the roadand with gentle slopes. However, non-forest cover isnot as agglomerated as in scenarios 200 and 300, con-sistent with increased forest harvesting activities in themore remote areas. At scenario 500, non-forest coveris highly diffuse and forest covers 93% of the land-scape.

By looking at differences between adjacent locationson the urban–rural gradient we can also examine howcontinued development could affect landscapes. Thatis, they can be used to show what would happen ifdevelopment moves the city outward by 10 km. Figure3 shows the changes in several landscape indices forthis type of development on both private and publiclands. This figure demonstrates most clearly that thereis little difference between scenarios on public lands,especially when compared with private lands. Changesin indices are almost an order of magnitude greater onthe private lands.

On private lands, changes in landscape structure

would not be constant over the urban–rural gradient.Total change in land cover is greatest at the two endsof the gradient: at the edge of the city and in areascurrently dominated by forest cover. Decreases in forestcover and increases in non-forest cover are also con-centrated at the two ends of the urban–rural gradient.Change in the non-forest index which weights non-forest cover by its proximity to the major stream in thelandscape also has a U-shape, indicating that the largestincreases in the non-forest index would also occur atthe urban and rural ends of the gradient. The greatestchanges in dominance and contagion occur where land-scapes shift from scenario 500 to scenario 400 and from400 to 300. Here the dominance of forest cover de-creases substantially and contagion falls as well. Bothsuggest an increase in the fragmentation of the resultinglandscapes. This is borne out in changes in the numberof patches as well as the weighted average patch sizebetween these scenarios.

DISCUSSION

Land-cover changes and landscape structure

Estimates of the land-cover change models and re-lated hypothesis tests provide insights into the structureof landscape dynamics. Our findings confirm findingsfrom previous studies in the southern Appalachians(Wear and Flamm 1993, Turner et al. 1996) and indicatethat forest disturbance regimes and land-cover changesare significantly related to site quality and location vari-ables, and that structural differences exist between pub-lic and private land-cover change regimes. The presentanalysis shows further that position on the urban–ruralgradient has a significant impact on land-cover changeregimes and gives rise to substantially different land-scape structures. Furthermore, position on the urban–rural gradient was found to have a compounding influ-ence on land-cover change (and therefore pattern) onprivate lands, by interacting with other site variablesin significant ways.

Our findings indicate that position on the urban–ruralgradient not only influences land-cover change re-gimes, but could have a substantial influence on re-sulting landscape patterns. For nearly every landscapemeasure we found significant differences between spa-tially adjacent scenarios on private lands. These find-ings suggest that institutions and location interact todefine unique ‘‘landscape signatures.’’

Water quality implications

Land use is clearly one of the most important factorsdetermining water quality (Allan and Flecker 1993).The phrase ‘‘water quality’’ refers to the biophysicalconstituents affecting the character of the aquatic en-vironment. Aquatic environments (or streams) in thestudy area are naturally cool, clear, typically shaded,and of high chemical quality. As such, they are highlysensitive to nutrient enrichment, temperature altera-


FIG. 3. Changes in landscape measures on private and public lands resulting from shifting the hypothetical landscape10 000 m (100 cell lengths) closer to Franklin, North Carolina. These show the implied effects of development that wouldmove the city edge outward by 10 km.

tions, introduction of suspended solids, and acid pre-cipitation. Removal of riparian vegetation, as well asalteration of organic inputs and hydrologic regimes byforest and agricultural activities and expanding urban-ization, generally results in increased erosion, in-creased algal production, changes to temperature re-gimes, and reduced concentrations of dissolved oxygen(Welch et al. 1998). Erosion of fine sediments are gen-

erally higher at locations where land use activities, suchas road construction, logging, agriculture, or grazingexpose soils. Algal communities respond positively tonutrient enrichment causing aesthetic, water quality,and habitat degradation. Changes in water temperatureregimes impact all aspects of the physiology, behavior,and life history strategies of aquatic organisms (Nai-man and Anderson 1997). The extent of oxygen de-


FIG. 3. Continued.

pletion can be severe but ultimately depends on themagnitude of organic matter and nutrient inputs, watertemperature, and discharge volume.

