Assessing current and projected suitable habitats for tree-of-heaven along the Appalachian Trail.

Downloaded from http://rstb.royalsocietypublishing.org/ on February 2, 2015

Assessing current and projected suitable habitats for tree-of-heaven along the Appalachian Trail John Clark, Yeqiao Wang and Peter V. August

rstb.royalsocietypublishing.org

Research Cite this article: Clark J, Wang Y, August PV. 2014 Assessing current and projected suitable habitats for tree-of-heaven along the Appalachian Trail. Phil. Trans. R. Soc. B 369: 20130192. http://dx.doi.org/10.1098/rstb.2013.0192 One contribution of 9 to a Theme Issue ‘Satellite remote sensing for biodiversity research and conservation applications’. Subject Areas: ecology, environmental science Keywords: tree-of-heaven, terrestrial observation prediction system, Appalachian trail, habitat suitability model, climate change Author for correspondence: Yeqiao Wang e-mail: [email protected]

Department of Natural Resources Science, University of Rhode Island, Kingston, RI 02881, USA The invasion of ecosystems by non-native species is a major driver of biodiversity loss worldwide. A critical component of effective land management to control invasion is the identification and active protection of areas at high risk of future invasion. The Appalachian Trail Decision Support System (A.T.-DSS) was developed to inform regional natural resource management by integrating remote sensing data, ground-based measurements and predictive modelling products. By incorporating NASA’s remote sensing data and modelling capacities from the Terrestrial Observation and Prediction System (TOPS), this study examined the current habitat suitability and projected suitable habitat for the invasive species tree-of-heaven (Ailanthus altissima) as a prototype application of the A.T.-DSS. Species observations from forest surveys, geospatial data, climatic projections and maximum entropy modelling were used to identify regions potentially susceptible to tree-of-heaven invasion. The modelling result predicted a 48% increase in suitable area over the study area, with significant expansion along the northern extremes of the Appalachian Trail.

1. Introduction The Appalachian National Scenic Trail (A.T.) is a footpath stretching from Springer Mountain in Georgia to Mount Katahdin in Maine and spans over 3500 km of peaks, valleys and ridges along the Appalachian Mountains (figure 1). The trail intersects 14 states; eight National Forests; six units of the National Park System; more than 70 State Park, Forest and Game Management units; and 287 local jurisdictions. The A.T. passes through some of the largest and least fragmented forest blocks remaining in the eastern United States [1], forests containing rich biological diversity and the headwaters of important water resources. The A.T.’s north –south alignment and gradients of elevation, latitude and moisture represent a continental-scale cross-section ‘MEGA-Transect’ of eastern US forest and alpine areas, offering a setting for collecting scientific data on the structure, function, species composition and condition of ecosystems [1]. The high elevation setting of the A.T. provides an ideal landscape for early detection of changes in natural resources of the eastern United States; for example, development encroachment, acid precipitation, invasions of exotic species and climate change impacts. The Appalachian Trail Decision Support System (A.T.-DSS) creates a coherent framework by integrating NASA multi-platform sensor data, Terrestrial Observation and Prediction System (TOPS) models [2] and in situ measurements. The goal of the A.T.-DSS is to improve the effectiveness of decisionmaking in the management of A.T. natural resources and for biodiversity by integrating NASA data and modelling products that link climate models (e.g. through TOPS) and ecological models (e.g. habitat suitability) with in situ observations. The A.T.-DSS provides critical geospatial information, in particular the past trends of land surface phenology (LSP) derived from remote sensing observation, the climate projection data from TOPS models and in situ measurements from United States Department of Agriculture (USDA) Forest Service

& 2014 The Author(s) Published by the Royal Society. All rights reserved.


2

M221 221

Phil. Trans. R. Soc. B 369: 20130192

211

221

221

N 231 M221

HUC-10 shell

231

A.T. centreline ecoregion province M211: Adirondack-New England mixed forest 221: Eastern broadleaf forest

221

M221: Central Appalachian broadleaf forest 211: Northeastern mixed forest 231: Southeastern mixed forest

231 0

75

150

300

450

600 km

Figure 1. The spatial extent of the A.T.-DSS was established by selecting all 10-digit hydrological unit code (HUC-10) watersheds within 8 km of A.T. The resulting area is termed the A.T. HUC-10 shell, or A.T.-shell. In addition, Bailey’s ecoregion provinces intersecting the A.T.-shell subdivide the region into units with similar climates and vegetation. (Online version in colour.)

Forest Inventory and Analysis (FIA) data [3]. These data were the basis for our assessment of the pattern and trends of the changing habitat suitability for Ailanthus from a megatransect perspective along the Appalachian Mountains using integrated modelling for management decision support. The economic and environmental costs of biological invasion can be significant [4,5]. Invasive species disrupt the balance of ecosystems by outcompeting and displacing native species [6]. The net result is the loss of habitat, rare and endangered species, and native biodiversity [7]. As ecological communities change, essential ecological functions can be altered and ecosystem services [8] can be degraded [8–10]. Commonly referred to as tree-of-heaven, Ailanthus altissima (Mill.) Swingle is a deciduous member of the Simaroubaceae family native to the temperate regions of central China. Ailanthus altissima (henceforth Ailanthus) is an exotic tree species pervasive throughout the USA owing to its rapid growth, high fecundity, hardy tolerance and strong


211

competitive ability [11,12]. Ailanthus is susceptible to frost damage, particularly juveniles, thus restricting Ailanthus from higher latitudes and elevations. Ailanthus is relatively drought hardy, though extended dry periods exclude the species from extremely arid regions. The species is also very tolerant of harsh environmental pollutants and thrives across a wide range of poor soils [13]. Ailanthus is shadeintolerant and must exploit gaps in the canopy to become established [14] and is resistant to herbivore browsing [15]. These traits make Ailanthus ideally suited to colonize the ruderal conditions found in human-impacted or otherwise disturbed areas. It is commonly found in urban areas and transportation corridors throughout the USA. Strong competitive ability enables Ailanthus to severely impact native communities within its introduced range [16,17]. Removing Ailanthus allows native vegetation to recover and prevents further dispersal. However, management efforts are confounded by the tree’s ability to


to maximize the probability, or gain, of the observed species presence points [37]. Projected climate data can then be substituted for current conditions to examine potential shifts in the distribution of suitable Ailanthus habitats [44].

2. Methods (a) Defining the study area

(b) Forest inventory and analysis data Ground-based vegetation samples were provided by the FIA programme of the USDA Forest Service. The FIA programme provides a nationwide systematic sample of forested ecosystems well-suited for broad-scale ecological analysis [47]. Measurements include the species, size and condition of trees within the plot as well as physiographic site attributes [3]. Within the A.T.-shell, data for 3926 FIA plots were available between 2002 and 2010. Among those, observations of Ailanthus were recorded at 136 locations (figure 2). In addition to the plot coordinates, several attributes were retained from the FIA records to examine the characteristics of sites colonized by Ailanthus and compare them with the overall study area. Plot attributes included elevation, aspect, slope, distance to improved road, land ownership, water on plot, physiographic class, stand age, stand size and basal area of live trees [3].

(c) Remote sensing of past trends of land surface phenology LSP has been studied in the context of ecosystem responses to climate change [48] and for monitoring changes in vegetation life cycle events. Because LSP is based on remote sensing observations at regional to global scales, it serves as a key biological indicator for detecting the response of terrestrial ecosystems to climatic variation. LSP metrics are primarily based on timeseries of vegetation indices derived from sensors such as the advanced very high-resolution radiometer, SPOT-VEGETATION (VGT) and moderate resolution imaging spectroradiometer (MODIS). Because of the spatial resolution of the remote sensing data, LSP is an indicator of mixtures of land covers and is distinct from the traditional notion of species-centric phenology, such as seasonal flowering or budburst [49]. LSP metrics typically retrieve the time of onset greenness as the start of the season (SOS); onset of senescence or time of end of greenness as the end of the season (EOS), timing of maximum of the growing season by peak vegetation indices and the length of growing season (LOS) or duration of greenness. An increasing number of studies have reported shifts in timing and length of the growing season based on remote sensing data and climatological studies [50,51]. Knowing the pattern change and trends of LSP in the past allows us to evaluate the effects on habitat suitability and environmental variation of the region. To obtain the patterns and trends of LSP metrics and the relationship to climatic variations, we used time-series data between 1982 and 2006 from the global inventory modelling and mapping studies (GIMMS) [52] and the surface observation and gridding system (SOGS) through the TOPS.

Phil. Trans. R. Soc. B 369: 20130192

The spatial extent of the A.T.-DSS is based on a boundary defined by the National Park Service and United States Geological Survey (USGS). It was established by selecting all 10-digit hydrological unit code (HUC-10) watersheds within five statute miles (8 km) of the trail; the resulting area is termed the A.T. HUC-10 shell [45]. The shell provides an ecologically relevant corridor boundary around the A.T. for habitat suitability modelling. In addition, Bailey’s ecoregion provinces [46] were used to subdivide the A.T.-shell into units with similar climates and vegetation (figure 1).

3


re-sprout from roots and stumps. Early detection is crucial for minimizing the costs of control programmes and the risks of further dispersal and establishment. While Ailanthus is currently absent from many sensitive ecosystems, such as the boreal forests of White Mountain National Forest in New Hampshire, changing conditions could lead to impacts in these regions. A greater understanding of the processes and patterns of Ailanthus invasion within the landscape is needed to inform effective management programmes [18,19]. Climate change is a widely recognized phenomenon [20] with significant implications for the spread, impact and management of invasive species [21–23]. Climate change alters temperature and precipitation patterns, resource availability (CO2, N), and affects management decisions and practices in land-cover and land-use [24]. Warming trends have been predicted to correspond with horizontal migrations of vegetation averaging 0.43 km yr21 across a wide variety of ecosystems [25]. In particular, for the Appalachian Mountains region, temperatures are predicted to increase by 2–68C by the end of the twenty-first century [26]. An objective of this study is the assessment of the direct effects of climate change on Ailanthus habitat suitability based on in situ observations, TOPS data products and predictive modelling [27]. Innovative statistical methods and advances in geographical information system (GIS) technology have led to the emergence of species distribution modelling as an important ecological tool [28]. Ecological niche theory examines the relationship between species fitness and environmental conditions [29]; species distribution modelling extends this paradigm into geographical space by linking species distribution to spatial variability [30 –32]. These models often generate spatially explicit predictions of habitat suitability, typically by comparing environmental variables between species presence and absence locations. Modelling the suitable habitats of an invasive species presents a unique challenge. A major assumption underlying most models is that the absence of a species from a particular area indicates that conditions found there are unsuitable for the species [33]. However, by definition, the distribution of an invasive population may not have reached equilibrium within the landscape [34,35]. The absence of an invasive species from a particular location may not indicate unsuitable conditions, but rather that the species simply has not been introduced or dispersed into that area. These characteristics of invasive populations necessitate the use of ‘presence-only’ species distribution modelling techniques [36]. Maximum entropy (Maxent) modelling is a machinelearning-based method for predicting species geographical distributions from presence-only data [37]. Several comparative studies of species distribution modelling methods have ranked Maxent among the top modelling approaches [38 –41]. With Maxent, the true distribution of a species is estimated as a probability distribution across all sites within the study area. The probability distribution adheres to a set of constraints derived from the presence data while maximizing entropy. The maximum entropy distribution is that which draws the least inferences beyond the available information [37]. With species distribution modelling, the set of constraints are functions of environmental variables. That is, the mean environmental conditions predicted by the model should be close to the conditions observed at presence locations [42,43]. Maxent begins with a uniform probability distribution and repeatedly adjusts the weights of features


