Predicting abundance of 80 tree species following climate change in the eastern United States

465

Ecological Monographs, 68(4), 1998, pp. 465–485q 1998 by the Ecological Society of America

PREDICTING ABUNDANCE OF 80 TREE SPECIES FOLLOWING CLIMATECHANGE IN THE EASTERN UNITED STATES

LOUIS R. IVERSON AND ANANTHA M. PRASAD

U.S. Department of Agriculture Forest Service, Northeastern Research Station, 359 Main Road,Delaware, Ohio 43015 USA

Abstract. Projected climate warming will potentially have profound effects on theearth’s biota, including a large redistribution of tree species. We developed models toevaluate potential shifts for 80 individual tree species in the eastern United States. First,environmental factors associated with current ranges of tree species were assessed usinggeographic information systems (GIS) in conjunction with regression tree analysis (RTA).The method was then extended to better understand the potential of species to survive and/or migrate under a changed climate. We collected, summarized, and analyzed data forclimate, soils, land use, elevation, and species assemblages for .2100 counties east of the100th meridian. Forest Inventory Analysis (FIA) data for .100 000 forested plots in theEast provided the tree species range and abundance information for the trees. RTA wasused to devise prediction rules from current species–environment relationships, which werethen used to replicate the current distribution as well as predict the future potential distri-butions under two scenarios of climate change with twofold increases in the level of at-mospheric CO2. Validation measures prove the utility of the RTA modeling approach formapping current tree importance values across large areas, leading to increased confidencein the predictions of potential future species distributions.

With our analysis of potential effects, we show that roughly 30 species could expandtheir range and/or weighted importance at least 10%, while an additional 30 species coulddecrease by at least 10%, following equilibrium after a changed climate. Depending on theglobal change scenario used, 4–9 species would potentially move out of the United Statesto the north. Nearly half of the species assessed (36 out of 80) showed the potential forthe ecological optima to shift at least 100 km to the north, including seven that could move.250 km. Given these potential future distributions, actual species redistributions will becontrolled by migration rates possible through fragmented landscapes.

Key words: climate change; envelope analysis; forest inventory; geographic information systems(GIS); global change; landscape ecology; predictive vegetation mapping; regression tree analysis(RTA); species–environment relationships; tree species distribution; tree species migration; tree speciesranges.

INTRODUCTION

The buildup of anthropogenic greenhouse gases isrelated to a general warming trend on the planet, andcurrent estimates from global circulation models(GCMs) predict 18–4.58C temperature increases as aresult of doubling atmospheric CO2, probably by theend of the next century (Kattenberg et al. 1996). Thistrend suggests a potential to trigger major changes inthe earth’s living systems, including temperate forests.Many research lines, including this one, are devoted topredicting these potential major changes in the earth’sbiota, especially with respect to the potential necessarymigration of plant species.

Studies of the Holocene, when there was climaticwarming at a lower rate than is projected by currentGCMs, show the following points with respect to spe-cies migration (Jacobson et al. 1987, Delcourt and Del-

Manuscript received 24 February 1997; revised 14 Novem-ber 1997; accepted 15 November 1997; final version received10 April 1998.

court 1988, Davis and Zabinski 1992, Webb and Bart-lein 1992, Malanson 1993): (1) Species did shift theirgeographical ranges, generally northward. (2) Species’responses were individualistic—the rates and directionof migration differed among taxa and species assem-blages did not remain the same. (3) Species respondedto climate change in an equilibrium manner, and, at thecontinental scale of evaluation, competition and dis-persal mechanisms did not seem to play a large role inthe responses of species. Migrations were occurringover thousands of years and over a relatively uninter-rupted landscape. Under current greenhouse warmingscenarios, however, the climate is projected to changeat a faster rate. In today’s fragmented landscapes, com-petition, dispersal ability, and nonequilibrium re-sponses may be critical in the final species assemblage,and individual-species models are necessary to accountfor the expected individualistic responses.

With advances in the use of geographic informationsystems (GIS) and the increased abundance of dataavailable for landscapes, several techniques can now

466 LOUIS R. IVERSON AND ANANTHA M. PRASAD Ecological MonographsVol. 68, No. 4

be used to predict the geographic distribution of veg-etation from mapped environmental variables (Franklin1995). For continuous data, these include regressionmodels, general linear models, general additive models,and regression tree models. Especially important driv-ers to vegetation models are climatic, edaphic, and to-pographic variables. At local scales, modeling of veg-etation pattern depends largely on local variations oftopography and geomorphology (e.g., Reed et al. 1993,Ertsen et al. 1995, Iverson et al. 1996b, 1997). At re-gional scales, overall vegetation patterns have been as-sumed to depend more on general climatic patterns(e.g., Booth 1990, Prentice et al. 1992, Woodward1992, Box et al. 1993, Neilson 1995). To some extent,we evaluate both in this study.

Various GCMs are being built to simulate the futureclimate system, which can then be used to predict pos-sible changes in the earth’s forest biota. Several ap-proaches are being taken at a number of scales andlocations. Because such models cannot be truly vali-dated (Rastetter 1996), multiple avenues of researchare encouraged, with the hope of eventual convergence.Much discussion centers on the various approaches andthe comparisons among them (e.g., McGuire et al.1993, Hobbs 1994, Melillo et al. 1995). Several modelsassess changes in global biomes (e.g., Prentice et al.1992, Woodward 1992, Neilson and Marks 1994), in-cluding recent models that couple vegetation effectsdirectly into the GCMs so that feedbacks are contin-uously incorporated into the outcome (e.g., Foley et al.1996), while others assess a selected number of speciesover a regional scale (e.g., Bonan and Sirois 1992,Mackey and Sims 1993, Burton and Cumming 1995,Dyer 1995, Hughes et al. 1996, Starfield and Chapin1996). For example, Huntley et al. (1995) used a three-way climate response surface to model the present andfuture distributions of eight species in Europe.

However, there have been few or no studies thatmodel the entire suite of major tree species, individ-ually, across their continental ranges. Such an approachis the only way to begin to identify the possible newoutcomes in species assemblages after climate change.Most of the previously cited studies also deal only withrange shifts, not with projected shifts in the importanceof species. In this study, we empirically model each of80 species across the eastern United States, so that thedifferent habitat requirements among species are ac-commodated by the models. We statistically evaluate,using regression tree analysis (RTA), the relationshipof 33 environmental variables to tree importance valuesin the eastern United States, and then use the derivedrelationships to predict their present and potential fu-ture ranges and importance values. Our approach isunique in the sense that it is a multiple-species effortat a continental scale.

METHODS

Database generationData were extracted from several sources for land

east of the 100th meridian (Fig. 1). The county was

chosen as the mapping unit because it is the reportingunit for many sources of data and, for the most partexcept for some northern counties, has roughly thesame area across the study region. The Modifiable AreaUnit Problem, where spatial aggregations can lead toproblems with scaling and zoning (Jelinski and Wu1996), was an issue, though it was minimized by usingpercentages and area-weighted averages as input vari-ables. We evaluated .100 environmental/land use/so-cioeconomic variables for each of nearly 2500 countiesin the East. Because of missing tree information onsome counties, only 2124 counties were finallymapped. To reduce autocorrelation and enable betterinterpretation, the final number of variables was re-duced to 33 (Table 1) by (a) removing highly redundantvariables as deduced via correlation analysis; (b) drop-ping variables selected by experts as excessively dif-ficult to interpret; and (c) scoring 65 variables for theirvalue in an earlier RTA run on all species and droppingthe variables of lower importance. To evaluate overallimportance of each of the resulting 33 variables, theywere scored for frequency of occurrence and rank-based weight in the RTA outputs for the 80 species(Table 1).

Tree ranges and importance values.—The USDAForest Service periodically determines the extent, con-dition, and volume of timber, growth, and removals ofthe Nation’s forest land by the work of six Forest Ser-vice Forest Inventory and Analysis (FIA) units. FourFIA units produced a database of standard format calledthe Eastwide Data Base (EWDB) for the 37 states fromNorth Dakota to Texas and east. These data are storedin three record types (Hansen et al. 1992): county data,plot data, and tree data. Plot locations are not preciselylocated but county location was provided for each plot.We used the data from .100 000 plots and nearly 3 3106 trees to summarize the desired county-level infor-mation needed for this study. We evaluated 196 speciesof trees from the EWDB, but as described later, only80 species were modeled due to sample restrictions.

We first summarized the information for individualforested plots. Tree records represented observationsof seedlings, saplings, and overstory trees. Tree spe-cies, tree status, and diameter at breast height werecombined with information on plot size to compute asingle summary record for each plot. This record con-tained, by species, estimates of the average number ofstems and total basal area of understory and overstorytrees per unit area. From this information, we generatedimportance values (IV) for each species as follows:

100BA(x) 100NS(x)IV(x) 5 1

BA(all species) NS(all species)

where x is a particular species on a plot, BA is basalarea, and NS is number of stems (summed for overstoryand understory trees). In monotypic stands, the IVwould reach the maximum of 200. The IVs were round-ed to decimal numbers with one exception. If 1 . IV

November 1998 467TREE RANGE SHIFTS AFTER CLIMATE CHANGE

FIG. 1. Study area is the United States east of the 100th meridian. Horizontal lines correspond to the row number forthe study area, in increments of 10 km. For example, row 50 corresponds to 500 km above the southern limit of the studyarea. Latitude and longitude (in degrees) are also given in dashed lines.

. 0, it was assigned to 1, since rounding would havefalsely turned species-present counties to species-ab-sent counties.

For each species, average IV per county was cal-culated by aggregating plot-level information. Thesevalues were linked with a county coverage of the Unit-ed States (ESRI 1992) for mapping into density slicesof IV. With these maps, biogeographical characteristics(such as absolute and optimum range) of the speciescan be visualized. Further details on the methodology,along with a table summarizing the number of plots,number of species, percentage of forest, and dates ofthe inventories by state, are in Iverson et al. (1996a).

