69 Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA. HABITAT RELATIONSHIPS AND POTENTIAL RESTORATION SITES FOR QUERCUS PACIFICA AND Q. TOMENTELLA ON CATALINA ISLAND Janet Franklin School of Geographical Sciences & Urban Planning Arizona State University, Tempe AZ 85287-5302 [email protected]and Denise A. Knapp University of California, Santa Barbara Ecology, Evolution, and Marine Biology Department Santa Barbara, CA 93106-9610 [email protected]ABSTRACT: The goals of this project were to develop statistical models determining environmental factors important to two endemic oak species on Catalina Island, Quercus pacifica (Island scrub oak), and Q. tomentella (Island oak), and identify suitable habitat for restoration. Both presence/absence and abundance data were used to develop these models. Three types of models were used: Classification and Regression Trees (CT, RT), Generalized Additive Models (GAM), and Generalized Linear Models (GLM), and predictive maps were produced in ArcView Geographic Information System (GIS). Presence/absence modeling was found to be much more robust than modeling of species parameters such as cover, density, and size. Models of Q. pacifica presence/absence predicted the highest probability of species presence at intermediate values of elevation, slope steepness, topographic moisture, radiation, and distance from the coast (flow length). Greater Q. pacifica abundance (cover) was associated with elevations above 400 m, low A-horizon pH, shallow and silty soils, north-facing slopes, and locations in the island‟s interior (greater distance to the coast). Models of Q. tomentella presence/absence indicate that this species tends to be present on igneous and sedimentary-derived substrates with a medium and gravelly texture, at high values of topographic moisture (TMI), intermediate slopes and flow lengths (distance to coast) and low to intermediate levels of solar radiation. Q. tomentella tree density was highest at moderate elevations (300-350 m), high values of winter radiation, steeper slopes, and lower flow accumulation, while reproduction is highest on concave slopes and at higher elevations; the strong spatial structure of the extant Q. tomentella populations on the island, with two metapopulations significantly larger in size, has likely confounded these relationships, however. Resulting maps of predicted habitat suitability can be used to assist in restoration efforts. KEYWORDS: Classification Trees; Generalized Linear Models; Generalized Additive Models; Quercus pacifica; Quercus tomentella; rare species; restoration; species distribution modeling. INTRODUCTION Poor regeneration, invasive animal and plant species, and large-scale dieback (in the case of Quercus pacifica) threaten Catalina‟s oaks and related ecosystems, which cover nearly one quarter of the island. The Catalina Island Conservancy has sought the collaborative help of a wide range of scientists to gather the data necessary to restore this endemic oak habitat. One of the key objectives identified by a group of these experts in 2003 was to describe the relationship of oak distribution to environmental patterns.
26
Embed
HABITAT RELATIONSHIPS AND POTENTIAL ......69 Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
69
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
HABITAT RELATIONSHIPS AND POTENTIAL RESTORATION SITES FOR QUERCUS
ABSTRACT: The goals of this project were to develop statistical models determining environmental
factors important to two endemic oak species on Catalina Island, Quercus pacifica (Island scrub oak), and Q. tomentella (Island oak), and identify suitable habitat for restoration. Both presence/absence and
abundance data were used to develop these models. Three types of models were used: Classification and
Regression Trees (CT, RT), Generalized Additive Models (GAM), and Generalized Linear Models (GLM), and predictive maps were produced in ArcView Geographic Information System (GIS).
Presence/absence modeling was found to be much more robust than modeling of species parameters such
as cover, density, and size. Models of Q. pacifica presence/absence predicted the highest probability of
species presence at intermediate values of elevation, slope steepness, topographic moisture, radiation, and distance from the coast (flow length). Greater Q. pacifica abundance (cover) was associated with
elevations above 400 m, low A-horizon pH, shallow and silty soils, north-facing slopes, and locations in
the island‟s interior (greater distance to the coast). Models of Q. tomentella presence/absence indicate that this species tends to be present on igneous and sedimentary-derived substrates with a medium and
gravelly texture, at high values of topographic moisture (TMI), intermediate slopes and flow lengths
(distance to coast) and low to intermediate levels of solar radiation. Q. tomentella tree density was highest
at moderate elevations (300-350 m), high values of winter radiation, steeper slopes, and lower flow accumulation, while reproduction is highest on concave slopes and at higher elevations; the strong spatial
structure of the extant Q. tomentella populations on the island, with two metapopulations significantly
larger in size, has likely confounded these relationships, however. Resulting maps of predicted habitat suitability can be used to assist in restoration efforts.
KEYWORDS: Classification Trees; Generalized Linear Models; Generalized Additive Models; Quercus
pacifica; Quercus tomentella; rare species; restoration; species distribution modeling.
INTRODUCTION
Poor regeneration, invasive animal and plant species, and large-scale dieback (in the case of Quercus
pacifica) threaten Catalina‟s oaks and related ecosystems, which cover nearly one quarter of the island. The Catalina Island Conservancy has sought the collaborative help of a wide range of scientists to gather
the data necessary to restore this endemic oak habitat. One of the key objectives identified by a group of
these experts in 2003 was to describe the relationship of oak distribution to environmental patterns.