N

4


perturbed FIA plot locations and Ailanthus presences

Phil. Trans. R. Soc. B 369: 20130192

A.T. centreline HUC-10 shell Ailanthus presence FIA plot location 0

75

150

300

450

600 km

Figure 2. Forest inventory and analysis (FIA) plots located within the A.T.-shell. Between 2002 and 2010, 3926 plots were surveyed and Ailanthus was observed at 136 locations. A photo of a female Ailanthus specimen laden with seeds is inset (retrieved from http://upload.wikimedia.org/wikipedia/commons/8/84/ Ailanthus_altissima_RJB.jpg). (Online version in colour.) GIMMS data have been corrected for calibration, view geometry, volcanic aerosols and other effects not related to vegetation change [52 – 54]. GIMMS data were fitted yearly with a Gaussian function to generate smoothed data for each one of the 25 years. We used monthly composites of GIMMS data for the A.T.-shell and the segments within the province level of ecological subregions to calculate the LSP metrics. SOGS is a climate gridding system that uses measurements of precipitation and maximum, minimum, and dewpoint temperatures from meteorological stations to create spatially continuous surfaces for air temperatures, precipitation, vapour pressure deficits and incident radiation [52]. To create a gridded climate surface, the automated TOPS data acquisition system retrieves the necessary weather data for all available weather stations in the region of interest. The data are then pre-processed and converted into data structures suitable for processing by the SOGS [2]. The GIMMS and the SOGS data were prepared and packaged in HDF5 format. Each dataset included subdatasets, and the

subdatasets included embedded data layers. We first converted the HDF5 data into common GIS data formats. We divided the data into three segments corresponding to boundaries of ecological subregion provinces (figure 1). We adopted the asymmetrical Gaussian approach for fitting Normalized Difference Vegetation Index (NDVI) data. We calculated the SOS and EOS for each year and obtained LOS as the difference between SOS and EOS in each grid point. We applied the non-parametric Mann – Kendall trend tests for improvement in trend identification. We used the Sen’s slope for obtaining the average of the magnitude and the image time-series for the trend analysis [55].

(d) Climate data and forecasts Data from TOPS provided baseline and projected climate data downscaled from an ensemble of 16 individual general circulation models (GCMs). The GCMs are a component of the Intergovernmental Panel on Climate Change (IPCC) Fifth Assessment


(a)

5

change in annual precipitation

(b)

precipitation (mm) high: +384.6

low: +3.5

low: +70.6

N

0

75

150

baseline: 1950–2005 | projection: CMIP5 RCP6.0 2095–2099

300 km

Figure 3. Distribution of changes in (a) annual mean temperature and (b) annual precipitation from the 1950– 2005 baseline to 2090– 2095 based on the ensemble average of the downscaled CMIP5 climate projections for RCP6.0. (Online version in colour.)

Report (AR5) Coupled Model Intercomparison Project Phase 5 (CMIP5) [56] and predict future conditions under a set of alternative scenarios defined by representative concentration pathways (RCPs). RCPs provide representative pathways of greenhouse gas concentrations with varying global emission rates and levels of mitigation effort. Each RCP included in the CMIP5 experiments is named for the radiative forcing equivalent (in watts per square metre) for the RCP greenhouse gas concentration in the year 2100 [57]. RCP6.0 was selected for Ailanthus modelling, as it represents a moderate increase in radiative forcing that stabilizes by 2100 owing to technologies and strategies for reducing greenhouse gas emissions. The broad-scale GCM data were downscaled to 800 m by interpolating historical climate observations and were corrected for local topography [58]. The data were then subset to the A.T.-shell for two time periods, a 1950 – 2005 baseline and projections for 2090– 2095 (figure 3). Multidimensional data for the ensemble average of monthly maximum temperature, minimum temperature and

precipitation were created, with an individual band for each month. In addition to annual mean temperature and precipitation, a set of 19 bioclimatic variables were derived reflecting annual trends, seasonality and extreme or limiting environmental factors (table 1). The bioclimatic variables were calculated for both the current and projected sets of climate data using the ‘biovars’ function of the R package ‘dismo’ [59]. The variables are intended to provide more biologically meaningful measures of conditions that are likely to restrict the range of Ailanthus.

(e) Ancillary geospatial data Topographic information within the A.T.-shell was adopted from the National Elevation Dataset (NED), a 30 m resolution digital elevation model (DEM) produced by the USGS with seamless coverage across the conterminous United States. The NED was compiled from the best publically available elevation data and

Phil. Trans. R. Soc. B 369: 20130192

temperature (°C) high: +4.7


change in annual mean temperature


Table 1. List of environmental variables evaluated for habitat modelling.

6

units

source

agdist

distance to agricultural landcover

m

NLCD06

agsum bio1

sum of agricultural landcover annual mean temperature

% 8C

NLCD06 TOPS

bio2 bio3

mean diurnal range (mean of monthly (max temp2min temp)) isothermality (bio2/bio7) (100)

8C %

TOPS TOPS

bio4 bio5

temperature seasonality (standard deviation100) max temperature of warmest month

1 10238C 8C

TOPS TOPS

bio6

min temperature of coldest month

8C

TOPS

bio7 bio8

temperature annual range (bio5 – bio6) mean temperature of wettest quarter

8C 8C

TOPS TOPS

bio9 bio10

mean temperature of driest quarter mean temperature of warmest quarter

8C 8C

TOPS TOPS

bio11

mean temperature of coldest quarter

8C

TOPS 24

bio12 bio13

annual precipitation precipitation of wettest month

1 10 m 1 1024 m

TOPS TOPS

bio14 bio15

precipitation of driest month precipitation seasonality (coefficient of variation)

1 1024 m %

TOPS TOPS

bio16 bio17

precipitation of wettest quarter precipitation of driest quarter

1 1024 m 1 1024 m

TOPS TOPS

bio18

precipitation of warmest quarter

1 1024 m

TOPS

24

bio19 cti

precipitation of coldest quarter compound topographic index

1 10 n.a.

dem devdist

elevation distance to developed landcover

m m

DEM NLCD06

devsum

sum of developed landcover

%

NLCD06

hli lfcc

heat load index canopy cover mean

n.a. %

DEM NLCD01

lfcc_min lfcc_std

canopy cover minimum canopy cover standard deviation

% %

NLCD01 NLCD01

nlcd_wet

wetland and open water landcover

0–1

NLCD06

slope slopepos

slope slope position (800 m radius focal mean elevation – elevation)

8 m

DEM DEM

soil_drain soil_fldfreq

soil drainage class flood frequency

0–7 0–4

USSOILS USSOILS

soil_hydric trasp

hydric soils topographic radiation aspect index

0–1 0–1

USSOILS DEM

undergoes rigorous accuracy assessments [60]. Individual tiles spanning the study area were acquired from, mosaicked and clipped to the boundary of the A.T.-shell. Additional variables were derived from the NED to further characterize topography [61]. Slope calculates the maximum rate of change in elevation from the focal raster cell to its neighbours within an 800 m radius. Slope position subtracts a focal mean of elevation from the original elevation raster. The compound topographic index (cti) is a steady-state wetness index and is a function of slope and upstream contributing area. The topographic radiation aspect index (trasp) transforms circular aspect into a continuous variable better suited for modelling. Cooler and wetter north – northeast orientations are assigned

m

TOPS DEM

values close to zero, whereas hotter and dryer south– southwest orientations are closer to unity. The heat load index (hli) is similar to TRASP, but also accounts for slope steepness (table 1). Land-cover classes were extracted from the 2006 National Land Cover Database (NLCD) [62], including developed, agricultural, wetland and open water areas. Layers were also generated measuring the distance from each pixel to the nearest agricultural and developed feature, respectively. In addition, a layer for percentage tree canopy was used to examine the shade intolerance of Ailanthus. Finally, data layers for soil drainage class, flood frequency and hydric soils were extracted from the USSOILS dataset [63]. USSOILS is a raster coverage derived from the State Soil Geographic Database (STATSGO) by the USGS (table 1).

Phil. Trans. R. Soc. B 369: 20130192

title


variable


(f ) Maxent model data preparation

(ii) Model evaluation

(g) Maximum entropy distribution modelling The predictive performance of Maxent is influenced by the selection of environmental variables, the feature algorithms fitted to them and the regularization constants used to control over-fitting. While complex models may accurately predict populations at equilibrium, simpler models are likely better suited for range shifting populations [44,64]. The default Maxent settings for feature classes and regularization are adapted from a study that tuned parameters based on datasets for 226 species across six regions, and have been shown perform to well across a wide range of applications [40,42]. These default settings were retained with the exception that feature classes were limited to hinge features. Hinge features form piece-wise linear functions and have been shown to improve model performance [42], while providing simple, smooth models appropriate for predicting the projected distribution of range shifting species [44].