Climatic factors.—Monthly means (averaged from1948 to 1987) of precipitation, temperature, and po-tential evapotranspiration for the current climate wereextracted from a database generated by the USEPA(1993). The data had been interpolated into 10 3 10km grid cells for the conterminous United States. Fromthese data, we extracted January and July temperatures,calculated annual means, and derived two attributesbased on their physiological importance to tree growthfor this region: July–August ratio of precipitation topotential evapotranspiration (PET) (the time most

prone to drought stress in the eastern United States),and May–September (i.e., growing season) mean tem-perature. The data were then transformed to countyaverages via area-weighted averaging.

Two scenarios of future monthly temperature andprecipitation under an equilibrium state of twice thepresent levels of atmospheric CO2 were used for pre-dictions of potential future species distributions: theGeophysical Fluid Dynamics Laboratory (GFDL)(Wetherald and Manabe 1988) and Goddard Instituteof Space Studies (GISS) (Hansen et al. 1988) models,which depict a divergent set of possible outcomes.These GCM output data were prepared by USEPA(1993) into future equilibrium estimates, by 10 3 10km grid, for monthly precipitation, temperature, andpotential evapotranspiration. These variables, alongwith the derived variables mentioned above, were sub-stituted into the database for predicting potential futurespecies distributions.

Soil factors.—The State Soil Geographic Data Base(STATSGO) was developed by the USDA Soil Con-servation Service (now Natural Resource ConservationService) to help achieve their mandate to collect, store,maintain, and distribute soil survey information for


TABLE 1. County environmental and land-use variables reported for each county and used inthe RTA process. Frequency indicates the number of times the variable appeared in the RTAmodels, while rank-weighted score is the sum of (1/rank in the RTA model) for all 80 species.

Abbreviation Variable Frequency

Rank-weightscore

Climatic factorsAVGTJANTJULTPPTPETMAYSEPTJARPPET

Mean annual temperature (8C)Mean January temperature (8C)Mean July temperature (8C)Annual precipitation (mm)Potential evapotranspiration (mm/mo)Mean May–September temperature (8C)July–August ratio of precipitation to PET

47115

4479703027

12.034.8

9.612.917.6

9.16.1

Soil factorsTAWCCECPHPERMCLAYBDKFFACTOMROCKFRAG

Total available water capacity (cm, to 152 cm)Cation exchange capacitySoil pHSoil permeability rate (cm/h)Percent clay (,0.002 mm size)Soil bulk density (g/m2)Soil erodibility factor, rock fragments freeOrganic matter content (% by mass)Percent mass of rock fragments 8–25 cm

475845756650926054

6.98.58.9

15.412.4

5.723.213.0

9.7NO10NO200ROCKDEPSLOPEORDALFISOLINCEPTSLMOLLISOLSPODOSOL

Percent passing sieve Number 10 (coarse)Percent passing sieve Number 200 (fine)Depth to bedrock (cm)Soil slope (%)Potential soil productivity (m3 of timber/ha)Alfisol (%)Inceptisol (%)Mollisol (%)Spodosol (%)

434629576134143013

6.510.5

2.815.712.2

9.31.97.92.6

Land use/cover factorsFORST.LNDCROPSGRAZE.PSTDIST.LND

Forest land (%)Cropland (%)Grazing pasture land (%)Disturbed land (%)

54465435

6.713.9

8.64.0

ElevationMAX.ELVMIN.ELVELV.CV

Maximum elevation (m)Minimum elevation (m)Elevation coefficient of variation

674646

18.36.68.0

Landscape patternED Edge density (m/ha) 50 6.4

U.S. lands. STATSGO data contain physical and chem-ical soil properties for ;18 000 soil series recognizedin the Nation (Soil Conservation Service 1991).STATSGO maps were compiled by generalizing moredetailed soil-survey maps into soil associations in ascale (1:250 000) more appropriate for regional anal-ysis. We selected 14 soil variables related to tree spe-cies’ habitat (Table 1). Weighted averages by depth andby area were calculated for county estimates of the soilvariables, as detailed in Iverson et al. (1996a). Addi-tional soil information was obtained from the GEO-ECOLOGY databases (Olson et al. 1980), includingpercentage of the county in each of four soil orders(Table 1).

Land use/cover factors.—GEOECOLOGY (Olson etal. 1980) data were used for estimations of percentageforest, crop, grazing/pasture, and disturbed land (Table1). These estimates originated from the USDA Soil

Conservation Service’s National Resources Inventoryfor 1977.

Elevation.—Maximum, minimum, and variation ofelevation were derived for each county from 1:250 000U.S. Geological Survey (USGS) Digital ElevationModel (DEM) files obtained from the USGS internetsite (U.S. Geological Survey 1987).

Landscape pattern.—The 1-km AVHRR forest covermap (U.S. Forest Service 1993) was used to generatestatistics on forest-cover pattern by county. Severallandscape pattern indices were calculated usingFRAGSTATS (McGarigal and Marks 1995), but onlyedge density was used in the final analysis.

Regression tree analysis

We use regression tree analysis (RTA, also knownas classification and regression trees, or CART), to de-cipher the relationships between environmental factors


and species distribution. The methods can either predictclasses (classification trees) or average values (regres-sion trees), depending on the nature of the responsevariable. Developed by Breiman et al. (1984), and usedfirst in the medical field, the methods have only beenused in ecological studies since about 1987 (Verbyla1987). Not incidentally, their use has coincided withthe use of geographic information systems, which allowmodel outputs to be readily mapped across landscapes.Since then, they have been used primarily in classifi-cation (e.g., Borchert et al. 1989, Lees and Ritman1991, Moore et al. 1991, Baker 1993, Lynn et al. 1995),though examples of the ecological use of regressiontrees are found in Davis et al. (1990) and Michaelsenet al. (1994).

RTA is a recursive data partitioning algorithm thatinitially splits the data set into two subsets based on asingle best predictor variable (the variable that mini-mizes the variance in the response). It then does thesame on each of the subsets and so on recursively. Theoutput is a tree with branches and terminal nodes. Thepredicted value at each terminal node is the average atthat node, which is relatively homogeneous (Chambersand Hastie 1993). RTA is therefore much more flexiblein uncovering structure in data that have variables thatmay be hierarchical, nonlinear, nonadditive, or cate-gorical in nature. RTA is rapidly gaining popularity asa means of devising prediction rules for rapid and re-peated evaluation, as a screening method for variables,as a diagnostic technique to assess the adequacy oflinear models, and for summarizing large multivariatedata sets (Clark and Pregibon 1992).

There are several key advantages to using RTA inour application, which covers such a wide spatial do-main, over classical statistical methods (Verbyla 1987,Clark and Pregibon 1992, Breiman et al. 1984, Mi-chaelsen et al. 1994). First, RTA is adept at capturingnonadditive behavior, where relationships between theresponse variable and some predictor variables are con-ditional on the values of other predictors. RTA dealswith interaction effects by subsetting data withoutspecifying the interaction terms in advance of the sta-tistical analysis (as is necessary in multiple linear re-gression). For example, in our study, the factors as-sociated with the northern range limits for a speciesmay be quite different from the factors regulating thesouthern limit of the species. This advantage allows,in effect, a stratification of the country so that somevariables may be most related to the IV of species Afor a particular region of the country, but a differentset of variables may drive its importance elsewhere.Second, numerical and categorical variables can easilybe used together and interpreted, because RTA essen-tially converts continuous data into two categories ateach node. The outcome is a set of step functions thatprovides a better capturing of nonlinear relationships,while also providing a reasonable solution for linearrelationships. Last, the variables that operate at large

scales are used for splitting criteria early in the model,while variables that influence the response variable lo-cally are used in decision rules near the terminal nodes(Moore et al. 1991). Thus we could expect that broadclimatic patterns are captured higher up on the tree,while more local effects (soil, elevation, etc.) determinemore local distributional variations. It should, however,be recognized that since our data set is aggregated toa county-level scale, RTA cannot capture the environ-mental drivers that operate on species at a very finescale (e.g., individual slopes or valley bottoms).

There are limitations of RTA, however. While RTAmay be the most appropriate tool for analyzing datasets with many possibly interacting predictor variablesand for separating macro-level and micro-level effectson the response, the response ideally should be ap-proximately normal, since least squares and the groupmean are used in deriving the best split (Clark andPregibon 1992). Even though the IV response in ourexample more closely resembles a Poisson distributiondue to the high frequency of zeros, the effect of longtails is much more confined than in a linear regressionmodel as the regression is local (B. D. Ripley, personalcommunication). Also, since we are mainly interestedin variables that are driving the distribution geograph-ically, we believe this is a reasonable approach; trans-forming the response would complicate the interpret-ability of the results. We are taking a further step andpredicting the effect of a changed climate based on themodel. While there is danger of extrapolating the re-sults beyond the model’s predictive ability for somespecies, it does provide a reasonable estimate of thespecies’ potential migration under changed climaticconditions, whose preferences we know a priori.

Prediction of current species importance values.—Regression trees were generated in S-PLUS (StatisticalSciences 1993) for each tree species. Species impor-tance value (IV, based on basal area and number ofstems) was the response variable, with the 33 predictorvariables (see Table 1). The regression trees were gen-erated on a split data set—a random selection of 80%of the data—so that 20% of the data remained for val-idation efforts. RTA models were generated after prun-ing the full tree to 12–20 nodes, depending on the rateof change of deviance explained. When the additionaldeviance explained was small for added nodes, the RTAmodel building was stopped. The resulting model wasused to generate predictions of IV of each species foreach county. While this pruning effort carried somesubjectivity, it is the most reasonable approach whileevaluating 80 tree species; picking nodes through crossvalidation by resampling is not without some majorlimitations (Venables and Ripley 1994).

In addition to the RTA models using all the 33 pre-dictors, we built models using only the seven climaticvariables (Table 1). This analysis was done to compareprediction effectiveness with that using edaphic andland use variables. For several species, maps were also


generated that depict the counties that fall along par-ticular branches of the RTA tree. The environmentalvariables most responsible for the predicted IV areshown geographically, which is a major strength ofRTA.