Patterns of land use/cover serve as useful indicatorsof overall water quality. Hunsaker and Levine (1995)found, in river basins in Texas and Illinois, that thepercentage of land in forest and other uses were thebest predictors of overall water quality. Similarly,Swank and Bolstad (1994), working in the southernAppalachians (in Coweeta Creek, which is containedin our study area), found the percentage of land in non-forest cover and the density of paved roads among themost important variables influencing baseline waterquality, and additionally defined a strong link betweenland cover and water quality changes associated withmajor storm events. Their findings suggest that evensmall changes in non-forest land cover have importantimplications for water quality in their headwater drain-age (no site in their study area has .6% non-forestcover). Hunsaker and Levine also found some evidencethat the spatial organization of land cover, measuredby contagion and dominance, may also have a bearingon water quality. Changes in ‘‘landscape signatures’’may therefore have important implications for water

quality. Understanding that these are only rough in-dicators of potential effects on water quality, the resultsof the simulation exercise may provide some insightsinto where water could be disproportionately impactedby continued development.

The development sequence defined by moving fromdistance scenario 500 (rural) to distance scenario 0 (ur-ban) shows that the impacts of development on land-cover patterns would vary across the urban–rural gra-dient (Fig. 3). However, the greatest impacts of de-velopment may occur in two positions: (1) where landis moving into an urban use (i.e., moving from distancescenario 100 to distance scenario 0) and (2) where landis moving out of a remote use (i.e, from 500 to 400).In these two locations, total land-cover change, as wellas declines in forest cover are at their greatest. To aneven greater extent, increases in the non-forest indexare concentrated at the two ends of the gradient. In-creases in unvegetated cover are likewise concentratedin these areas. To the extent that unvegetated cover isassociated with built up areas and impervious surfacearea, water quality may be heavily influenced by thesechanges. Impervious cover has especially importantimpacts on storm surges and stream concentrations of


associated pollutants (Swank and Bolstad 1994, Welchet al. 1998). All of these findings suggest that potentialwater quality impacts are not spread evenly throughoutthe watershed, nor are they a simple decreasing func-tion of distance from the city. Rather they would beconcentrated at the two ends of the urban–rural gra-dient.

The methodology applied here could be extended.For example, the hypothetical landscape, while con-structed to reflect a range of conditions observed in thestudy area, is essentially arbitrary. Landscape analysiscould be constructed for actual places with similar tech-niques. In addition, these types of models can be ap-plied directly to existing landscapes to define whereand how land cover is most likely to change (see Wearet al. 1996). Resulting ‘‘hazard maps’’ could be es-pecially useful for defining short-run implications ofdevelopment within an area. These models are basedon observations of change over relatively short periods.The rate as well as the pattern of change may be vari-able even within the relatively short period of fifteenyears (Turner et al. 1996). Work that extends the anal-ysis to address different epochs of human use couldimprove our understanding of the full range of effectsand of the evolving interactions between people andtheir environment. A longer time series or more spatialbreadth could provide data and insights into the for-mation of value through land markets and, therefore,more direct information on the economic mechanismsbehind land-use and cover changes (see Bockstael1996).

The methods we describe provide one approach tounderstanding how human processes act to structurelandscapes and ecosystems. Better information on theinteractions between human populations and nature,both in terms of how people impact their environmentand how environmental quality contributes to humanwell-being, should improve our ability to structure ra-tional policies and management strategies for protect-ing environmental quality in developing areas.

ACKNOWLEDGMENTS

This work was supported by the U.S. Man and the Bio-sphere Program and by the Economics of Forest Protectionand Management Research Work Unit, Southern ResearchStation, U.S.D.A. Forest Service. Our work has benefittedfrom discussions with Robert G. Lee, Robert R. Gottfried,Richard O. Flamm, and Michael Berry. We thank DianeRiggsbee and John Pye for their assistance in preparing thefigures.

LITERATURE CITED

Allan, J. D., and A. S. Flecker. 1993. Biodiversity conser-vation in running waters. BioScience 43:32–43.

Bockstael, N. E. 1996. Modeling economics and ecology:the importance of a spatial perspective. American Journalof Agricultural Economics 78:1168–1180.

Greene, W. H. 1995. LIMDEP: User’s manual and referenceguide, version 7.0. Econometric Software, Bellport, NewYork, USA.