(i) Variable selection For an initial assessment of Ailanthus habitat characteristics, FIA plot locations within the A.T.-shell were used to append values from the collection of environmental variables to the point data attributes. Histograms were constructed to compare the distributions of Ailanthus presence and absence points across the sets of FIA attributes and environmental variables. Incorporating a large number of variables into a Maxent model may lead to complex solutions that obscure important ecological relationships, resulting in unexpected or erroneous behaviour when extrapolating the model to future conditions [43,65]. Pearson correlation coefficients were calculated for every pairwise combination of topographic and land-cover environmental variables as well as bioclimatic variables using ENMtools [66]. The number of variables was reduced iteratively by evaluating an initial model incorporating many variables, eliminating variables with weak or inapt contributions and running the revised model. An extensive array of tools was used to evaluate variable performance, including variable response curves, percentage contribution and permutation, jackknifing as well as Pearson correlation coefficients. Marginal variable response curves plot the change in logistic prediction from varying the value of one environmental variable while holding all other variables constant at their average sample value. Strongly correlated variables may confound the interpretation of marginal response curves, as the actual Maxent model can incorporate features where variables change together. Isolated variable response curves represent a model incorporating

For each candidate set of variables, a 10-fold (replicate) crossvalidated model was generated with 122 of the Ailanthus presence points used for model training and the remaining 14 for testing. As with variable selection, a variety of methods were used to assess model performance. Models were evaluated based on their performance on test data, parameter complexity, ecological consistency and degree of extrapolation required when projecting to future conditions. To compare model performance, the Maxent package determines the area under the curve (AUC) of the receiver operating characteristics (ROC) for each model. The ROC curve is constructed by plotting model sensitivity and specificity. Sensitivity is a function of the omission rate, i.e. the rate training or test presence points incorrectly classified as unsuitable. Specificity is typically the commission rate, or rate of absence points incorrectly classified as suitable. However, given the lack of absence data for presenceonly modelling, specificity is instead derived from the fraction of the study area predicted as suitable [68]. While the application of AUC to presence-only modelling is not without limitations [70], it provides a threshold-independent measure of model predictive performance on withheld test data. In addition, comparing predictive performance on training versus test data, as well as the standard deviation of scores across replicates, provides an indication of model transferability. A model with high training, but low test AUC, is likely overfit to the training data, and may perform poorly when extrapolated to new environmental space (e.g. climate projections) [65]. Model complexity was assessed using sample-size-corrected Akaike information criteria (AICc), as proposed by Warren & Seiftert [65] and implemented within ENMtools [66]. AICc is determined from model log-likelihood (the product of suitability scores across all presence points) penalized by the number of parameters (the complexity of features applied to the environmental variables). Models with lower AICc scores balance high predictive performance with low complexity, and are likely more appropriate for extrapolating to future conditions. Hinge features are penalized more heavily than other feature classes as they incorporate more parameters. Therefore, AICc scores were compared only between models including the same selection of feature classes [65].

(iii) Projecting suitability and examining trends Once a model was selected, TOPS downscaled CMIP5 climate scenario data were substituted for current climate variables, and suitability was recalculated. When projecting the model to

7

Phil. Trans. R. Soc. B 369: 20130192

only the focal variable and may be more informative when dealing with highly correlated variables [37]. The shape of the response curves is highly informative. Variables with complex surfaces may indicate over-fitting, whereas sharp increases or decreases near the limits of the environmental range sampled increase uncertainty when extrapolating the model to projected conditions. Finally, the appropriateness of the response curve shape should be considered in the context of the ecological understanding of Ailanthus [30,44,67]. The percentage contribution and permutation importance of each variable are provided with the Maxent model output. Maxent also evaluates the set of input environmental variables by performing jackknife tests. Variables may perform well in isolation, but make little difference on the overall model prediction, indicating that they are good predictors but contain little information not present in the other variables. In other cases, jackknife tests may indicate that withholding the variable actually increases model performance. Finally, a variable that performs well for the training data but badly with the test data, is poor at generalizing and therefore less transferable [43,68]. Additional details of model fitting in this study are provided in reference [69].


Prior to Maxent modelling, all environmental data layers were pre-processed to conform to a uniform spatial extent, resolution and geographical projection, and converted into grid format and projected to the Albers equal-area conic projection. The data layers were clipped using the A.T.-shell boundary with a 1.6 km buffer. Retaining data within a buffer around the shell circumvents distortions caused by edge effects near the boundary of the study area. Operations were performed at this stage to derive additional layers (e.g. hli from elevation) and transform layers (e.g. natural log of distance to development). The data layers were then resampled to a 300 m cell size using a snap raster template to ensure that cell alignment agreed perfectly between data layers. Bilinear interpolation was used for downscaling continuous variables with resolutions coarser than 300 m. The layers were then adjusted to reflect the uncertainty introduced by FIA plot location perturbation by calculating focal statistics with a moving window radius of 800 m. For continuous variables, the mean value of cells within 800 m of the focal cell was calculated, and for categorical variables, the total counts for each category of interest were tallied. Finally, the data layers were clipped to the A.T.-shell and exported to grid format.


cumulative threshold þ 1.6 fractional predicted area. The threshold was applied to both current and projected distributions by reclassifying the two raster files, and their total suitable area, mean latitude and mean elevation were calculated. The previous metrics were also calculated within each ecological province intersecting the overall A.T.-shell, facilitating the comparison of regions with similar conditions across time.

3. Results

(c) Model selection and performance

(a) Trends in land surface phenology

Model ‘4bio_4topo’ was selected as the highest performer from an array of 15 models incorporating alternative sets of environmental variables (table 3). This model had the lowest AICc score (3707.5), a high mean test AUC (0.85) and low test AUC standard deviation (0.034) between replicates. The model’s AICc score outranked models incorporating both more and fewer variables, suggesting its level of complexity was nearest to optimal, whereas the low standard deviation indicates a robust model with high transferability. The model incorporated a limited set of environmental variables with clear ecological interpretations (table 4). Mean temperature of coldest quarter (bio11) made the largest contribution to the model (40.1%), with a unimodal isolated variable response curve indicating high suitability for sites with warmer winters, but decreasing rapidly again within the humid subtropical conditions of the trail’s southern extremes (figure 4). Temperature seasonality (bio4) contributed 27.1% of regularized gain, with a response curve reflecting a preference for the moderate seasonal variation over the seasonal extremes of the northern A.T. or the steady warmth of to the south (figure 4). Mean temperature of wettest quarter (bio8) made the third largest contribution (22.7%) with a marginal response curve similar to bio11, suitability increasing with temperature with a sharp decrease at the upper extreme of temperatures sampled (figure 4). Slope and aspect (trasp) contributed 4.6% and 3.1% of regularized gain, respectively, with a preference for moderate slopes and drier, sunnier aspects (figure 4). While the remaining variables appear to have contributed very little (nlcd_wet ¼ 0.9%, bio19 ¼ 0.8%, dem ¼ 0.6%), their importance may have been obscured owing to correlation with other variables.

We obtained time-series data to illustrate spatial dynamics of LSP metrics between 1982 and 2006. The results indicate that the trend of SOS delayed 0.99 days within the A.T.-shell area in the Adirondack–New England Mixed Forest–Coniferous Forest–Alpine Meadow Province (M211) from 1982 to 2006. On the other hand, the trend of EOS was delayed by 13.2 days. The most extended LOS occurred in the M211 province by 12.1 days during the 25 years. The LOS in the M211 province was 19 days shorter in 1988 and 27 days longer in 1998 than the average. The trend of SOS delayed 0.70 days within the A.T.-shell in the Eastern Broadleaf Forest Province (221) between 1982 and 2006. The extreme late EOS occurred in 1986, 1990, 1994, 2002 and 2005. The LOS trend extended 8.99 days during the 25 years. The trends of SOS delayed 1.71 days within the Central Appalachian Broadleaf Forest–Coniferous Forest–Meadow Province (M221) from 1982 to 2006. The trend of EOS delayed 5.62 days during the time period. The extreme late EOS occurred in 1984, 1988 and 2006. The least change in LOS occurred in the southern segment in this section of the A.T.-shell and extended 3.31 days during the 25 years. This segment-based LSP trend analysis provides background information for understanding of the habitat suitability and latitudinal variation within the study area.

(b) Representative environmental variables Comparing conditions throughout the study area with the subset of Ailanthus presence points revealed patterns that reflected the habitat preferences of Ailanthus as reported in the literature. From the FIA survey plot attributes, Ailanthus was generally observed at sites with lower elevations and moderate soil moisture, closer to roadways and in younger forest stands. A similar analysis of the environmental variables further illustrated a preference for lower elevations and mesic topography as well as warmer and dryer climates. The frequently reported association of Ailanthus with development, agriculture and canopy gaps was evident from the analysis of land-cover variables, though their contributions were minor within subsequent Maxent modelling.

(d) Predicted trends in invasive pressure Bioclimatic variables calculated from TOPS downscaled CMIP5 RCP6.0 climate scenarios data were substituted for current climate variables, and suitability was recalculated. From the 1950–2005 baseline to the 2095–2099 projection, mean temperature seasonality (bio4) increased by 4 10238C, mean

8

Phil. Trans. R. Soc. B 369: 20130192

Bt ¼ minimize 6 training omission rate þ 0.04

Nine highly correlated (r . 0.90) bioclimatic variables were removed (table 2a) from consideration, whereas no topographic or land-cover variables were eliminated (table 2b). Strongly performing or ecologically limiting variables were selected over generalized or erratically performing variables. Notable exceptions were the mean temperature of coldest quarter (bio11) and temperature seasonality (bio4). Both variables were retained despite high correlation (r ¼ 0.91) owing to the ecological significance of bio11 and exceptional performance of bio4 in preliminary jackknife tests. Maxent consistently selected climate variables as the most important predictors of suitability. Mean temperature of coldest quarter (bio11) ranked highest for percentage contribution and permutation importance across model runs. Jackknife tests for temperature seasonality (bio4) had the highest gain when used in isolation and the largest decrease in gain when omitted, indicating that seasonality contains information both most useful by itself and not present in other variables. Conversely, some variables reduced test gain, such as the CTI and distance to development, and were removed from subsequent models.