The response predicted by RTA for zero values ofIV was almost always a fraction ,1. Through testingacross all species, we determined that predicted IVscores ,1.0 were essentially zero and were set as such.Predicted IV scores between 1.0 and 3.0 were classed‘‘fuzzy’’ in that we were not confident in the predictionof these low IV scores. The predictions of IV classeswere then output to Arc/Info for mapping, using Unixtools and Arc/Info’s Arc Macro Language (ESRI 1993).

Validation/verification of the RTA outputs.—Toevaluate the model outputs, a comparison between pre-dicted current and actual (FIA) distributions was madeusing correlation, verification, and validation process-es. Scattergrams and correlations of IV were calculatedfor each species. Verification, to assess how well themodels recreate the distributions with the entire dataset, was performed by assessing for each species thenumber and proportion of counties found in (a) boththe actual and predicted scenarios, (b) only the actual,and (c) only the predicted. This allowed calculation ofthe percent correct assignment ([a/(a 1 b 1 c)] 3 100,hereafter referred to as classification accuracy), andalso the calculation of omission errors, where the modelfailed to predict the presence of the target species eventhough it was recorded in the FIA data ([a/(a 1 b)] 3100, hereafter referred to as omission accuracy). Wewere especially interested in the omission accuracy.The error was deemed less serious if the model pre-dicted a county to contain a species that was not re-corded because it is quite possible that the FIA sam-pling may have missed it. FIA attempts to place a plotfor about every 2000 ha of forest land; uncommon orhabitat-specific species could easily be missed by theplots. A validation comparison was made by using arandomly selected portion of the counties (80%) to pre-dict the IV of the remaining 20% (n 5 424). Similarcalculations of error were then made on this 20% dataset.

Prediction of future species importance values.—Once the regression trees were generated, they wereused to generate not only predictive maps of currentdistributions, but also potential future distributions un-der a scenario of a changed climate. We did this byreplacing the climate-related variables in our predictorvariable set with those based on the GFDL and GISSmodels. The replaced variables were (see Table 1 forthe description): MAYSEPT, JARPPET, JANT, JULT,AVGT, PET, and PPT. The previously established re-gression trees then were used with the new predictivevariables, and the data output to Arc/Info as before.

Assessment of potential species changes.—We de-vised several metrics to assess the potential speciesshifts as a result of the modeled changes in IV of each

species. The first metric, change in area, is the differ-ence, in percent, between the predicted current arealextent and the predicted potential future areal extentunder the two climate change scenarios. So, a score of100 means no change, scores ,100 represent a con-traction of the range, and scores .100 are an expansionof the species range. Because of relatively large un-certainty in the ‘‘fuzzy’’ class of predicted IV (1.0–3.0), we took a conservative approach and calculatedthis metric only for counties with IV . 3.0.

The second metric, change in weighted average ofIV, is a weighted (by area of counties) average of IVfor all counties calculated for predicted current vs. thetwo climate change scenarios; as above, a score below100 indicates the overall importance of the species willdecline (though areal extent may or may not decline).In this case, we included the ‘‘fuzzy’’ class of predictedIV in the calculation, since county areas with low IVwould not contribute much to the overall score.

The third metric is the change in northern and south-ern range limits. Because we were not interested inspurious shifts out on the tails of the distribution, westatistically identified the range limits for predicted cur-rent and future ranges. The country was subdividedinto a grid cell network of 10 3 10 km cells; a totalof 277 rows, or latitudinal strips, was generated (Fig.1). The relative IVs were calculated for each row (totalIV/total area in row), and box plots were generated, inS-PLUS (Statistical Sciences 1993), for the summedIV for each species. From this distribution, the IV ofthe first quartile was identified; the first and last rowwhere this IV was exceeded was deemed the northernand southern limit of the distribution, respectively.Shifts in the northern or southern limits were simplycalculated by taking the difference between current andprojected future limits.

The fourth metric is the change in ecological opti-mum. The mean of the interquartile range was deemedthe ‘‘optimum’’ latitude range for the species. Both ofthe latter metrics used only those areas with predictedIV values .3.0; the ‘‘fuzzy’’ class was not included.

Model assumptions and limitations

Here we attempt to state our assumptions explicitlyso that progress can continue as uncertainties are re-duced. First, uncertainty remains in projected climatechange under a doubled CO2 scenario, especially as itplays out spatially across the continent. The impactsof this uncertainty on our results are reduced by usingtwo scenarios, but changes in estimates of future cli-mate will continue to occur. Second, any time multipleGIS layers from disparate sources and scales are over-laid, errors will propagate through the data (Goodchildand Gopal 1989). This impact is minimized in thisstudy by using a large sampling unit, the county, asthe common spatial unit. On the other hand, some coun-ties are very diverse, and some important ecologicalfactors could be averaged out at this scale. For example,


FIG. 2. Regression tree for quaking aspen (Populus tremuloides), using a nonuniform tree of 12 nodes. If the rule at thetop of a branch is true, follow the left branch; if false, follow the right branch. The numbers at the termination points referto predicted IV for that particular branch of the model, which is the average IV at that node. The length of the vertical linebelow each true–false split corresponds to the importance of the variable.

small zones of high elevation, bottomlands, or uniquesoils, which harbor specialized flora, may be lost in theaveraging process. There is also error associated withthe FIA sampling of trees; occasionally species that doin fact reside in the county will be missed by the FIAsampling plots. Third, the method described here doesnot account for changes in physiological and species-interaction effects in the model outputs. Therefore,there is no way to assess changes in competition amongthe ‘‘new’’ species mix, nor is there a way to accountfor whatever gains in water-use efficiency may accom-pany elevated CO2 (Neilson 1995). Fourth, in a criti-cism of model-based assessments of climate changeeffects on forests, Loehle and LeBlanc (1996) note thatmany forest simulation models assume that tree speciesoccur in all environments where it is possible for themto survive, and that they cannot survive outside theclimatic conditions of their current range (fundamentalvs. realized niche). The RTA models here reduce thisproblem by considering a wide range of variables andonly trying to evaluate potential range changes due toclimate change. These models assume equilibrium con-ditions, and that there are no barriers to migration.Finally, RTA does have limitations, and spurious ornoncausative relationships will appear, especially whenRTA methodology is applied to 80 species without fine-tuning for individual species preferences. Improvement

of models may be possible for many of the species, ifindividual characteristics and spatial trends are takeninto account.

RESULTS AND DISCUSSION

Regression trees

Regression trees were generated for each species; anexample for quaking aspen (Populus tremuloides) isshown in Fig. 2. At the terminal ends of each branch,a predicted IV is presented. In this example, the rela-tively large importance of mean annual temperature isapparent, for the distance between nodes is in propor-tion to the variance explained by the variable. The high-est predicted importance occurs on the left branch,where a maximum IV of 91 was obtained if the meanannual temperature was ,4.48C, the coefficient of vari-ation of elevation in the county was ,17.8% (i.e., thecounty is homogeneously rather flat), the average po-tential evapotranspiration was ,60.0 mm/mo, and.92.3% of soil can pass a number 10 sieve (i.e., thesoils are primarily sands and finer material). Only eightcounties matched these criteria. On the other branch,the lowest IV, 0.14, was found for counties that had anaverage temperature .8.128C; this condition occurs in1704 counties. Because we used a cutoff of 1.0 for aspecies to be considered present, none of the 1704counties was predicted to have quaking aspen.


→

FIG. 3. Mapped environmental variables most related to the distribution of three species: (a) southern red oak (Quercusfalcata var. falcata), (b) Virginia pine (Pinus virginiana), and (c) American beech (Fagus grandifolia). The indented legendshows the hierarchical structure of the data; see Table 1 for definitions of the variables in the first column. The second columnis the number of counties that the equation pertains to, and the third column is the average IV for the species for thosecounties. An asterisk indicates that the branch is a terminal node for the tree.

The distribution of variables most responsible for thepredicted IV for three species shows the additional val-ue of the RTA in that different variables operate incontrolling the distribution (Fig. 3). For example, withsouthern red oak (Quercus falcata var. falcata), thereis essentially no representation of the species in 1222counties with an average January temperature,0.348C. In this example, the maximum IV (8.0) isfound in 94 counties that have: a January temperature.0.348C, an average slope in the county exceeding2.3%, average total water-holding capacity exceeding19.5 cm, and a maximum elevation .160 m (Fig. 3).The northern band of southern red oak presence isfound where January temperature is ,3.28C, indicatingthat temperature may be the key factor regulating thenorthern boundary of the species.

For Virginia pine (Pinus virginiana), pH is the firstcontrolling variable, with the species found only wherepH , 4.35, except for 26 counties with higher pH butwith average slopes exceeding 21% (Fig. 3). If the pHcriterion is satisfied and January temperatures ,0.78C,then the species is found (IV . 1) only in countieswhere slopes . 20.8%. This could be a situation wherethe species can survive colder winters in protectedsteep valleys or on more southerly exposures. Virginiapine prefers warmer temperatures in January (.0.78C)but relatively cooler temperatures in July (,25.18C),unless the maximum elevation in the county exceeds889 m (Fig. 3). The importance of the species is thusdistributed spatially by soil and climate factors.

American beech (Fagus grandifolia) reaches itsmaximum IVs when potential evapotranspiration (PET)is ,46.5 mm/mo and when maximum elevation ex-ceeds 663 m, conditions met by 76 counties in thenorthern Appalachians. For lower elevations in thenorth, beech can be prevalent if precipitation exceeds850 mm. For counties with higher PET, beech is mostlyfound in the high sloping regions of the southern Ap-palachians (Fig. 3).