Hunsaker, C. T., and D. A. Levine. 1995. Hierarchal ap-

proaches to the study of water quality in rivers. BioScience.45:193–203.

Judge, G. G., W. E. Griffiths, R. C. Hill, W. E. Lutkepolh,and T. C. Lee. 1985. The theory and practice of econo-metrics. John Wiley and Sons, New York, New York, USA.

Lee, R. G., R. O. Flamm, M. G. Turner, C. Bledsoe, P. Chan-dler, C. DeFerrari, R. Gottfried, R. J. Naiman, N. Schu-maker, and D. Wear. 1992. Integrating sustainable devel-opment and environmental vitality. Pages 499–521 in R.J. Naiman, editor. New perspectives in watershed manage-ment. Springer-Verlag, New York, New York, USA.

Li, H., and J. F. Reynolds. 1993. A new contagion index toquantify spatial patterns of landscapes. Landscape Ecology8:155–162.

McDonnell, M. J., and S. T. A. Pickett. 1990. Ecosystemstructure and function along urban–rural gradients: an un-exploited opportunity for ecology. Ecological Applications71:1232–1237.

Naiman, R. J., J. J. Magnuson, D. M. McKnight, and J. A.Stanford, editors. 1995. The freshwater imperative. IslandPress, Washington, D.C., USA.

Naiman, R. J., and E. C. Anderson. 1997. Streams and rivers:their physical and biological variability. Pages 131–148 inP. K. Schoonmaker, B. von Hagen, and E. C. Wolf, editors.The rain forests of home: profile of a North American biore-gion. Island Press, Washington, D.C., USA.

O’Neil, R. V., J. R. Krummel, R. H. Gardner, G. Sugihara,B. Jackson, D. L. DeAngelis, B. T. Milne, M. G. Turner,B. Zygmunt, S. Christensen, V. H. Dale, and R. L. Graham.1988. Indices of landscape pattern. Landscape Ecology 1:153–162.

SAMAB (Southern Appalachian Man and the Biosphere).1996a. The Southern Appalachian Assessment AquaticsTechnical Report. Report 2 of 5. U.S. Forest Service, At-lanta, Georgia, USA.

. 1996b. The Southern Appalachian Assessment So-cial, Cultural, and Economic Technical Report. Report 4of 5. U.S. Forest Service, Atlanta, Georgia, USA.

SAS Institute. 1990. SAS/STAT User’s Guide, Version 6,Volume 2. SAS Institute, Cary, North Carolina, USA.

Swank, W. T., and P. V. Bolstad. 1994. Cumulative effectsof land use practices on water quality. Pages 409–421 inN. E. Peters, R. J. Allan, and V. V. Tsirkunov, editors.Hydrochemistry 1993: hydrological, chemical and biolog-ical processes affecting the transformation and the transportof contaminants in aquatic environments. Proceedings ofthe Rostov-on-don Symposium, May 1993. IAHS Publi-cation 219. IAHS Press, Oxfordshire, UK.

Turner, M. G. 1990. Spatial and temporal analysis of land-scape patterns. Landscape Ecology 4:21–30.

Turner, M. G., D. N. Wear, and R. O. Flamm. 1996. Influenceof landownership on land cover dynamics in the SouthernAppalachian Highlands and the Olympic Peninsula. Eco-logical Applications 6:1150–1172.

USACERL. 1991. Geographic resources analysis supportsystem, version 4.0. U.S. Army Corps of Engineers, Con-struction Engineering Research Laboratory, Champaign, Il-linois, USA.

Wear, D. N., and R. O. Flamm. 1993. Public and privatedisturbance regimes in the Southern Appalachians. NaturalResource Modeling 7:379–397.

Wear, D. N., M. G. Turner, and R. O. Flamm. 1996. Eco-system management with multiple owners: landscape dy-namics in a Southern Appalachian watershed. EcologicalApplications 6:1173–1188.

Welch, E. B., J. M. Jacoby, and C. W. May. 1998. Streamquality. In R. J. Naiman and R. E. Bilby, editors. Riverecology and management: lessons from the Pacific CoastalEcoregion. Springer-Verlag, New York, New York, USA,in press.

Land Cover Along an Urban-Rural Gradient: Implications for Water Quality

Documents

Transcript of Land Cover Along an Urban-Rural Gradient: Implications for Water Quality