future conditions, variable ‘clamping’ was used to restrict the values of projected variables within the range of values encountered while training the model under current conditions. In addition to the projected distribution, Maxent provides outputs for evaluating divergence of current and projected variables as well as the influence of variable clamping. A threshold must be applied to the continuous probability distribution to provide a binary map of suitable and unsuitable locations. The Maxent output includes several common thresholds and their omission rates. Of these, ‘balance training omission, predicted area and threshold value’, henceforth, the balance threshold (Bt) provided a compromise between over-fitting and over-commissioning:

bio1

0.78

hli

lfcc_mean

lfcc_min

0.28

lfcc_std

0.87

0.84

*

(b)

soil_hydric

soil_fldfreq

soil_drain

slopepos

slope

nlcd_wet

lfcc_std

lfcc_min

lfcc_mean

hli

devsum

devdist

dem

cti

agsum

agdist

trasp

0.00

0.09

20.21

0.23

0.01

0.02

20.23 20.03

20.28

20.38 0.16

0.54

20.01

20.01

0.04

devdist

dem

20.34 0.46

0.20 20.44 0.43

0.12 20.20

0.66

0.25

0.43

20.39

20.23

0.43

20.44

0.50

20.43

0.08

20.02

20.50

20.78

0.09

20.02

0.17

20.29

20.02

0.01

20.23

0.03

20.14

0.01

20.18

20.43

0.07

20.18 20.17

20.29 20.37

20.04

0.08

20.20

0.39

20.09

0.16

0.00

nlcd_wet

0.18

20.38

20.40

0.16

0.31

20.27

0.01

20.46 20.15

20.42 20.21

0.62

0.43

0.40

0.45

20.21

0.22

0.77

20.37

20.27

0.11

0.00

slopepos

*

0.76

0.85

0.02

20.21

0.15

20.56

20.24

0.03

20.28

20.20

0.07

20.17

20.07

bio14

0.55

0.57

0.74

20.25

0.15

0.68

20.49

20.40

20.04

20.03

slope

*

*

*

0.97

0.25

20.09

20.11 0.21

0.51

20.55

20.46

0.22

20.18

20.48

0.43

0.14

0.13

bio13

0.44

20.55

* devsum

0.61

20.43

0.20

20.20

20.44

0.36

0.07

0.10

bio12

bio19

cti

0.45

20.06

20.88

0.99

0.80

20.91

0.85

0.44

0.98

bio11

*

*

20.59

20.81

0.88

0.98

20.14

0.53 0.81

0.45 0.32

20.45 20.90

20.60

0.58

0.35

0.95

bio10

*

*

0.79

20.88

0.88

0.53

0.78

bio9

20.10

0.05

20.03

0.35

bio8

0.96

20.79

20.18

20.80

bio7

bio18

*

*

0.80 20.90

0.54

20.50

20.92

0.98 0.34

0.89

0.44

20.83

bio6

20.43

bio5

bio17

agsum

*

0.74

bio4

*

agdist

0.42

bio3

bio16

bio15

bio14

bio13

trasp

*

bio11

bio12

*

*

bio10

bio9

bio8

bio7

bio6

bio5

bio4

bio3

bio2

bio2

0.92

0.89

0.13 20.23 20.13 0.74

0.17 0.15 0.51

20.19 20.03

0.12

20.12

20.13 20.37

0.02 20.07 0.14

20.18 0.12 0.25

0.28 20.06 0.00

0.13 20.06

0.11 20.26

20.31

0.43 20.09 0.00

0.00

0.01 20.13

soil_hydric

soil_fldfreq

*

20.06

20.10

20.05

0.06

20.15

0.01

0.14

20.19

20.01

20.01

0.22

0.01

soil_drain

0.19

20.21 0.24

*

0.72

0.85

0.91

0.93

0.18

0.95

0.16

0.27

20.16

0.75

20.13

20.07

0.38

20.35

20.40

0.34

0.52

0.89

20.56

20.52

20.43

0.96

20.38

20.47

0.17

20.25

0.99

0.15

0.24

0.98

20.21

20.15

20.43

0.35

0.26 20.37

0.44 20.50

0.06 0.05

0.05 20.02

0.15 0.15

bio18

bio17

bio16

0.24

0.05

0.24

0.12

0.41

0.09

20.22

0.20

0.09

20.30

0.40

0.36

0.21

bio15

Phil. Trans. R. Soc. B 369: 20130192

bio1


(a)

0.85

0.89

0.97

0.16

0.75

0.95

0.96

0.33

0.00

0.61

20.58

20.50

0.30

20.08

20.54

0.51

0.23

0.22

bio19

Table 2. Matrices of Pearson’s correlation coefficients calculated for (a) topographic and landcover variables and (b) bioclimatic variables. Asterisks indicate highly correlated (r . 0.90) pairs. Variables in italics were included in the final distribution model. See table 1 for variable descriptions.


9


climate variables

model

topographic variables

landcover variables

loglikelihood

parameters

AICc score

mean test AUC

AUC SD

4

4

0

21751.24

56.8

3707.533

0.85

0.034

2bio_1topo 4bioalt_4topo_2lc

2 4

1 4

0 2

21796.72 21749.24

40.9 61.8

3713.173 3733.685

0.812 0.847

0.035 0.034

5bioalt_5topo 10bioalt_6topo

5 10

5 6

0 0

21747.56 21735.15

68.2 72.2

3777.135 3788.426

0.848 0.855

0.045 0.047


5 4

4 3

0 0

21747.74 21752.98

69.8 67.8

3794.235 3796.161

0.849 0.85

0.044 0.041

5bioalt2_3topo

5

3

0

21750.84

69.4

3804.34

0.852

0.04

6bioalt_6topo allbio_alltopo

6 19

6 10

0 0

21743.5 21732.14

73.1 76.2

3817.243 3821.316

0.848 0.847

0.046 0.039

5bio_5topo 10bio_5topo

5 10

5 5

0 0

21751.37 21739.94

74.4 79.4

3862.39 3886.28

0.848 0.851

0.045 0.048

4bio_5topo

4

5

0

21752.8

76.9

3886.495

0.842

0.046

10 19

5 10

4 7

21734.39 21722.52

88.3 92.5

3997.353 4079.972

0.844 0.844

0.046 0.049

10bio_5topo_4lc allbio_alltopo_alllc

Table 4. The relative contributions of the environmental variables for the final model (4bio_4topo). Percentage contribution is calculated as the training algorithm iterates by adding the increase in regularized gain or subtracting if the absolute value of lambda is negative. Permutation importance is determined by randomly changing the values of the focal variable for the training and background data, revaluating the model with each permutated variable in turn and recording the corresponding drop in training AUC normalized to a percentage. Interpretation of variable contributions may be confounded by high correlation. See table 1 for variable descriptions. variable

contribution (%)

permutation importance

bio11

40.1

34.7

bio4 bio8

27.1 22.7

43.9 6

slope

4.6

5.1

trasp nlcd_wet

3.1 0.9

2 0.7

bio19 dem

0.8 0.6

2.4 5.3

temperature of coldest quarter (bio8) increased by 28C, mean temperature of wettest quarter (bio11) increased by 48C and mean precipitation of coldest quarter (bio19) increased by 532 1024 m (table 5). The multivariate similarity surface (MESS) revealed that the projected variables were outside the range encountered during training in three regions. Mean temperature of coldest quarter (bio8) was the most dissimilar

variable across the New Jersey and Virginia trail sections, whereas mean temperature of wettest quarter (bio11) was most dissimilar in the southern section. If clamping had not restricted projected variables to the range of values encountered during training, then the model would predict a dramatic decrease in suitability throughout the mid-Atlantic [69]. The Bt threshold value provided by Maxent was used to reclassify the continuous raster surfaces into two discrete suitability classes (figure 5). All cells with a logistic probability greater than the threshold value of 0.047 were predicted as suitable, and all cells below unsuitable. For the current distribution, this produced a fractional predicted area of 0.560 and a training omission rate of zero. The threshold was also applied to the projected distribution, and a change map was created to visualize shifts in suitability. The Maxent model of current conditions predicts that 60 044 km2, or 56%, of the A.T.-shell is potentially suitable for Ailanthus colonization. By 2095, the suitable areas are projected to expand to 89 066 km2 (82%), an increase of 48% (figure 5). Ecoregion province M211—Adirondack–New England Mixed Forest exhibited the most dramatic increase, from 2% to 50% total area. Suitable area also increased in provinces 221— Eastern Broadleaf Forest and M221—Central Appalachian Broadleaf Forest by þ15 and þ25%, respectively (table 6). The significance of any trends observed in provinces 211 and 231 is limited due to the small portion of the A.T.-shell they encompass (7% of A.T.-shell, combined). The mean elevation of suitable areas is projected to increase by 59 m (from 391 to 449 m) and the mean latitude to shift north by 49 km. Elevation increased in province M211 by 28% (96 m). Ranges shifted north in provinces M211 (108 km) and 221 (36 km), and south in M221 (61 km; table 4).

Phil. Trans. R. Soc. B 369: 20130192

4bio_4topo

10


Table 3. Parameters of candidate Maxent models and the metrics used to evaluate their performance. Sample-size-corrected Akaike information criteria (AICc) is determined from model log-likelihood (the product of suitability scores across all presence points) penalized by the number of parameters (the complexity of features applied to the environmental variables). The AUC of the ROC provides a threshold-independent measure of model predictive performance on withheld test data. The values listed below are averaged across the 10 replications performed for each model. Models are titled based on the composition of variables they incorporate; see appendix I for a detailed description of each model.


4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0 –8

–6 –4 –2 bio11_300 m

0

2

4

6

8

response of Ailanthus to bio8_300 m

600

700

900 bio4_300 m

1000

1100

1200

response of Ailanthus to slope_800 mn_300 m_atshell

(d)

6.0 5.5 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0

800

1.1 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 –5

(e)

0

5

10 bio8_300 m

15

20

25

–2

0

(f)

response of Ailanthus to trasp_800 mn_300 m_atshell

2

4

6 8 10 12 14 16 slope_800 mn_300 m_atshell

18

20

22

24

1.0

1.1

response of Ailanthus to nlcd_wet_300 m_atshell

raw output (×10–4)

1.3 1.2

1.0

1.1

0.9

1.0

0.8

0.9

0.7

0.8 0.7

0.6

0.6

0.5

0.5

0.4

0.4

0.3

0.3 –0.1

0

(g)

0.1

0.2

0.3 0.4 0.5 0.6 0.7 0.8 trasp_800 mn_300 m_atshell

0.9

1.0

1.1

0

(h)

response of Ailanthus to bio19_300 m 1.4

0.1

0.2

0.3 0.4 0.5 0.6 0.7 nlcd_wet_300 m_atshell

0.8

0.9

response of Ailanthus to dem_300 m_atshell 1.4

1.2 raw output (×10–4)

–0.1

1.2

1.0

1.0

0.8

0.8

0.6 0.6 0.4 0.4 0.2 0.2 0 1000

2000

3000

4000

5000 6000 bio19_300 m

7000

8000

9000

0

500

1000 dem_300 m_atshell

1500

2000

Figure 4. Isolated variable response curves plot the change in logistic prediction for a model incorporating only the focal variable. See table 1 for variable descriptions and units. (Online version in colour.)

4. Discussion (a) Increasing pressure on regional biodiversity The distribution of suitable habitats (figure 5) largely coincides with the existing knowledge of Ailanthus distribution within

the eastern United States [71,72]. The majority of locations with high suitability fall within the Virginian and mid-Atlantic sections of the A.T. which is expected given the distribution of FIA presence points. Conditions in these regions are ideal, with moderate-to-low rainfall, low elevations, mild winters

Phil. Trans. R. Soc. B 369: 20130192


11

response of Ailanthus to bio4_300 m 5.0

–16 –14 –12 –10

(c)

(b)

response of Ailanthus to bio11_300 m 2.4 2.2 2.0 1.8 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0



(a)

532 3286 894 2479 2038 bio19 (1 1024 m)

441

9310

9679

369

850

244

3818

2 4 17 2 15 22 2 21 7 5 26 29 29 214 bio8 (8C) bio11 (8C)

3 5

24 6

28 9

4 3

8 4

4 901 897 229 111 707 665 bio4 (1 10238C)

42

1143

1100

243

82

change RCP6.0 current current current variable

RCP6.0

change

current

RCP6.0

change

RCP6.0

change

mean s.d. max. min.