As can be seen from the legend of Fig. 3, a com-bination of climate and edaphic variables was neces-sary to achieve the best RTA models. This pattern heldtrue for nearly all species. Scores of frequency of oc-currence in RTA outputs for all 80 species showedJanuary temperature to be the most influential variableoverall. It appeared 115 times at RTA nodes summedacross all species (a variable can appear more than oncein any given RTA model), and had a rank-weightedscore (sum of 1/rank for all species, with rank beingthe node position of the variable in the model) of 34.8(Table 1). Next most important was the K factor, or

wind erodibility of the soil. It relates to texture, beinghighest with silty soils. By weighted score, the nextmost important variables, in order, were maximum el-evation, potential evapotranspiration, slope, soil per-meability, percent cropland, percent organic matter,precipitation, and percent clay (Table 1). Edaphic fac-tors thus occupy 6 of the top 10 positions in relativeimportance. Most work of this type has ignored edaphicfactors (Loehle and LeBlanc 1996).

Prediction of current species distributions

Tree models work best if sufficient samples are avail-able (Moore et al. 1991, Clark and Pregibon 1992).Analysis of the verification data set revealed there wasa significant correlation between the number of coun-ties for which the target species was recorded in theFIA data and the classification accuracy (Table 2).Those species with larger samples tend to have in-creased model success based on correlations betweenpredicted and actual IV. Rarer species also tend to havemore specific habitat requirements that would be dif-ficult for the coarse, county-level data to capture. Somespecies that are not so rare, with more specific habitatrequirements, also fall into this category. It should beborne in mind that the model was not fine-tuned forindividual species preferences, as we were interestedin macro-scale comparison among several species.Also, the added benefit of the analysis is that it high-lights species that require more individual attentionand/or finer resolution data. In light of the above con-siderations, only those species with a recorded mini-mum IV of 3.0 in each of at least 100 counties (out ofa possible 2124, or 4.7% of the counties) are includedin the reporting here. A total of 80 species matchesthis criterion (Table 2).

Predicted current distributions match FIA data rea-sonably well for most species. For example, for quak-ing aspen (Populus tremuloides, Fig. 4), the FIA datarecorded the species in a total of 338 counties. Totalclassification accuracy was 71% on the entire data set,and 67% on the 20% validation data set. If only con-sidering error of omission (the RTA model predictedno target species when the FIA sampling recorded it),the accuracies are 81 and 77%, respectively (Table 2).The correlation of IVs (for counties with IV at least3.0) between predicted and FIA data is 0.81. The spe-cies is a generalist, being the most widely dispersedspecies in the United States, and can grow on a widevariety of soil types (Elias 1980, Perala 1990). Thecounty level of resolution is, therefore, adequate tocapture the major environmental variables driving its


TABLE 2. Verification/validation table for species known from FIA plot data to be present in .100 counties, and with anIV of at least 3.0.

Species

Verification

PF PP PB CA (%) EOA (%) CORR

Validation

PF PP PB CA (%) EOA (%)

Abies balsameaAcer negundoAcer pensylvanicumAcer rubrumAcer saccharinum

9170

3453

161

45250

12339

91

104317

551341

124

65.843.054.577.433.0

92.065.161.896.243.5

0.870.700.720.750.68

531

89

35

661

86918

187015

26220

62.143.248.477.127.4

78.369.365.296.736.4

Acer saccharumBetula alleghaniensisBetula lentaBetula nigraBetula papyrifera

10666553147

23813162018

731868832

154

68.052.155.338.670.3

87.356.661.550.876.6

0.610.770.730.190.61

311715

99

4127

1413

1542118

132

68.152.545.0

4.259.3

83.255.354.510.078.0

Carpinus carolinianaCarya cordiformisCarya glabraCarya ovataCarya sp.

357104

7076

188

2049

9171196

95139

88234761

20.147.652.748.666.5

21.057.255.775.580.2

0.300.260.520.630.64

6329192342

912

62128

20191139

149

21.731.730.647.068.0

24.139.636.762.978.0

Carya tomentosaCeltis laevigataCeltis occidentalisCercis canadensisCornus florida

476872

108176

1812

22160

109

8452

351108776

56.439.454.539.173.1

64.143.383.050.081.5

0.550.800.730.430.56

1014273444

1115451730

14146712

139

40.032.648.219.065.3

58.350.071.326.176.0

Crataegus sp.Diospyros virginiana

153161

8613

16858

41.325.0

52.326.5

0.590.55

3036

277

416

41.812.2

57.714.3

Fagus grandifoliaFraxinus americanaFraxinus nigraFraxinus pennsylvanicaFraxinus sp.

236206

56139

90

87150

18389

35

308612

94452108

48.863.256.046.146.4

56.674.862.776.554.5

0.630.650.570.810.36

3847123814

3431

68214

68123

157813

48.661.245.539.431.7

64.272.455.667.248.1

Gleditsia triacanthosIlex opacaJuglans nigraJuniperus virginianaLiquidambar styraciflua

12688

231120

85

528

102322

92

10961

205287689

38.038.938.139.479.6

46.440.947.070.589.0

0.520.850.600.570.75

2720372216

164

287321

191

5060

120

30.64.0

43.538.776.4

41.34.8

57.573.288.2

Liriodendron tulipiferaMaclura pomiferaMagnolia virginianaMorus rubraNyssa aquatica

1727254

10059

144382038

9

56395

1245236

64.146.362.627.434.6

76.656.969.734.237.9

0.650.490.580.770.84

26161122

8

311910

55

1092015

85

65.736.441.722.927.8

80.755.657.726.738.5

Nyssa bifloraNyssa sylvaticaOstrya virginianaOxydendrum arboreumPinus echinata

4496

36487

154

36201

634938

193675130233321

70.769.423.363.162.6

81.487.526.372.867.6

0.770.400.570.660.73

824731827

832211724

29118

334267

64.467.826.054.556.8

78.483.131.170.071.3

Pinus elliottiiPinus palustrisPinus resinosaPinus strobusPinus taeda

393655

13268

4560295986

176169

78177591

67.763.848.148.179.3

81.982.458.657.389.7

0.860.710.670.720.75

858

294

498

1122

33341734

108

73.370.851.545.980.6

80.587.268.054.096.4

Pinus virginianaPlatanus occidentalisPopulus deltoidesPopulus grandidentataPopulus tremuloides

83159147107

57

10129

1135

41

21276

14270

240

53.528.835.338.571.0

71.932.349.139.580.8

0.710.380.820.710.81

3236422716

181727

510

3112221252

38.318.524.227.366.7

49.225.034.430.876.5

Prunus serotinaQuercus albaQuercus coccineaQuercus falcata

var. falcata

392130134

89

208411117

93

5571130

249

443

48.167.649.8

70.9

58.789.765.0

83.3

0.730.590.47

0.41

593436

25

8071

9

19

129199

30

77

48.165.540.0

63.6

68.685.445.5

75.5Quercus falcata

var. pagodaefoliaQuercus laurifoliaQuercus macrocarpaQuercus marilandicaQuercus muehlenbergii

3554957191

2122

1091912

77166211

7859

57.968.650.846.436.4

68.875.569.052.339.3

0.590.710.720.760.59

186

181521

106

471110

4334417

8

12.573.340.439.520.5

18.284.671.053.127.6

Quercus nigraQuercus palustrisQuercus phellosquercus prinusQuercus rubra

8055

128161161

45161627

256

4214665

265857

77.139.331.158.567.3

84.045.533.762.284.2

0.620.440.750.680.64

2113192235

1031

1445

6669

60179

68.027.331.062.569.1

75.931.632.173.283.6


TABLE 2. Continued.

Species

Verification

PF PP PB CA (%) EOA (%) CORR

Validation

PF PP PB CA (%) EOA (%)

Quercus stellataQuercus velutinaRobinia pseudoacaciaSalix nigraSalix sp.

89133182164

75

312311

2613

9

410692125

4228

50.660.937.519.225.0

82.283.940.720.427.2

0.860.760.330.650.56

829432915

48752018

7

90143

2495

61.657.927.616.118.5

91.883.135.823.725.0

Sassafras albidumTaxodium distichumTaxodium distichum

var. nutansThuja occidentalis

18340

2126

14354

2116

33175

6875

50.444.4

61.864.1

64.465.2

76.474.3

0.490.82

0.780.77

4014

41

179

99

596

1615

50.920.7

55.260.0

59.630.0

80.093.8

Tilia americanaTsuga canadensisUlmus alataUlmus americana

1195751

207

1474882

169

258148257674

49.258.565.964.2

68.472.283.476.5

0.670.700.690.68

19151858

536

2036

642850

130

47.157.156.858.0

77.165.173.569.1

Ulmus rubraUlmus sp.

83161

2847

44565

54.827.9

84.328.8

0.420.54

2626

6212

6815

43.628.3

72.336.6

Mean 52.24 63.93 0.64 45.59 59.15

Note: PF 5 number of counties present in FIA data only; PP 5 number of counties present in predictive model only; PB5 number of counties present in both; CA 5 classification accuracy (PB/[PB 1 PF 1 PP] 3 100); EOA 5 error of omissionaccuracy (PB/[PB + PF] 3 100); CORR 5 correlation between actual and predicted IV.

distribution, and conditions represented by county av-erages are adequate to model the species.

For flowering dogwood (Cornus florida) (Fig. 5), themodel has classification accuracies of 73 and 65% forthe full and 20% data set, and omission accuracies of82 and 76%, respectively (Table 2). This species toois a generalist, but favors lighter soils with good drain-age (McLemore 1990).

For all 80 species, classification accuracies rangedwidely, from 19% for black willow (Salix nigra) to80% for sweetgum (Liquidambar styraciflua); the av-erage was 52% in the full verification data set (Table2). The omission accuracies were somewhat better,ranging from 20 to 96%, with an average accuracyof 63%. In general, the more specialized the speciesis with respect to edaphic conditions, the less ac-curacy in the RTA model predictions. In fact, in near-ly all situations where the accuracy was less than;40%, the species is known to be a habitat specialist,preferring bottomlands, alkaline areas, or disturbedsystems. These are not attributes readily captured inthe coarse-resolution data set used here. Severalpoorly classified species, including Morus rubra,Nyssa aquatica, Platanus occidentalis, Populusgrandidentata, Salix nigra, and Taxodium distichum,are all bottomland species. Bottomland habitatswould occupy a small minority of most counties, andcounty-resolution data would not be expected to con-sistently capture the appropriate information in theRTA process. Lower classification accuracies werealso apparent in taxa that were reported at the genuslevel (for example, Salix sp. and Ulmus sp., Table2). Because these taxa include a number of differentspecies, with varying habitat requirements, the RTAwould be less likely to uncover precise trends. An

exception is Carya sp., in which several species ofhickory have similar habitat requirements.