(b) Characterizing Ailanthus habitats (i) Climate The performance of various environmental variables assessed throughout Maxent modelling (table 3) indicates that the distribution of suitable Ailanthus habitats is primarily constrained by climate conditions at a regional scale. The mean temperature of the coldest quarter (bio11) and temperature seasonality (bio4) were particularly significant (table 4), with a preference for warmer and milder conditions (figure 4). Ailanthus saplings are highly vulnerable to frost mortality [13,17] and annual die-backs may restrict occurrence to lower elevations and warmer regions [73], a limiting factor that the marginal response curve for bio11 appears to reflect (figure 4). Kowarik & Saumel [17] also note that while Ailanthus tolerates a wide range of climatic conditions, temperature seasonality strongly affects growth, dispersal and survival. Variables relating temperature to precipitation also performed well. While annual mean temperature (bio1) and annual precipitation (bio12) performed well during preliminary modelling, the more nuanced climate variables were more capable of capturing the extreme factors that limit the success of Ailanthus. Ailanthus’s preference for high mean temperatures during the wettest quarter (bio8; figure 4) may indicate increased mortality owing to frost stress and mechanical damage associated with winter storms [74], or it may simply reflect a broader preference for warmer climates. While there are conflicting accounts of Ailanthus flood vulnerability [13,75], its exceptional drought tolerance is well established [17,76]. This trait is evident within the response curve for the

12

Phil. Trans. R. Soc. B 369: 20130192

and abundant development. Suitability decreases as the trail moves south into the Smoky Mountains and elevation and precipitation increase, and development thins. To the north, suitability again decreases as elevation increases and temperature drops. The northeast is predicted to contain the least suitable areas along the A.T. While Ailanthus invasions are reported throughout the northeast [11,71,72], and are historically abundant in the New England region, these records predominately occur within the low elevations and dense population centres along the Atlantic seaboard, rather than the remote, mountainous regions through which the A.T. passes. The increase in suitability predicted as the A.T. leaves the Kittatinny Mountains in New Jersey and approaches the New York City metro area is supportive of this conclusion. Estimating the future distribution of suitable Ailanthus habitats by integrating climate projections (figure 5) reveals several interesting trends. Overall, there is a 48% increase in suitable area, representing a dramatic increase in the potential extent of Ailanthus invasion. Subdividing the A.T.-shell by ecological province delineates the area into units with similar environmental conditions and ecological communities, providing insights into the processes driving this expansion. The most dramatic increase occurs in the A.T. section within the Adirondack–New England Mixed Forest–Coniferous Forest–Alpine Meadow ecological province, where warmer temperatures allow Ailanthus to expand north with a 49 km increase in mean latitude as well as to higher elevations. Conversely, the average latitude in the Central Appalachian Broadleaf Forest–Coniferous Forest– Meadow ecological province actually shifts south, and average elevation increases, as Ailanthus migrates into the Great Smoky Mountains (table 6).


Table 5. Change from the 1950– 2005 baseline to the projection for 2090– 2095 using the bioclimatic variables used in the final model and derived from the ensemble average of the downscaled CMIP5 climate projections for RCP6.0. See table 1 for variable descriptions.



(c)

suitability high

low

N

0

100 200

400 km

low suitability currently suitable predicted suitable

Figure 5. Continuous habitat suitability distributions for the (a) 1950– 2005 baseline and (b) 2090– 2095 ensemble average of the downscaled CMIP5 climate projections for RCP6.0. A threshold value was applied to derive binary suitability classes and map the changes in suitable habitats (c). (Online version in colour.) rainfall of the coldest quarter (bio19; figure 4) and appears to support the frequently reported preference for drier soils.

(ii) Topography Topographic variables were also included in the final distribution model (table 4). Slope was a significant factor, with suitability being highest at moderate gradients (figure 4). An extensive root system allows Ailanthus to colonize rough terrain and steep slopes [73]. These extreme areas may coincide with decreases in canopy density and increased access to direct sunlight, a primary Ailanthus habitat requirement [17]. However, it should be noted that the maximum slope was only 268 owing to the variable’s 300 m pixel resolution, and the adjustments made to reflect FIA plot location perturbation. The heat load index (hli) was discarded in favour of the topographic radiation aspect index (trasp), as hli’s incorporation of slope was inappropriate given the inclusion of slope as a separate variable. The variable response curves for trasp indicate that Ailanthus prefers the sun exposure (figure 4), and therefore increased temperature and decreased humidity, of the south– southwest-facing slopes. Suitability decreased in the presence of wetlands (nlcd_wet; figure 4), further indicating a preference for dryer sites. While compound topographic index (cti) was theoretically well suited to identify wet and dry positions in the landscape, its importance was diminished by the coarse resolution of the analysis and the locational fuzzing of the FIA data. The response curve for elevation clearly reflects Ailanthus’s characteristic association with low elevation, mild, heavily

developed areas (figure 4). While elevation made the lowest contribution, it had the fourth highest permutation importance. Its influence was likely diminished owing to correlation with other variables better suited to discriminate the underlying mechanisms disrupting Ailanthus establishment, such as low temperatures and frost mortality.

(iii) Land cover Land-cover variables, while potentially significant, proved to be difficult to incorporate into the model. The association between Ailanthus and urban areas and canopy cover is very prevalent throughout the literature [11,13,77] and apparent from the analysis of FIA plot data. Plotting an isolated variable response curve illustrates the relationship: suitability decreases exponentially as the distance to development increases. However, when distance to development (devdist) and canopy cover standard deviation (lfcc_std) were added to the final model, they each only made an infinitesimal contribution of 0.6%. Furthermore, the overall test AUC of the model decreased when the two land-cover variables were included. One factor suppressing the importance of land-cover variables may be correlation with bioclimatic and topographic variables. The regions within the A.T.-shell furthest removed from urban areas also contain some of its most extreme conditions. Maine to the north and the Smoky Mountains to the south contain remote areas, but are also at very high elevations with low temperatures and high rainfall, respectively, as well as large tracts of forest. In other words, the broad spatial extent of the

Phil. Trans. R. Soc. B 369: 20130192

A.T. centreline HUC-10 shell Ailanthus presence Ailanthus presence (beyond shell)

13

suitability change

Maxent model: 4bio_4topo_hinge—projection: CMIP5 RCP6.0 2095–2099

projected suitability (2095–2099)

(b)


current suitability (1950–2005)

(a)

49

261 37.68

39.35 38.91

38.23 14.2

15.1 59 449 391

491 24.8

48.3 29 022

9750 96.3

82.4 89 066

49 098 77.1

55.6 60 044

561 39 348 51 004

10 8072

M221

A.T.-shell

70

2128 108 36.02 43.46 37.17 42.49 30.1 28.2 267 340 79.1 2298.3 1247 14 345 99.7 50.3 2824 14 969 55.7 2.1

289 195 84.8 15.1 1283 17 211

28.7 86.0

2372 19 802

53.0 98.9

1089 2591

% change % projected % current

348 436 1577 624 2831 29 746 231 M211

81 96

17 36 41.41 40.63 41.26 40.31 43.3 3.0 4478 20 013 211 221

414 201

current total area province

125 6

change (km) projected current % change mean elevation (m) suitable area (km2)

projected

(iv) Scale considerations One of spatial ecology’s fundamental quandaries is reconciling information obtained at disparate spatial scales into a common resolution for analysis [78 –81]. There is often a large degree of variation in the spatial resolutions of the datasets readily available for analysis. Model resolution is limited by the coarsest dataset, the locational uncertainty of FIA species occurrences in this instance, and predictor variables with finer grains (i.e. smaller pixel size) must be downscaled to coarser scales. A 300 m pixel size was selected for modelling as an optimal compromise between fine-grain landcover and topography data, coarse bioclimatic data, and the spatial uncertainty of FIA plot locations. Information is inevitably lost or altered with each manipulation, whether it be the intensity of development or the complexity of the wildland–urban interface [78], resulting in a distorted representation of the underlying regional patterns that the variable strives to reflect. Features within the landscape, such as transportation corridors or ridges, may be exaggerated or suppressed depending on the algorithms used [81]. The order in which operations are performed can have a dramatic effect; aggregating the urban areas to 300 m before calculating distance gives a substantially different result. Transforming predictor variables to new scales requires a clear understanding of the operations applied to the data as well as the ecological process that the output is attempting to characterize. However, scale issues may arise even when transformations are carefully managed to minimize distortions, as the underlying ecological processes may themselves be dependent on the scale of observation [79,80,82]. For instance, the observed patterns of light shift radically when moving from the perspective of a mite amid topsoil to a raptor circling high above the landscape. While the pattern of light can be accurately measured throughout the intervening scales, only a limited domain is relevant to the canopy cover processes that influence Ailanthus establishment [83]. A related consideration is the extent of the study area, as mentioned previously. Continental-scale distributions are typically driven by broad climatic patterns, but have little predictive power on localized models, where variations in topography and land-cover variables exert far more influence [28,36]. The poor performance of the finer-scaled predictors is likely due to a combination of these factors. The model’s coarse scale may lie beyond the domain where observed patterns reflect the ecological processes relevant to Ailanthus. The patterns of vegetation cover, soil saturation and human disturbance within one 300 m pixel can vary widely, and do so at scales that are likely to influence Ailanthus. Coupled with distortions from downscaling and FIA spatial error, it is unsurprising that incorporating land-cover variables into the final model decreased its overall performance. The Maxent distribution is fit to training data misrepresenting the actual conditions, referred to as forced-matching [84]. The drop in performance would likely be more pronounced if the extent were reduced while retaining the coarse grain. Uncertainty also arises from the model’s reliance on climate projections. While the methods used to downscale the climate data did attempt to correct for finer-scale gradients introduced by local topography [58], any residual effects of

14

Phil. Trans. R. Soc. B 369: 20130192

mean latitude

A.T.-shell contains regional-scale sociogeographical patterns that obscure potential finer-scale relationships between land cover and Ailanthus habitat suitability.


Table 6. Change in the area, elevation and latitude of suitable habitats from the 1950 – 2005 baseline to the projection for 2090 – 2095 derived from the ensemble average of the downscaled CMIP5 climate projections for RCP6.0. Ecoregion provinces in italics incorporated only a small portion of the trail and were not interpreted. See figure 1 for the locations of Bailey’s ecoregion provinces intersecting the study area.



This distribution model makes significant assumptions by relying primarily on climate data. While the importance of biological, cultural and topographic features is evident from the literature [73,77,87] and analysis of Ailanthus FIA records, these variables proved difficult to implement within the Maxent model for several reasons. Circumventing these issues by increasing the resolution of variables and acquiring locationally accurate plot records would allow the model to discriminate suitable patches within the broad regions predicted by this model. Unfortunately, model resolution is constrained by current hardware and software performance limitations ( particularly across a study area as expansive as the A.T.) and federal privacy regulations precluded the use of the true FIA plot locations. However, hierarchical modelling approaches may hold the key to integrating both smalland large-scale processes across broad spatial extents, and are currently an active topic of discussion [82,88,89]. Projections of the distribution model are similarly impaired by their lack of biological [90] and anthropogenic interactions. Land-cover change will affect Ailanthus dispersal pathways and establishment opportunities [91], but specific patterns are difficult to predict with certainty. Biological interactions, such as interspecific competition, may also limit invasion. However, the hardy traits of Ailanthus (e.g. rapid growth, high fecundity and robust environmental tolerances) suggest its competitive advantage over native species will only increase as climate change alters the frequency and magnitude of disturbances [92,93]. While incorporating biological interactions would augment this model, a climatic projection is suitable for an initial investigation of potential shifts in the distribution of suitable Ailanthus habitats. Bioclimatic variables with projected values extending beyond the range of current values encountered while training the Maxent model further complicate predictions. In these cases, the model must either extrapolate features beyond the range within which they were parametrized or ‘clamp’ the bioclimatic values and hold them constant at the upper limit of current conditions. Expanding the study area to incorporate a broader range of conditions would partially mitigate this issue, but only at the cost of decreasing the model’s ability to discriminate within the A.T.-shell [44]. Similarly, comparing Maxent distribution models with additional modelling techniques or constructing model ensembles is increasingly prevalent [94,95], and may provide valuable

(d) Remote sensing for biodiversity conservation: obstacles and opportunities The model projection indicates the extent of Ailanthus invasion is predicted to increase significantly as the climate changes. Mapping the distribution of suitable Ailanthus habitats facilitates a quantitative assessment of future range expansions into novel environments. This habitat suitability model successfully integrates the resources of the A.T.-DSS to select a set of environmental variables that define Ailanthus habitat suitability, map the current distribution of suitable habitats, and estimate future range expansions driven by climate change. These predictive modelling results complement the existing framework of the A.T.-DSS, which facilitates further investigation within areas of concern using TOPS remote sensing data products. The FIA database provided a systematic, abundant and detailed ground survey of Ailanthus populations. Geospatial data from the A.T.-DSS proved to be valuable for determining the environmental factors restricting the range of Ailanthus. In particular, the seamless climate data products provided by TOPS were a powerful and accessible resource. As a prototype application of the A.T.-DSS, this research demonstrates the utility of coupling in situ and remotely sensed geospatial data with innovative statistical techniques to investigate important ecological processes within the landscape. Modelling the climatic envelope of Ailanthus will provide insights into the future distribution of suitable habitat and potential ecological impacts. This approach establishes a framework that can be effectively adopted to examine the distribution of habitats for additional important species in the region and inform efforts to conserve natural resources and biodiversity. Acknowledgements. The authors acknowledge the contributions from the project team members, in particular Roland Duhaime, Christopher Damon, Fu Luo, Charles LaBash, Peter Paton and Jianjun Zhao from the University of Rhode Island; Forrest Melton, Hirofumi Hashimoto, Samuel Hiatt and Ramakrishna Nemani from the NASA Ames Research Center; Fred Dieffenbach, Matt Robison, Casey Reese and Brian Mitchell from the National Park Service; Ken Stolte from the USDA Forest Service; Glenn Holcomb and Marcia McNiff from the USGS; and Paul Mitchell from the Appalachian Trail Conservancy. We appreciate the comments, suggestions and insights from three anonymous reviewers that helped significantly improve the quality of this manuscript. Funding statement. This study was supported by the Applied Sciences Programme of NASA’s Science Mission Directorate (ROSES-2008) under Decision Support through Earth Science Research Results (grant no. NNX09AV82G).

Appendix I: matrix of the environmental variables incorporated into each candidate habitat suitability model. Model names are derived from the number of bioclimatic, topographic and land-cover variables selected. For descriptions of the variables, see table 1.

15

Phil. Trans. R. Soc. B 369: 20130192

(c) Model limitations

insights into the behaviour and accuracy of predicting the suitable habitats of Ailanthus. In particular, the predictive performance of a model may be enhanced by incorporating mechanics that explicitly address the population dynamics of a migrating species [27].


spatial uncertainty would be more pronounced within the mountainous terrain characterizing the A.T. [85]. Uncertainty is also increased by the long time span of the projection [86]. Finally, the use of an ensemble average may obscure the impact of climate variability, such as extreme winter storms that may prevent Ailanthus establishment. Ultimately, the intended purpose of a model should determine its scale within the limitations imposed by the spatial accuracy of the species observation data [67,82]. While increasing model resolution may reveal fine-scale ecological processes, a priority of this study was to examine the potential influence of climate change. To that end, the broad extent and coarsegrained predictor variables used for this model were wellsuited to investigate the influence of extensive environmental gradients on Ailanthus habitats within the A.T.-shell.

variables included

bio13

x

bio11

x

x x

x x

x x

x

x

model name

allbio_alltopo_alllc

allbio_alltopo 10bio_5topo_4lc

10bio_5topo 10bioalt_6topo

6bioalt_6topo 5bio_5topo

5bioalt_5topo

5bioalt_4topo

x

x

bio12

bio14

bio15

bio16

x

x x

x

x x

x x

x x

x

x

x

x x

x

x x

4bioalt_4topo_2lc 4bioalt_3topo

2bio_1topo

x

x x

4bio_5topo 4bio_4topo

x

x

x

x

x x

bio4

x

x x


x

x

x x

bio3

5bioalt2_3topo

x

6bioalt_6topo 5bio_5topo

x x x

x x

bio2

x x

x

bio1


x

allbio_alltopo_alllc allbio_alltopo

agsum

10bio_5topo_4lc

agdist

model name

x x

x x

x

bio17

x x

x

x

x

bio18

bio5 x x

bio6

x

x

x

x

x

x

bio19

x

x

x

x x

bio7

x

x

cti

x x

x

x

x x

x

x

x

x x

(Continued.)

x

x x x x

x

x

x

x x

devdist

x x x

x x x

x x

bio10

dem

x x

x x

x

bio9

bio8

Phil. Trans. R. Soc. B 369: 20130192

variables included


Appendix I:


16

x x

x


4bioalt_3topo

2bio_1topo

4bioalt_3topo

x

x

x x


x x

x x

x x

x

x

slope

5bioalt2_3topo 4bio_5topo

x

x

x

nlcd_wet

x

x

x

lfcc_std

bio17

5bioalt_4topo

x

lfcc_min

x

bio16

x x

x

x

lfcc

x

bio15


x

x

hli

bio14

x x

x

devsum

bio13


10bio_5topo_4lc 10bio_5topo

allbio_alltopo

allbio_alltopo_alllc

model name

x

bio12

variables included

x x

5bioalt2_3topo 4bio_5topo

2bio_1topo

bio11

model name

x

x

x x

x x

x x

x

x

slopepos

bio18

x

x x

x x

x x

x

x

soil_drain

x

x x

x

bio19

x x

x

x

soil_fldfreq

cti

Phil. Trans. R. Soc. B 369: 20130192

variables included

x x

x x

x

x x

x x

x

x x

x x

x x

trasp

x

devdist

soil_hydric

x

x

x x

x x

dem


Appendix I: (Continued.)


17


References

2.

3.

5.

6.

7.

8.

9.

10.

11. 12.

13.

14. Knapp LB, Canham CD. 2000 Invasion of an oldgrowth forest in New York by Ailanthus altissima: sapling growth and recruitment in canopy gaps. J. Torrey Bot. Soc. 127, 307– 315. (doi:10.2307/ 3088649) 15. Meloche C, Murphy SD. 2006 Managing tree-ofheaven (Ailanthus altissima) in parks and protected areas: a case study of Rondeau Provincial Park (Ontario, Canada). Environ. Manage. 37, 764–772. (doi:10.1007/s00267-003-0151-x) 16. Call LJ, Nilsen ET. 2005 Analysis of interactions between the invasive tree-of-heaven (Ailanthus altissima) and the native black locust (Robinia pseudoacacia). Plant Ecol. 176, 275–285. (doi:10. 1007/s11258-004-0338-0) 17. Kowarik I, Saumel I. 2007 Biological flora of Central Europe: Ailanthus altissima (Mill.) Swingle. Perspect. Plant Ecol. Evol. Syst. 8, 207–237. (doi:10.1016/j. ppees.2007.03.002) 18. Peterson AT, Vieglais DA. 2001 Predicting species invasions using ecological niche modeling: new approaches from bioinformatics attack a pressing problem. Bioscience 51, 363–371. (doi:10.1641/0006-3568(2001)051[0363:PSIUEN] 2.0.CO;2) 19. Thuiller W, Richardson DM, Pysˇek P, Midgley GF, Hughes GO, Rouget M. 2005 Niche-based modelling as a tool for predicting the risk of alien plant invasions at a global scale. Glob. Change Biol. 11, 2234 –2250. (doi:10.1111/j.1365-2486.2005. 001018.x) 20. IPCC. 2007 Climate change 2007: the physical science basis: contribution of Working Group I to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge, UK: Cambridge University Press. 21. Blois JL, Zarnetske PL, Fitzpatrick MC, Finnegan S. 2013 Climate change and the past, present, and future of biotic interactions. Science 341, 499–504. (doi:10.1126/science.1237184) 22. Walther G-R et al. 2009 Alien species in a warmer world: risks and opportunities. Trends Ecol. Evol. 24, 686 –693. (doi:10.1016/j.tree.2009.06.008) 23. Dukes JS. 2011 Responses of invasive species to a changing climate and atmosphere. In Fifty Years of Invasion Ecology (ed. DM Richardson), pp. 345– 357. Oxford, UK: Wiley-Blackwell. 24. Bradley BA, Blumenthal DM, Wilcove DS, Ziska LH. 2010 Predicting plant invasions in an era of global change. Trends Ecol. Evol. 25, 310– 318. (doi:10. 1016/j.tree.2009.12.003) 25. Loarie SR, Duffy PB, Hamilton H, Asner GP, Field CB, Ackerly DD. 2009 The velocity of climate change. Nature 462, 1052 –1055. (doi:10.1038/ nature08649) 26. Hashimoto H, Hiatt SH, Milesi C, Melton FS, Michaelis AR, Votava P, Wang W, Nemani RR. 2011 Monitoring and forecasting ecosystem climate impacts on ecosystem dynamics using the terrestrial observation and prediction system. In Remote

27.