Species that showed high classification and omissionaccuracies (greater than ;70% classification accuracyor 80% omission accuracy), on the other hand, tendedto be capable of occupying a large variety of habitatswithin their range of distribution. For example, Acerrubrum is one of the most common species in easternNorth America (Elias 1980, Golet et al. 1993). Liquid-ambar styraciflua, Nyssa sylvatica, Pinus taeda, Quer-cus alba, and Q. rubra are also generalist species, oc-cupying a wide variety of sites within their ranges (Eli-as 1980). County-level environmental data are muchmore likely to represent conditions for these speciesthan the specialist species.

When the RTA analysis using only climatic variablesas inputs was compared to the results described above,we found the inclusion of edaphic and land-use vari-ables to be vitally important for many species. Forexample the climate-only model failed for Taxodiumdistichum, whose distribution is mainly driven by el-evation and permeability. For 70 of the 80 species, thecorrelation between actual and predicted IV was im-proved with the addition of the extra variables. In ad-dition, when using only climate variables, one couldcause the resultant model to be overly sensitive to cli-mate under changed climate scenarios. These resultsindicate the importance of assessing edaphic con-straints in the prediction of species shifts followingclimate change.

Prediction of potential future speciesimportance and area

Projected potential species distributions, followingequilibrium of predicted climate changes, show major


FIG. 4. Example model outputs for quaking aspen (Populus tremuloides), including (a) actual county importance value(IV) as calculated from FIA data; (b) predicted current IV from the RTA model; (c) predicted potential future IV after climatechange according to the GISS GCM; (d) predicted potential future IV after climate change according to the GFDL GCM;(e) difference between predicted current and predicted future IV according to the GISS model; and (f) difference betweenpredicted current and predicted future IV according to the GFDL model. This example shows a species with the potentialto contract its range and importance in the United States under climate change.


FIG. 5. Example model outputs for flowering dogwood (Cornus florida), including (a) actual county importance value(IV) as calculated from FIA data; (b) predicted current IV from the RTA model; (c) predicted potential future IV after climatechange according to the GISS GCM; (d) predicted potential future IV after climate change according to the GFDL GCM;(e) difference between predicted current and predicted future IV according to the GISS model; and (f) difference betweenpredicted current and predicted future IV according to the GFDL model. This example shows a species with the potentialto expand its range and importance under climate change.


TABLE 3. Predicted potential changes of area and weighted average importance values, after GISS and GFDL, with twofoldincreases in CO2 levels. Scores above 100 correspond to increases; those less than 100 correspond to decreases.

Species

IV score

GISS GFDL

Area score

GISS GFDL Species

IV score

GISS GFDL

Area score

GISS GFDL

Abies balsameaAcer negundoAcer pensylvanicumAcer rubrumAcer saccharinum

1101

2765

104

0101

2534

104

3100

20100

98

0100

17100

98

Pinus echinataPinus elliottiiPinus palustrisPinus resinosaPinus strobus

149148187

7135

146149170

0134

123165208

492

119162204

084

Acer saccharumBetula alleghaniensisBetula lentaBetula papyriferaCarpinus caroliniana

8275114

117

00

490

129

95342

9230

00

410

289

Pinus taedaPinus virginianaPlatanus occidentalisPopulus deltoidesPopulus grandidentata

180141157119

2

179167156210

0

158108127112

2

158106130155

0Carya cordiformisCarya glabraCarya ovataCaya sp.Carya tomentosa

10137

12094

176

88378181

358

93108787

154

86109373

315

Populus tremuloidesPrunus serotinaQuercus albaQuercus coccineaQuercus falcata

var. falcata

10529790

230

047

10264

229

137

77105206

00

6963

206

Celtis laevigata

Celtis occidentalisCercis canadensisCornus florida

329

225144113

624

228191116

278

72180118

1031

65252124

Quercus falcatavar. pagodaefolia

Quercus laurifoliaQuercus macrocarpaQuercus marilandica

151

14961

180

148

14855

955

159

12048

184

154

12149

851Crataegus sp.Diospyros virginianaFagus grandifoliaFraxinus americanaFraxinus nigra

8196

246344

5217

116046

6330

305858

3350

215263

Quercus muehlenbergiiQuercus nigraQuercus palustrisQuercus phellosQuercus prinus

197158

32199101

212171

36216

98

233127

28167112

259127

33177112

Fraxinus pennsylvanicaFraxinus sp.Gleditsia triacanthosIlex opacaJuglans nigra

11291

16075

247

116120155

78717

10040

11952

143

100135113

52313

Quercus rubraQuercus stellataQuercus velutinaRobinia pseudoacaciaSalix nigra

68237

9494

100

52889108

87100

72105107

98100

70199

91109100

Juniperus virginianaLiquidambar styracifluaLiriodendron tulipiferaMaclura pomifera

362113

86279

579113

83526

242153

97202

325154

96370

Salix sp.Sassafras albidumTaxodium distichumTaxodium distichum

var. nutans

10580

102628

10563

102557

10070

100316

10044

100280

Magnolia virginianaMorus rubraNyssa aquaticaNyssa bifloraNyssa sylvatica

253114

79133120

242100

79133

99

181109

75100104

169100

75100104

Thuja occidentalisTilia americanaTsuga canadensisUlmus alataUlmus americana

03160

35998

02566

89899

02047

210100

02

48472100

Ostrya virginianaOxydendrum arboreum

6190

6184

2978

3075

Ulmus rubraUlmus sp.

101117

184117

87173

103173

shifts for many species (Table 3). For example, quakingaspen is projected to move north of the United Statesboundary under the GFDL scenario, and nearly so withthe GISS scenario (Fig. 4), while flowering dogwoodis projected to move to the northeast (Fig. 5). For mostspecies, area and weighted average IV shift together,though Acer rubrum is predicted to keep the same dis-tribution but decrease markedly in importance. A sum-mary of species predictions shows that approximatelyequal numbers of species would decrease or would in-crease in IV and area importance, according to thisanalysis (Fig. 6). The figure shows that 27 to 36 specieswould increase by at least 10% ($110 on the relativeIV or area scale), while 30 to 34 species would decreaseby at least 10% (#90). The remaining species wouldchange little in overall weighted importance or area.Further, most species behave similarly, regardless ofthe GCM model being used, as 28–33 species increased

in both GISS and GFDL models, while 31 species de-creased in both scenarios. Only one species, Carya ova-ta, showed a reversal of trend between GCM scenarios,where IV is projected to increase under GISS and de-crease under GFDL.

Comparisons of potential changes in distribution andabundance were made with published examples for afew species; this type of work has been done for onlya few eastern North American species. General agree-ment was found. For example, Jacobson and Dieffen-backer-Krall (1995) predicted that white pine (Pinusstrobus) would be favored, while spruce–fir would bedecreased under a climate change. We found a 35%increase predicted in area-weighted IV for white pine(even though it is likely to move substantially north-ward), but that balsam fir (Abies balsamea) would benearly eliminated from the United States as it migratednorthward (Table 3). Similarly, Flannigan and Wood-


FIG. 6. Number of species showing decreases or increasesof importance value (IV) or area for the two GCM scenariosevaluated. The total number of species is ,80 for the ‘‘Both’’categories, because those species that fall in different classesbetween GCM scenarios are not tallied.

TABLE 4. Predicted change in importance values (IV) fortree species from Vinton County, Ohio (GISS model only).

Species Change

Acer saccharumPrunus serotinaCrataegus sp.Fagus grandifoliaFraxinus americanaUlmus rubraPinus virginianaCercis canadensisPlatanus occidentalisRobinia pseudoacacia

214.5210.027.124.923.623.322.922.621.721.7

Asimina trilobaCarpinus carolinianaCarya glabraMalus sp.Populus grandidentataAcer negundoAcer rubrumCarya sp.Cornus floridaFraxinus pennsylvanica

21.321.221.121.121.0

00000

Liriodendron tulipiferaNyssa sylvaticaOstrya virginianaQuercus coccineaQuercus prinusQuercus rubraQuercus velutinaSassafras albidumUlmus americanaQuercus albaOxydendrum arboreum

0000000004.07.5

ward (1994) predicted red pine (Pinus resinosa) to mi-grate 600–800 km to the northeast, but with an increasein volume per unit area. We predict that the speciesmay migrate completely out of the United States intoCanada.

Overpeck et al. (1991) predicted the above generalpatterns for northern pines, as well as large increasesin oak abundance in the northern Great Lakes and NewEngland (also found by us for black, northern red, andwhite oak; Quercus velutina, Q. rubra, and Q. alba,respectively). They also predicted a severe northernshift for birch, especially with the GFDL over the GISSglobal change scenario. Again, these results agree withours for paper birch (Betula papyrifera). Finally, Over-peck et al. (1991) predicted a large northward expan-sion for southern pines, as exemplified by loblolly pine(Pinus taeda). Joyce et al. (1990) also report a predictedexpansion for loblolly. Our results concur by predictingan 80% increase in area-weighted IV, and a 58% in-crease in range area (Table 3).

Predicted changes in potential speciesimportance by county

Predicted changes in major species importance forany county can be generated with the RTA outputs(Table 4). For example, in Vinton County in southernOhio, the GISS climate projections show that, among15 species projected to decline, Acer saccharum andPrunus serotina would decline sharply, while Oxyden-drum arboreum and Quercus alba would increase inimportance. An additional 14 species are projected tochange very little (less than one IV unit, on average)under the changed climate.