28.

29.

30.

31.

32.

33.

34.

35.

36.

37.

38.

39.

40.

sensing of protected lands (ed. YQ Wang), pp. 525–542. Boca Raton, FL: CRC Press. Thuiller W et al. 2008 Predicting global change impacts on plant species’ distributions: future challenges. Perspect. Plant Ecol. Evol. Syst. 9, 137–152. (doi:10.1016/j.ppees.2007.09.004) Guisan A, Zimmermann NE. 2000 Predictive habitat distribution models in ecology. Ecol. Model. 135, 147–186. (doi:10.1016/S0304-3800(00)00354-9) Hutchinson GE. 1957 Concluding remarks. Cold Spring Harb. Symp. Quant. Biol. 22, 415 –427. (doi:10.1101/SQB.1957.022.01.039) Austin MP. 2002 Spatial prediction of species distribution: an interface between ecological theory and statistical modelling. Ecol. Model. 157, 101–118. (doi:10.1016/S0304-3800(02)00205-3) Brotons L, Thuiller W, Arau´jo MB, Hirzel AH. 2004 Presence –absence versus presence-only modelling methods for predicting bird habitat suitability. Ecography 27, 437– 448. (doi:10.1111/j.0906-7590. 2004.03764.x) Hirzel AH, Le Lay G. 2008 Habitat suitability modelling and niche theory. J. Appl. Ecol. 45, 1372– 1381. (doi:10.1111/j.1365-2664.2008. 01524.x) Lobo JM, Jimeńez-Valverde A, Hortal J. 2010 The uncertain nature of absences and their importance in species distribution modelling. Ecography 33, 103–114. (doi:10.1111/j.1600-0587.2009.06039.x) Sakai AK et al. 2001 The population biology of invasive species. Annu. Rev. Ecol. Syst. 32, 305–332. (doi:10.1146/annurev.ecolsys.32. 081501.114037) Robinson TP, van Klinken RD, Metternicht G. 2010 Comparison of alternative strategies for invasive species distribution modeling. Ecol. Model. 221, 2261– 2269. (doi:16/j.ecolmodel.2010.04.018) Elith J, Leathwick JR. 2009 Species distribution models: ecological explanation and prediction across space and time. Annu. Rev. Ecol. Evol. Syst. 40, 677–697. (doi:10.1146/annurev.ecolsys.110308. 120159) Phillips SJ, Anderson RP, Schapire RE. 2006 Maximum entropy modeling of species geographic distributions. Ecol. Model. 190, 231 –259. (doi:10. 1016/j.ecolmodel.2005.03.026) Hernandez PA, Graham CH, Master LL, Albert DL. 2006 The effect of sample size and species characteristics on performance of different species distribution modeling methods. Ecography 29, 773–785. (doi:10.1111/j.0906-7590.2006.04700.x) Hijmans RJ, Graham CH. 2006 The ability of climate envelope models to predict the effect of climate change on species distributions. Glob. Change Biol. 12, 2272 –2281. (doi:10.1111/j.1365-2486.2006. 01256.x) Elith J et al. 2006 Novel methods improve prediction of species’ distributions from occurrence data. Ecography 29, 129– 151. (doi:10.1111/j.2006.09067590.04596.x)

Phil. Trans. R. Soc. B 369: 20130192

4.

Dufour C, Crisfield E (eds) 2008 The Appalachian Trail MEGA-transect. Harpers Ferry, WV: Appalachian Trail Conservancy. Nemani R et al. 2009 Monitoring and forecasting ecosystem dynamics using the terrestrial observation and prediction system (TOPS). Remote Sens. Environ. 113, 1497– 1509. (doi:10.1016/j.rse. 2008.06.017) Woudenberg SW, Conkling BL, O’Connell BM, LaPoint EB, Turner JA, Waddell KL. 2010 The forest inventory and analysis database: database description and users manual version 4.0 for Phase 2. General technical report RMRS-GTR245. Fort Collins, CO: US Department of Agriculture, Forest Service, Rocky Mountain Research Station. Pimentel D, Zuniga R, Morrison D. 2005 Update on the environmental and economic costs associated with alien-invasive species in the United States. Ecol. Econ. 52, 273 –288. (doi:10.1016/j.ecolecon. 2004.10.002) Cardinale BJ et al. 2012 Biodiversity loss and its impact on humanity. Nature 486, 59 –67. (doi:10. 1038/nature11148) Mack RN, Simberloff D, Mark Lonsdale W, Evans H, Clout M, Bazzaz FA. 2000 Biotic invasions: causes, epidemiology, global consequences, and control. Ecol. Appl. 10, 689 –710. (doi:10.1890/10510761(2000)010[0689:BICEGC]2.0.CO;2) Benning TL, LaPointe D, Atkinson CT, Vitousek PM. 2002 Interactions of climate change with biological invasions and land use in the Hawaiian Islands: modeling the fate of endemic birds using a geographic information system. Proc. Natl Acad. Sci. USA 99, 14 246–14 249. (doi:10.1073/pnas. 162372399) Brauman KA, Daily GC, Duarte TK, Mooney HA. 2007 The nature and value of ecosystem services: an overview highlighting hydrologic services. Annu. Rev. Environ. Resour. 32, 67 – 98. (doi:10.1146/annurev.energy.32.031306. 102758) Charles H, Dukes JS. 2007 Impacts of invasive species on ecosystem services. In Biological Invasions (ed. W Nentwig), pp. 217–237. Berlin, Germany: Springer. Vila` M et al. 2011 Ecological impacts of invasive alien plants: a meta-analysis of their effects on species, communities and ecosystems. Ecol. Lett. 14, 702–708. (doi:10.1111/j.1461-0248.2011. 01628.x) Hu S. 1979 Ailanthus. Arnoldia 39, 29– 50. Hoshovsky MC. 1988 Element stewardship abstract for Ailanthus altissima. Arlington, VA: The Nature Conservancy. See http://www.invasive.org/gist/ esadocs/documnts/ailaalt.pdf. Miller JH. 1990 Ailanthus altissima (Mill.) Swingle ailanthus. In Silvics of North America vol. 2. Hardwoods (eds RM Burns, BH Honkala), pp. 101 –104. Washington, DC: USDA Forest Service.


1.

18


68.

69.

70.

71.

72.

73.

74.

75.

76.

77.

78.

79. 80.

81.

82.

possible new approaches. Ecol. Model. 200, 1–19. (doi:16/j.ecolmodel.2006.07.005) Phillips S. 2005 A brief tutorial on Maxent. AT&T Research. See http://www.cs.princeton.edu/ ~schapire/maxent/tutorial/tutorial.doc. Clark JM. 2013 Effects of climate change on the habitats of the invasive species Ailanthus altissima along the Appalachian Trail. Open Access Master’s Theses, Paper 79. Kingston, RI: University of Rhode Island. See http://digitalcommons.uri.edu/ theses/79. Lobo JM, Jimeńez-Valverde A, Real R. 2008 AUC: a misleading measure of the performance of predictive distribution models. Glob. Ecol. Biogeogr. 17, 145 –151. (doi:10.1111/j.1466-8238.2007. 00358.x) EDDMaps. 2011 Early detection & distribution mapping system. University of Georgia - Center for Invasive Species and Ecosystem Health. See http://www. invasiveplantatlas.org/subject.html?sub¼3003. USDA, NRCS. 2014 Plants Profile for Ailanthus altissima (tree of heaven). In The PLANTS Database. Greensboro, NC: USDA National Plant Data Team. See http://plants.usda.gov/core/ profile?symbol=AIAL. Fryer JL. 2010 Ailanthus altissima. In Fire effects information system. Fort Collins, CO: U.S. US Department of Agriculture, Forest Service, Rocky Mountain Research Station, Fire Sciences Laboratory. See http://www.fs.fed.us/database/feis/plants/tree/ ailalt/all.html. Lemon PC. 1961 Forest ecology of ice storms. Bull. Torrey Bot. Club 88, 21 –29. (doi:10.2307/ 2482410) Kowarik I, Saümel I. 2008 Water dispersal as an additional pathway to invasions by the primarily wind-dispersed tree Ailanthus altissima. Plant Ecol. 198, 241 –252. (doi:10.1007/s11258-0089398-x) Trifilo` P, Raimondo F, Nardini A, Gullo ML, Salleo S. 2004 Drought resistance of Ailanthus altissima: root hydraulics and water relations. Tree Physiol. 24, 107–114. (doi:10.1093/treephys/24.1.107) Landenberger RE, Warner TA, McGraw JB. 2009 Spatial patterns of female Ailanthus altissima across an urban-to-rural land use gradient. Urban Ecosyst. 12, 437 –448. (doi:10.1007/ s11252-009-0087-x) Turner MG, O’Neill RV, Gardner RH, Milne BT. 1989 Effects of changing spatial scale on the analysis of landscape pattern. Landsc. Ecol. 3, 153–162. (doi:10.1007/BF00131534) Wiens JA. 1989 Spatial scaling in ecology. Funct. Ecol. 3, 385– 397. (doi:10.2307/2389612) Levin SA. 1992 The problem of pattern and scale in ecology: the Robert H. MacArthur award lecture. Ecology 73, 1943– 1967. (doi:10.2307/1941447) Wu J. 2004 Effects of changing scale on landscape pattern analysis: scaling relations. Landsc. Ecol. 19, 125 –138. (doi:10.1023/B:LAND.0000021711. 40074.ae) Guisan A, Thuiller W. 2005 Predicting species distribution: offering more than simple habitat