Predicted changes in potential species boundariesand optima

Estimates of southern limit, northern limit, and zoneof ecological optimum, along with the estimated IV atthe ecological optimum, were calculated using quartiledistributions (Table 5). Though this statistical methodof calculating limits is flawed for some species by out-lier counties having predicted distributions accordingto the models, it gives us a metric for quantifying po-tential migration for most species. Because each num-ber corresponds to a latitudinal strip of 10 km startingfrom the southern tip of Florida, we can begin to es-timate potential shifts in the northern and southern lim-its. According to this analysis, four species (Betulapapyrifera, Pinus resinosa, Populus grandidentata, andThuja occidentalis) are projected by the GISS GCMmodel to have their southern optimum move north ofthe United States border, while five additional species(Abies balsamea, Acer saccharum, Betula alleghanien-sis, Populus tremuloides, and Prunus serotina) are pro-jected to disappear according to the GFDL model.Shifts of the southern optima are also apparent for sev-eral species that do stay in the United States: Acerpensylvanicum, Betula lenta, Carya glabra, Celtis oc-cidentalis, Cercis canadensis, Cornus florida, Cratae-gus sp., Diospyros virginiana, Oxydendrum arboreum,Pinus virginiana, Quercus muehlenbergii, Q. palustris,Tilia americana, and Tsuga canadensis show a sizable


TABLE 5. Calculated north–south ranges, shifts, and estimated importance values due to GISS and GFDL model forecasts.Scores of 299 indicate that the species is projected to become absent in the United States, while a score of 208 indicatesthat the species abuts the Canadian border, and a score of 70 indicates the species abuts the Gulf of Mexico.

Species

Southern optimum

FIA Pred. GISS GFDL

Northern optimum

FIA Pred. GISS GFDL

Abies balsamea 191 135 199 299 208 208 204 299

Acer negundoAcer pensylvanicumAcer rubrumAcer saccharinum

71118104

95

76143104118

76166

7076

76166

7076

208208208208

208208208208

208203208208

208203199208

Acer saccharum 120 110 148 299 208 208 208 299

Betula alleghaniensisBetula lentaBetula papyrifera

109120179

146144188

191157

299

299157

299

208202208

208208207

208208

299

299208

299

Carpinus caroliniana 70 70 70 70 206 205 191 208

Carya cordiformisCarya glabraCarya ovataCarya sp.Carya tomentosa

9686

1327376

140129110

7376

140153110

8276

140153110

8176

208191208201194

208179208204208

208179208208208

208179207206208

Celtis laevigataCeltis occidentalis

7085

70113

70140

70140

134208

128208

208208

208208

Cercis canadensis

Cornus florida

Crataegus sp.

86

70

70

104

70

153

121

89

194

121

89

185

183

201

208

182

186

208

208

201

203

208

197

203

Diospyros virginiana

Fagus grandifolia

Fraxinus americana

70

71

71

70

103

70

80

102

70

80

102

70

161

208

208

156

208

208

208

208

208

208

163

208Fraxinus nigraFraxinus pennsylvanicaFraxinus sp.

1507070

1817070

1857070

1857070

208208157

208208170

208208170

208208170

Gleditsia triacanthos

Ilex opaca

74

70

81

111

81

99

81

99

203

161

197

160

206

147

206

147

Juglans nigra

Juniperus virginiana

Liquidambar styraciflua

Liriodendron tulipifera

Maclura pomifera

110

70

70

74

93

70

85

70

92

70

70

70

70

94

70

81

70

70

94

71

207

208

173

189

189

198

208

173

208

202

208

208

202

208

201

208

208

205

208

208

Magnolia virginiana

Morus rubra

70

95

70

144

70

144

70

144

168

206

168

201

126

197

138

201

Nyssa aquaticaNyssa bifloraNyssa sylvatica

707070

777070

777070

777070

139143191

138195177

138160204

138162204

Ostrya virginiana

Oxydendrum arboreum

73

81

80

83

80

99

80

99

208

163

208

152

208

175

205

175

Pinus echinataPinus elliottiiPinus palustris

707070

707070

707070

707070

168115114

12991

111

190127154

191127180

Pinus resinosaPinus strobusPinus taedaPinus virginiana

157104

7089

176104

7075

299104

7091

299104

7093

208208153169

208208153204

299208171187

299208176190

Platanus occidentalis 72 86 86 86 191 171 177 202

Populus deltoidesPopulus grandidentataPopulus tremuloides

128152162

145135135

145299205

145299299

208208208

208208208

208299208

208299299

Prunus serotina 70 93 188 299 208 208 208 299


TABLE 5. Extended.

Latitudinal optimum

FIA Pred. GISS GFDL

IV at optimum latitude

FIA Pred. GISS GFDL

200 180 202 299 4.8 5.4 4.0 299

145170156161

144177156167

144182147163

144182127159

11.06.2

20.212.8

6.07.4

15.613.0

6.03.08.3

14.6

6.03.04.9

14.8164 161 183 299 16.2 12.8 10.0 299

169164198

182173198

200184

299

299184

299

6.08.56.6

6.49.1

11.0

4.68.1

299

2997.4

299

138 147 128 139 4.9 5.9 4.6 4.6

161152174133137

174150167130134

174165167137145

174165167140146

6.06.5

10.510.3

6.8

4.96.76.28.66.1

5.05.86.69.36.1

4.95.86.58.36.1

99156

97160

143174

139178

12.714.7

20.110.0

21.739.6

14.044.2

140

129

147

143

124

182

170

145

198

170

143

194

5.3

8.7

7.8

4.3

8.0

8.5

4.6

8.4

12.7

4.6

8.3

8.7

116

155

156

116

164

163

148

144

159

149

132

160

4.8

9.3

10.7

6.3

9.8

11.1

6.6

7.6

9.8

6.4

6.2

10.1185145117

150

117

195137118

164

134

196140122

173

123

196140117

173

123

6.815.0

5.4

7.7

7.2

7.210.4

3.6

6.9

10.7

7.711.5

4.1

14.1

6.2

7.511.8

3.5

14.3

6.2

167

151

114

134

146

135

150

118

141

143

140

139

128

142

134

150

139

132

142

144

8.5

13.5

18.6

9.3

12.2

5.5

8.3

18.3

9.8

15.6

21.5

12.6

13.0

9.0

18.0

27.8

13.7

13.6

8.4

18.7

114

162

105101132

109

173

103123122

97

169

103112132

101

173

103113132

5.4

10.1

6.412.3

7.0

5.2

16.9

8.29.35.3

10.8

20.1

7.317.6

6.6

11.0

16.9

7.317.5

5.4

143

119

1089093

164

119

1037992

147

137

12195

104

153

137

12098

121

6.7

6.7

11.712.6

9.6

9.3

6.1

13.213.1

7.2

8.9

10.0

16.717.9

8.9

16.0

9.6

17.719.010.3

185156109126

134

194155110126

127

299156112137

137

299156121140

140

7.510.435.411.6

6.2

6.910.735.2

7.5

5.5

29915.241.011.6

9.3

29915.536.914.5

9.6

173183188

177185186

175299206

177299299

27.16.69.1

26.49.07.9

21.8299

8.7

28.8299299

139 153 198 299 8.0 7.7 10.3 299


TABLE 5. Continued.

Species

Southern optimum

FIA Pred. GISS GFDL

Northern optimum

FIA Pred. GISS GFDL

Quercus albaQuercus coccineaQuercus falcata var. falcataQuercus falcata var. pagodaefolia

7070

10092

70709292

70739599

70719599

145158208201

141158208204

182196208198

182198208196

Quercus laurifoliaQuercus macrocarpaQuercus marilandicaQuercus muehlenbergiiQuercus nigra

70148

70105

70

70156

74106

70

70156

79127

70

70156

90127

70

116208171200154

113208159173120

131208208208154

138208208208171

Quercus palustrisQuercus phellosQuercus prinusQuercus rubraQuercus stellata

1237093

11074

1337093

10674

1737089

10977

1727089

13786

204151204208175

191128187208157

190191208208208

190208208208208

Quercus velutina 105 93 109 109 208 208 208 208

Robinia pseudoacacia

Salix nigraSalix sp.

Sassafras albidum

87

12370

79

106

16470

95

106

16470

95

106

16470

95

207

208207

206

198

207123

208

206

207122

182

198

207122

208

Taxodium distichumTaxodium distichum var. nutans

Thuja occidentalis

Tilia americana

Tsuga canadensis

7070

194

110

109

7070

144

150

109

7070

299

182

128

7070

299

175

128

123145

208

208

208

78193

208

208

208

108193

299

205

208

121193

299

205

208

Ulmus alataUlmus americanaUlmus rubraUlmus sp.

74837876

70708073

70707070

70707070

145208208206

146208208206

208208208206

208208208206

Note: FIA 5 Forest Inventory Analysis data; Pred. 5 predicted current; predicted future with GISS 5 Goddard Instituteof Space Sciences model; predicted future with GFDL 5 Geophysical Fluid Dynamics Laboratory model. Numbers correspondto row numbers shown in Fig. 1.

(.100 km) northward shift in their southern boundaries(Table 5). Of course, because of tree longevity andremnant refugia, it would take centuries for these shiftsto be realized (Loehle and LeBlanc 1996).

For a large portion of the species, the northern op-timum abuts the Canadian border so an estimate ofpotential migration is not possible (row number 208for northern optimum in Table 5). However, for thespecies where the northern optimum is projected to staysouth of the Canadian border, sizable shifts in northernoptima are apparent. For example, the following spe-cies are projected to shift their northern limit by at least100 km to the north: Celtis laevigata, Cercis canaden-sis, Cornus florida, Diospyros virginiana, Liquidambarstyraciflua, Nyssa sylvatica, Oxydendrum arboreum,Pinus echinata, P. elliottii, P. palustris, P. taeda, Quer-cus alba, Q. coccinea, Q. laurifolia, Q. marilandica,Q. muehlenbergii, Q. nigra, Q. phellos, Q. stellata, andUlmus alata. Many of these species do not concomi-tantly shift their southern limits northward. Differingenvironmental factors are related to the range bound-aries on the north vs. south. This geographic shift invariable importance shows the power of the RTA anal-ysis; for general linear or general additive methods, thevariables have to operate equally everywhere.

Because calculations of the ecological optimum lat-itude are not as prone to outliers and boundary effects,they may provide the best indicator of projected speciesmovement (Table 5). A total of 36 species are projectedto have their zone of maximum IV migrate at least 100km north. Of these, seven are estimated to move inexcess of 250 km: Cercis canadensis, Prunus serotina,Quercus coccinea, Q. marilandica, Q. stellata, Tsugacanadensis, and Ulmus alata. Interestingly, five taxaare projected to have their ecological optima shiftsouth: Carpinus caroliniana, Fagus grandifolia, Ju-niperus virginiana, Magnolia virginiana, and Sassafrasalbidum. In these cases, there seems to be a relativeconstriction of range to the higher elevations of thesouthern Appalachians.

Estimated IV for the ecological optima are also pre-sented for current and future projections (Table 5).Some species are projected to significantly increase inimportance at their ecological optima, including Celtisoccidentalis, Gleditsia triacanthos, Juglans nigra,Magnolia virginiana, Nyssa biflora, Pinus elliottii, Pi-nus strobus, Quercus stellata, Taxodium distichum, Til-ia americana, and Ulmus alata. On the other hand,significant decreases in importance are projected for


TABLE 5. Extended.

Latitudinal optimum

FIA Pred. GISS GFDL

IV at optimum latitude

FIA Pred. GISS GFDL

107111154146

104111150143

121136155147

120139154145

5.36.7

13.17.0

5.35.59.55.3

5.07.0

11.95.5

5.26.9

13.78.1

91180119149103

90183115147

92

105182150170108

107182146170117

6.318.6

7.56.98.5

6.520.7

8.87.99.0

10.815.310.2

8.39.4

10.915.312.6

8.49.9

162108142161120

16294

145157115

181129155172146

180135152172146

7.95.2

12.18.6

13.6

7.37.1

13.86.59.9

8.18.1

12.95.7

27.4

8.97.9

12.84.6

42.3156

153

173129

150

151

151

18498

148

163

151

18494

138

154

146

18494

145

10.0

6.5

10.99.3

6.9

6.8

9.4

20.313.2

5.9

7.5

9.4

20.314.7

6.2

11.0

9.6

20.314.7

6.3

88105

201

165

159

73110

188

179

159

88110

299

193

189

90110

299

190

189

4.27.1

8.6

7.0

7.8

5.06.4

7.0

6.6

6.3

17.15.0

299

13.5

9.3

15.55.0

299

16.6

9.8

105149156137

108141144138

136141142131

144140135131

9.213.5

8.06.7

6.810.9

5.08.9

11.510.6

9.18.9

16.710.713.2

8.9

several species, including Acer rubrum, Liquidambarstyraciflua, and Quercus macrocarpa.

CONCLUSIONS

The work described here differs from most in that itoperates at the species level over a large geographicregion for most of the major species in the easternUnited States. Regression tree analysis has been shownto be a valuable tool to improve understanding of spe-cies–environment relations for 80 tree species. Spacehas limited the information portrayed for each species,but the key variables correlated with the present IVdistributions of each species have been identified. Theoperating variables differ geographically, so that, aswe have shown, the northern range boundary can berelated to different variables than the southern rangeboundary. Accuracy of prediction of current IVs variedwidely among species. Species more related to distur-bance or specialized habitats area did not model as wellas those regionally controlled by environmental vari-ables. Model improvements could be expected, es-pecially for specialist species, with finer scale inputdata and fine tuning of individual models.

RTA was used here to predict potential migrationsof trees under a climate scenario associated with a two-

fold increase in the level of atmospheric CO2. We em-phasize that potential future ranges presented here donot represent forecasts, but rather an indication of thepotential impact on species distributions. With ouranalysis of potential effects, we show that 27–36 spe-cies would significantly increase in area and/or weight-ed IV, while an additional 30–34 species would de-crease by at least 10%, following equilibrium after achanged climate. Nearly half of the species assessed(36 out of 80) showed the potential for the ecologicaloptima to shift at least 100 km to the north, includingseven that could move .250 km. Depending on theGCM used, 4–9 species would potentially move out ofthe United States. Results obtained here compare fa-vorably with those obtained by other researchers (forsome species). However, only a few eastern U.S. spe-cies had been assessed for changes in potential distri-bution and abundance prior to this study.

Historic rates of migration (;10–50 km/100 yr) willnot likely occur with current fragmented habitat. Evenat historic rates, many species would not reach thepotential distributions predicted here within the nextcentury without major intervention. In work underwayin a related research effort (e.g., Schwartz 1996; L. R.Iverson, A. Prasad, and M. W. Schwartz, unpublished


manuscript), we are investigating more realistic mi-gration scenarios based on historic species migrationrates and actual landscapes.

ACKNOWLEDGMENTS

Sincere thanks to all the people who provided data for thiseffort, and to the USDA Forest Service Northern GlobalChange Program (R. Birdsey, Program Manager) for support.Thanks also to Mark Schwartz, David Mladenoff, CharlesScott, Thomas Jacob, William Baker, and two anonymousreviewers for technical reviews, and to Mary Buchanan foreditorial review.

The use of trade, firm, or corporation names in this pub-lication is for the information and convenience of the reader.Such use does not constitute and official endorsement or ap-proval by the U.S. Department of Agriculture or the ForestService of any product or service to the exclusion of othersthat may be suitable.

LITERATURE CITED

Baker, F. A. 1993. Classification and regression tree analysisfor assessing hazard of pine mortality caused by Hetero-basidion annosum. Plant Disease 77:136–139.

Bonan, G. B., and L. Sirois. 1992. Air temperature, treegrowth, and the northern and southern range limits to Piceamariana. Journal of Vegetation Science 3:495–506.

Booth, T. H. 1990. Mapping regions climatically suitable forparticular tree species at the global scale. Forest Ecologyand Management 36:47–60.

Borchert, M., F. W. Davis, J. Michaelsen, and L. D. Oyler.1989. Interactions of factors affecting seedling recruitmentof blue oak (Quercus douglasii) in California. Ecology 70:389–404.

Box, E. O., D. W. Crumpacker, and E. D. Hardin. 1993. Aclimatic model for location of plant species in Florida,U.S.A. Journal of Biogeography 20:629–644.

Breiman, L., J. Freidman, R. Olshen, and C. Stone. 1984.Classification and regression trees. Wadsworth, Belmont,California, USA.

Burton, P. J., and S. G. Cumming. 1995. Potential effects ofclimatic change on some western Canadian forests, basedon phenological enhancements to a patch model of forestsuccession. Water Air and Soil Pollution 82:401–414.

Chambers, J. M., and T. J. Hastie. 1993. Statistical modelsin S. Chapman and Hall, New York, New York, USA.

Clark, L. A., and D. Pregibon. 1992. Tree-based models.Pages 377–419 in T. J. Hastie, editor. Statistical models inS. Wadsworth, Pacific Grove, California, USA.

Davis, F. W., J. Michaelsen, R. Dubayah, and J. Dozier. 1990.Optimal terrain stratification for integrating ground datafrom FIFE. Pages 11–15 in Symposium on FIFE, FirstISLSCP Field Experiment. American Meteorological So-ciety, Boston, Massachusetts, USA.

Davis, M. B., and C. Zabinski. 1992. Changes in geograph-ical range resulting from greenhouse warming: effects onbiodiversity in forests. Pages 297–308 in R. L. Peters andT. E. Lovejoy, editors. Global warming and biological di-versity. Yale University Press, New Haven, Connecticut,USA.

Delcourt, H. R., and P. A. Delcourt. 1988. Quaternary land-scape ecology; relevant scales in space and time. LandscapeEcology 2:23–44.

Dyer, J. M. 1995. Assessment of climatic warming using amodel of forest species migration. Ecological Modelling79:199–219.

Elias, T. S. 1980. The complete trees of North America. VanNostrand Reinhold, New York, New York, USA.

ESRI. 1992. ArcUSA 1:2M, user’s guide and data reference.Environmental Systems Research Institute, Redlands, Cal-ifornia, USA.

. 1993. Arc/Info GRID command reference. Environ-mental Systems Research Institute, Redlands, California,USA.

Ertsen, A. C. D., J. W. Frens, J. W. Nieuwenhuis, and M. J.Wassen. 1995. An approach to modelling the relationshipbetween plant species and site conditions in terrestrial eco-systems. Landscape and Urban Planning 31:143–151.

Flannigan, M. D., and F. I. Woodward. 1994. Red pine abun-dance: current climatic control and responses to futurewarming. Canadian Journal of Forest Research 24:1166–1175.

Foley, J. A., I. C. Prentice, N. Ramankutty, S. Levis, D.Pollard, S. Sitch, and A. Haxeltine. 1996. An integratedbiosphere model of land surface processes, terrestrial car-bon balance and vegetation dynamics. Global Biogeochem-ical Cycles 10:603–628.

Franklin, J. 1995. Predictive vegetation mapping: geographicmodelling of biospatial patterns in relation to environmen-tal gradients. Progress in Physical Geography 19:474–499.

Golet, F. C., A. J. K. Calhoun, W. R. DeRagon, D. J. Lowry,and A. J. Gold. 1993. Ecology of red maple swamps inthe glaciated northeast: a community profile. BiologicalReport 12, U.S. Department of the Interior, Fish and Wild-life Service, Washington, D.C., USA.

Goodchild, M. F., and S. Gopal. 1989. Accuracy of spatialdatabases. Taylor and Francis, Bristol, Pennsylvania, USA.

Hansen, J., I. Fung, A. Lacis, D. Rind, S. Lebedeff, and R.Ruedy. 1988. Global climate changes as forecast by God-dard Institute for Space Studies three-dimensional model.Journal of Geophysical Research 93:9341–9364.