19

Phil. Trans. R. Soc. B 369: 20130192

54. Sobrino JA, Julien Y, Atitar M, Nerry F. 2008 NOAA-AVHRR orbital drift correction from solar zenithal angle data. IEEE Trans. Geosci. Remote Sens. 46, 4014–4019. (doi:10.1109/TGRS.2008. 2000798) 55. Zhao J, Wang Y, Hashimoto H, Melton FS, Hiatt SH, Zhang H, Nemani RR. 2013 The variation of land surface phenology from 1982 to 2006 along the Appalachian trail. IEEE Trans. Geosci. Remote Sens. 51, 2087–2095. (doi:10.1109/TGRS. 2012.2217149) 56. Taylor KE, Stouffer RJ, Meehl GA. 2012 An overview of CMIP5 and the experiment design. Bull. Am. Meteorol. Soc. 93, 485–498. (doi:10.1175/BAMS-D11-00094.1) 57. Vuuren DP et al. 2011 The representative concentration pathways: an overview. Clim. Change 109, 5 –31. (doi:10.1007/s10584-0110148-z) 58. Thrasher B, Xiong J, Wang W, Melton F, Michaelis A, Nemani R. 2013 Downscaled climate projections suitable for resource management. Eos Trans. Am. Geophys. Union 94, 321–323. (doi:10.1002/ 2013EO370002) 59. Hijmans RJ, Phillips S, Leathwick J, Elith J. 2012 Package ‘dismo’. See http://cran.r-project.org/ web/packages/dismo/dismo.pdf. 60. Gesch D, Oimoen M, Greenlee S, Nelson C, Steuck M, Tyler D. 2002 The national elevation dataset. Photogramm. Eng. Remote Sens. 68, 5 –11. 61. Evans JS, Oakleaf J. 2011 Geomorphometry and gradient metrics toolbox. See http://s3.amazonaws. com/nrcs-sgi/EvansModels/Geomorphometryþ %26þGradientþMetricsþþ(versionþa1.01).zip. 62. Fry JA, Xian G, Jin S, Dewitz JA, Homer CG, Yang L, Barnes CA, Herold ND, Wickham JD. 2011 Completion of the 2006 national land cover database for the conterminous United States. Photogramm. Eng. Remote Sens. 77, 858 –864. 63. Schwarz GE, Alexander RB. 1995 Soils data for the conterminous United States derived from the NRCS state soil geographic (STATSGO) data base. Reston, VA: US Geological Survey OpenFile Report 95-449. See http://water.usgs.gov/GIS/ metadata/usgswrd/XML/ussoils.xml. 64. Jimeńez-Valverde A, Peterson AT, Soberoń J, Overton JM, Aragoń P, Lobo JM. 2011 Use of niche models in invasive species risk assessments. Biol. Invasions 13, 2785– 2797. (doi:10.1007/s10530011-9963-4) 65. Warren DL, Seifert SN. 2011 Ecological niche modeling in Maxent: the importance of model complexity and the performance of model selection criteria. Ecol. Appl. 21, 335–342. (doi:10.1890/101171.1) 66. Warren DL, Glor RE, Turelli M. 2010 ENMTools: a toolbox for comparative studies of environmental niche models. Ecography 33, 607–611. (doi:10. 1111/j.1600-0587.2009.06142.x) 67. Austin M. 2007 Species distribution models and ecological theory: a critical assessment and some


41. Elith J, Graham CH. 2009 Do they? How do they? Why do they differ? On finding reasons for differing performances of species distribution models. Ecography 32, 66 – 77. (doi:10.1111/j.1600-0587. 2008.05505.x) 42. Phillips SJ, Dudik M. 2008 Modeling of species distributions with Maxent: new extensions and a comprehensive evaluation. Ecography 31, 161 –175. (doi:10.1111/j.0906-7590. 2008.5203.x) 43. Elith J, Phillips SJ, Hastie T, Dudı´k M, Chee YE, Yates CJ. 2011 A statistical explanation of MaxEnt for ecologists. Divers. Distrib. 17, 43 –57. (doi:10.1111/ j.1472-4642.2010.00725.x) 44. Elith J, Kearney M, Phillips S. 2010 The art of modelling range-shifting species. Methods Ecol. Evol. 1, 330 –342. (doi:10.1111/j.2041-210X.2010. 00036.x) 45. Dieffenbach F. 2003 APPA_HUC10_Shell Metadata. Woodstock, VT: National Park Service Northeast Temperate Network. See http://www.edc.uri.edu/ ATMT-DSS/images/APPA_HUC10_Shell.pdf. 46. Bailey RG. 1980 Description of the ecoregions of the United States. Miscellaneous publication no. 1391. Ogden, UT: US Department of Agriculture, Forest Service. 47. Powell SL, Cohen WB, Healey SP, Kennedy RE, Moisen GG, Pierce KB, Ohmann JL. 2010 Quantification of live aboveground forest biomass dynamics with Landsat time-series and field inventory data: a comparison of empirical modeling approaches. Remote Sens. Environ. 114, 1053–1068. (doi:10.1016/j.rse.2009.12.018) 48. Cleland E, Chuine I, Menzel A, Mooney H, Schwartz M. 2007 Shifting plant phenology in response to global change. Trends Ecol. Evol. 22, 357 –365. (doi:10.1016/j.tree.2007.04.003) 49. De Beurs KM, Henebry GM. 2004 Land surface phenology, climatic variation, and institutional change: analyzing agricultural land cover change in Kazakhstan. Remote Sens. Environ. 89, 497–509. (doi:10.1016/j.rse.2003.11.006) 50. Linderholm HW. 2006 Growing season changes in the last century. Agric. For. Meteorol. 137, 1– 14. (doi:10.1016/j.agrformet.2006.03.006) 51. White MA et al. 2009 Intercomparison, interpretation, and assessment of spring phenology in North America estimated from remote sensing for 1982 –2006. Glob. Change Biol. 15, 2335 –2359. (doi:10.1111/j.1365-2486. 2009.01910.x) 52. Tucker C, Pinzon J, Brown M. 2004 Global inventory modeling and mapping studies. (GIMMS) NA94apr15b.n11-VIg, 2.0. College Park, MD: University of Maryland Global Land Cover Facility. See http://gcmd.nasa.gov/records/ GCMD_GLCF_GIMMS.html. 53. Tucker CJ, Pinzon JE, Brown ME, Slayback DA, Pak EW, Mahoney R, Vermote EF, El Saleous N. 2005 An extended AVHRR 8-km NDVI dataset compatible with MODIS and SPOT vegetation NDVI data. Int. J. Remote Sens. 26, 4485 –4498. (doi:10.1080/ 01431160500168686)


84.

86.

91. Dale VH. 1997 The relationship between land-use change and climate change. Ecol. Appl. 7, 753–769. (doi:10.1890/1051-0761(1997)007[0753: TRBLUC]2.0.CO;2) 92. Dukes JS, Mooney HA. 1999 Does global change increase the success of biological invaders? Trends Ecol. Evol. 14, 135–139. (doi:10.1016/S01695347(98)01554-7) 93. Dale VH et al. 2001 Climate change and forest disturbances. Bioscience 51, 723–734. (doi:10. 1641/0006-3568(2001)051[0723:CCAFD]2.0.CO;2) 94. Araujo M, New M. 2007 Ensemble forecasting of species distributions. Trends Ecol. Evol. 22, 42 –47. (doi:10.1016/j.tree.2006.09.010) 95. Stohlgren TJ, Ma P, Kumar S, Rocca M, Morisette JT, Jarnevich CS, Benson N. 2010 Ensemble habitat mapping of invasive plant species. Risk Anal. 30, 224–235. (doi:10.1111/j.1539-6924. 2009.01343.x)

20

Phil. Trans. R. Soc. B 369: 20130192

85.

87. Kota NL, Landenberger RE, McGraw JB. 2006 Germination and early growth of Ailanthus and tulip poplar in three levels of forest disturbance. Biol. Invasions 9, 197–211. (doi:10.1007/s10530006-9026-4) 88. Pearson RG, Dawson TP, Liu C. 2004 Modelling species distributions in Britain: a hierarchical integration of climate and land-cover data. Ecography 27, 285–298. (doi:10.1111/j.0906-7590. 2004.03740.x) 89. Jones CC, Acker SA, Halpern CB. 2010 Combining local- and large-scale models to predict the distributions of invasive plant species. Ecol. Appl. 20, 311–326. (doi:10.1890/08-2261.1) 90. Arau´jo MB, Luoto M. 2007 The importance of biotic interactions for modelling species distributions under climate change. Glob. Ecol. Biogeogr. 16, 743– 753. (doi:10.1111/j.1466-8238.2007. 00359.x)


83.

models. Ecol. Lett. 8, 993 –1009. (doi:10.1111/j. 1461-0248.2005.00792.x) Austin MP, Van Niel KP. 2011 Improving species distribution models for climate change studies: variable selection and scale. J. Biogeogr. 38, 1–8. (doi:10.1111/j.1365-2699.2010.02416.x) Guisan A, Graham CH, Elith J, Huettmann F, the NCEAS Species Distribution Modelling Group. 2007 Sensitivity of predictive species distribution models to change in grain size. Divers. Distrib. 13, 332–340. (doi:10.1111/j.1472-4642.2007. 00342.x) Pielke RA, Avissar R. 1990 Influence of landscape structure on local and regional climate. Landsc. Ecol. 4, 133–155. (doi:10.1007/BF00132857) Jarnevich CS, Stohlgren TJ. 2008 Near term climate projections for invasive species distributions. Biol. Invasions 11, 1373–1379. (doi:10.1007/s10530008-9345-8)

Medical Applications for 3D Printing: Current and Projected Uses.

Suitable Days for Plant Growth Disappear under Projected Climate Change: Potential Human and Biotic Vulnerability.

The Current and Projected Taxpayer Shares of US Health Costs.

Early events along the trail leading to identification of Rubisco.

Suitable environmental ranges for potential coral reef habitats in the tropical ocean.

Assessing competencies using milestones along the way.

Are mountain habitats becoming more suitable for generalist than cold-adapted lizards thermoregulation?

Appalachian Trail hikers' ability to recognize Lyme disease by visual stimulus photographs.

Is yoga a suitable treatment for rheumatoid arthritis: current opinion.

Re-establishing the pecking order: Niche models reliably predict suitable habitats for the reintroduction of red-billed oxpeckers.

Suitable habitats for endangered frugivorous mammals: small-scale comparison, regeneration forest and chimpanzee density in Kibale National Park, Uganda.

The hydroclimatic and ecophysiological basis of cloud forest distributions under current and projected climates.

TRAIL combinations: The new 'trail' for cancer therapy (Review).

Braden Scale is not suitable for assessing pressure ulcer risk in individuals aged 80 and older.

Current tools for assessing heart function and perfusion adequacy.

Clinical prediction model suitable for assessing hospital quality for patients undergoing carotid endarterectomy.

Methylotrophs in natural habitats: current insights through metagenomics.

Current perspectives on unconventional shale gas extraction in the Appalachian Basin.

Testing projected wild bee distributions in agricultural habitats: predictive power depends on species traits and habitat type.

Comparing the trends of elevated blood pressure in appalachian and non-Appalachian regions.

Meaning, "nerves" and the urban-Appalachian family.

Diversity and ecological structure of vibrios in benthic and pelagic habitats along a latitudinal gradient in the Southwest Atlantic Ocean.

Development of an evaluation framework suitable for assessing humanitarian workforce competencies during crisis simulation exercises.

Current and Projected Heat-Related Morbidity and Mortality in Rhode Island.