Hansen, M. H., T. Frieswyk, J. F. Glover, and J. F. Kelly.1992. The eastwide forest inventory data base: users man-ual. U.S. Forest Service North Central Forest ExperimentStation General Technical Report NC-151.

Hobbs, Richard J. 1994. Dynamics of vegetation mosaics:can we predict responses to global change? Ecoscience1(4):346–356.

Hughes, L., E. M. Cawsey, and M. Westoby. 1996. Climaticrange sizes of Eucalyptus species in relation to future cli-mate change. Global Ecology and Biogeography Letters 5:23–29.

Huntley, B., P. Berry, W. Cramer, and A. P. McDonald. 1995.Modelling present and potential future ranges of some Eu-ropean higher plants using climate response surfaces. Jour-nal of Biogeography 22:967–1001.

Iverson, L. R., M. E. Dale, C. T. Scott, and A. Prasad. 1997.A GIS-derived integrated moisture index to predict forestcomposition and productivity in Ohio forests. LandscapeEcology 12:331–348.

Iverson, L. R., A. Prasad, and C. T. Scott. 1996a. Preparationof forest inventory and analysis (FIA) and state soil geo-graphic data base (STATSGO) data for global change re-search in the eastern United States. Pages 209–214 in J.Hom, R. Birdsey, and K. O’Brian, editors. Proceedings,1995 meeting of the Northern Global Change Program. U.S.Forest Service Northeastern Forest Experiment StationGeneral Technical Report NE-214.

Iverson, L. R., C. T. Scott, M. Dale, and A. M. G. Prasad.1996b. Development of an integrated moisture index forpredicting species composition. Pages 101–116 in M. Kohland G. Z. Gertner, editors. Caring for the forest: researchin a changing world. Statistics, mathematics, and comput-ers. Swiss Federal Institute for Forest, Snow and LandscapeResearch, Birmensdorf, Switzerland.

Jacobson, G. L., Jr., and A. Dieffenbacker-Krall. 1995. Whitepine and climate change: insights from the past. Journal ofForestry 93:39–42.

Jacobson, G. L., Jr., T. Webb III, and E. C. Grimm. 1987.Patterns and rates of vegetation change during the degla-ciation of eastern North America. Pages 277–288 in W. F.


Ruddiman and H. E. Wright, Jr., editors. North Americaand adjacent oceans during the last deglaciation. GeologicalSociety of America, Boulder, Colorado. USA.

Jelinski, D. E., and J. Wu. 1996. The modifiable areal unitproblem and implications for landscape ecology. Land-scape Ecology 11:129–140.

Joyce, L. A., M. A. Fosberg, and J. M. Comanor. 1990. Cli-mate change and America’s forests. U.S. Forest ServiceRocky Mountain Forest and Range Experiment StationGeneral Technical Report GTR RM-187.

Kattenberg, A., F. Giorgi, H. Grassl, G. A. Meehl, J. F. B.Mitchell, R. J. Stouffer, T. Tokioka, A. J. Weaver, and T.M. L. Wigley. 1996. Climate models—projections of fu-ture climate. Pages 285–357 in J. T. Houghton et al., editors.Climate Change 1995: the science of climate change. In-tergovernmental Panel on Climate Change. Cambridge Uni-versity Press, Cambridge, UK.

Lees, B. G., and K. Ritman. 1991. Decision-tree and rule-induction approach to integration of remotely sensed andGIS data in mapping vegetation in disturbed or hilly en-vironments. Environmental Management 15:823–831.

Loehle, C., and D. LeBlanc. 1996. Model-based assessmentsof climate change effects on forests: a critical review. Eco-logical Modelling 90:1–31.

Lynn, H., C. L. Mohler, S. D. DeGloria, and C. E. McCulloch.1995. Error assessment in decision-tree models applied tovegetation analysis. Landscape Ecology 10:323–335.

Mackey, B. G., and R. A. Sims. 1993. A climatic analysisof selected boreal tree species, and potential responses toglobal climate change. World Resource Review 5:469–487.

Malanson, G. P. 1993. Comment on modeling ecologicalresponse to climatic change. Climatic Change 23:95–105.

McGarigal, K., and B. J. Marks. 1995. FRAGSTATS: spatialpattern analysis program for quantifying landscape struc-ture. U.S. Forest Service Pacific Northwest Research Sta-tion General Technical Report PNW-GTR-351.

McGuire, A. D., L. A. Joyce, D. W. Kicklighter, J. M. Melillo,G. Esser, and C. J. Vorosmarty. 1993. Productivity re-sponse of climax temperate forests to elevated temperatureand carbon dioxide: a North American comparison betweentwo global models. Climatic Change 24:287–310.

McLemore, B. F. 1990. Cornus florida L. Flowering dog-wood. Pages 278–283 in R. M. Burns and B. H. Honkala,coordinators. Silvics of North America. Volume 2: Hard-woods. U.S. Forest Service Agriculture Handbook 654.

Melillo, J. M., J. Borchers, J. Chaney, H. Fisher, S. Fox, andA. Haxeltine. 1995. Vegetation/ecosystem modeling andanalysis project: comparing biogeography and biogeo-chemistry models in a continental-scale study of terrestrialecosystem responses to climate change and CO2 doubling.Global Biogeochemical Cycles 9:407–437.

Michaelsen, J., D. S. Schimel, M. A. Friedl, F. W. Davis, andR. C. Dubayah. 1994. Regression tree analysis of satelliteand terrain data to guide vegetation sampling and surveys.Journal of Vegetation Science 5:673–686.

Moore, D. E., B. G. Lees, and S. M. Davey. 1991. A newmethod for predicting vegetation distributions using deci-sion tree analysis in a geographic information system. Jour-nal of Environmental Management 15:59–71.

Neilson, R. P. 1995. A model for predicting continental-scalevegetation distribution and water balance. Ecological Ap-plications 5:362–385.

Neilson, R. P., and D. Marks. 1994. A global perspective ofregional vegetation and hydrologic sensitivities from cli-matic change. Journal of Vegetation Science 5:715–730.

Olson, R. J., C. J. Emerson, and M. K. Nungesser. 1980.Geoecology: a county-level environmental data base forthe conterminous United States. Publication Number 1537.

Oak Ridge National Laboratory Environmental SciencesDivision, Oak Ridge, Tennessee, USA.

Overpeck, J. T., P. J. Bartlein, and T. Webb III. 1991. Po-tential magnitude of future vegetation change in EasternNorth America: comparisons with the past. Science 254:692–695.

Perala, D. A. 1990. Populus tremuloides Michx. Quakingaspen. Pages 555–569 in R. M. Burns and B. H. Honkala,coordinators. Silvics of North America. Volume 2: Hard-woods. U.S. Forest Service Agriculture Handbook 654.

Prentice, I. C., W. Cramer, S. P. Harrison, R. Leemans, R. A.Monserud, and A. M. Solomon. 1992. A global biomemodel based on plant physiology and dominance, soil prop-erties and climate. Journal of Biogeography 19:117–134.

Rastetter, E. B. 1996. Validating models of ecosystem re-sponse to global change. BioScience 46:190–197.

Reed, R. A., R. K. Peet, M. W. Palmer, and P. S. White. 1993.Scale dependence of vegetation–environment correlation:a case study of a North Carolina piedmont woodland. Jour-nal of Vegetation Science 4:329–340.

Schwartz, M. W. 1996. Assessing the ability of plants torespond to climatic change through distribution shifts.Pages 184–191 in J. Hom, R. Birdsey, and K. O’Brian,editors. Proceedings, 1995 meeting of the Northern GlobalChange Program. U.S. Forest Service Northeastern ForestExperiment Station General Technical Report NE-214.

Soil Conservation Service. 1991. State soil geographic database (STATSGO) data users guide. USDA Soil Conser-vation Service Miscellaneous Publication 1492.

Starfield, A. M., and F. S. Chapin, III. 1996. Model of tran-sient changes in arctic and boreal vegetation in responseto climate and land use change. Ecological Applications 6:842–864.

Statistical Sciences. 1993. S-PLUS guide to statistical andmathematical analysis, version 3.2. StatSci, Seattle, Wash-ington, USA.

U.S. Geological Survey. 1987. Digital elevation models:U.S. Geological Survey data user’s guide 5. U.S. Geolog-ical Survey, Reston, Virginia, USA.

U.S. Forest Service. 1993. Forest type groups of the UnitedStates. Map produced by Z. Zhu, D. L. Evans, and K.Winterberger. Southern Forest Experiment Station, Starks-ville, Mississippi, USA.

USEPA. 1993. EPA-Corvallis model-derived climate data-base and 2 3 CO2 predictions for long-term mean monthlytemperature, vapor pressure, wind velocity and potentialevapotranspiration from the Regional Water Balance Modeland precipitation from the PRISM model, for the conter-minous United States. Digital raster data on a 10 3 10 km,470 3 295 Albers Equal Area grid, in ‘‘Image ProcessingWorkbench’’ format. USEPA Environmental Research Lab-oratory. Corvallis, Oregon, USA.

Venables, W. N., and B. D. Ripley. 1994. Modern appliedstatistics with S-Plus. Springer-Verlag, New York, NewYork, USA.

Verbyla, D. L. 1987. Classification trees: a new discrimi-nation tool. Canadian Journal of Forest Research 17:1150–1152.

Webb, T., III, and P. J. Bartlein. 1992. Global changes duringthe last 3 million years: climatic controls and biotic re-sponses. Annual Review of Ecology and Systematics 23:141–173.

Wetherald, R. T., and S. Manabe. 1988. Cloud feedback pro-cesses in a general circulation model. Journal of Atmo-spheric Science 45:1397–1415.

Woodward, F. I. 1992. Tansley Review no. 41. Predictingplant responses to global environmental change. New Phy-tologist 122:239–251.

Predicting abundance of 80 tree species following climate change in the eastern United States

Documents

Transcript of Predicting abundance of 80 tree species following climate change in the eastern United States