70
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
The goals of this project were to develop statistical models that would identify environmental factors important to two endemic oak species of interest on Catalina Island, Q. pacifica (Island scrub oak), and Q.
tomentella (Island oak), and identify suitable habitat for restoration. Both presence/absence data and
abundance variables such as cover, density, size and reproduction were used as the response variables.
Three types of models were used for continuous response variables: Regression Trees (RT), Generalized Additive Models (GAM), and Generalized Linear Models (GLM). Classification trees (CT) were used for
categorical variables. GAMs are useful for exploring the shape of the response functions, GLMs allow
those response functions to be parameterized and their significance tested, and classification and regression trees are particularly useful for exploring the effects of categorical predictors (Franklin 1998).
In this study we compared the results of multiple types of models in order to gain insight into complex
ecological relationships between oak distributions and environment.
The distribution of oak woodland, chaparral, and other California habitat types found on Catalina Island
such as coastal sage scrub, grassland and Catalina ironwood (Lyonothamnus floribundus ssp. floribundus)
groves have been found to be directly influenced by substratum characteristics, moisture availability, temperature and precipitation, and disturbance history (such as fire and grazing) (Wells 1962, 1968;
Harrison et al. 1971; Kirkpatrick & Hutchinson 1980; Ng & Miller 1980; Poole & Miller 1981; Westman
1981, 1991; Miller et al. 1983; Davis et al. 1986; Davis & Goetz 1990; Franklin 1998; Davis et al. 2007; Keeley and Davis 2007; reviewed in Franklin et al. 2000). Moisture availability is, in turn, influenced by
slope angle, slope aspect, soil depth, solar radiation, and climate (precipitation, fog, and temperatures
during the winter and spring growing seasons). Variables that have been particularly successful at explaining patterns of chaparral distribution, diversity, and species composition include elevation and
potential solar insolation (Franklin et al. 2000, Keeley et al. 2005, Keeley and Davis 2007). The oak
woodland community in California and its various dominant oak species appear to have strong
associations with soil depth and type (Griffin 1988). Wells (1962) noted the importance of depth and texture of soils to the distribution of different physiognomic types of vegetation, reporting a dominance of
woody types on sandy or shallow, rocky soils in central coastal California.
Quercus dumosa, the former classification of the Channel Islands endemic scrub oak Q. pacifica (Nixon
& Muller 1994), has been found to be “sensitive to slight differences in environment,” preferring
northerly exposures and almost never being found on westerly or southerly slope aspects (Bauer 1936). A
study of oak distributions on Santa Cruz Island in relation to slope, aspect, and substrate (Jones et al. 1993) revealed that oaks such as Q. pacifica (then classified as Q. dumosa) were not found to correlate
strongly with slope and aspect, although exhibiting some preference for north-facing slopes. Q. dumosa
has also been shown to be deeply rooted and limited to deep soil conditions (Kummerow & Mangan 1981). Poole & Miller (1981) found moderately high (less negative) xylem pressure potentials for Q.
dumosa when compared with other chaparral species Rhus spp., Adenostoma fasciculatum,
Arctostaphylos spp., and Ceanothus greggii, indicating that these other species are less affected by soil drought. Q. dumosa was also found to occupy areas with intermediate to low precipitation and
temperatures throughout its range in California (Westman 1991).
Q. tomentella occupies a wide range of geologic substrates and soils on Santa Cruz and Santa Rosa Islands, while preferring steep northwest- and north-facing slopes at higher elevations, canyon bottoms,
and foggy locations (Kindsvater 2006). Oak species in general on Santa Cruz Island were found only on
the “less faulted highly weathered substrates, especially the Santa Cruz Island schist and Willows diorite,” and were “absent from all disturbed substrates” (Jones et al. 1993). Maritime influences (such as fog),
elevation, and exposure appeared to produce patterns found in vegetation cover there (Jones et al. 1993).
On Santa Rosa Island, significant differences were found in soil factors for Q. tomentella stands with and
without seedlings, with regenerating sites having soils with higher pH, higher exchangeable potassium,
71
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
lower percent clay, higher percent sand, and lower phosphorous (Kindsvater 2006). The current, limited
distribution and recruitment of Q. tomentella on Santa Rosa and Santa Cruz Islands is expected to be the result of heavy grazing by introduced herbivores (Kindsvater 2006).
Environmental predictor variables used in the statistical modeling were selected based on these known
correlations and either measured in the field or derived from other sources, as described in the following section.
METHODS
Species data used for these analyses included presence/absence information for each oak species, as well
as cover, density, size, recruitment and mortality information derived from plot data for Q. pacifica and survey data for Q. tomentella. Data collection methods are described below.
Species data collection
Species presence/absence A Q. pacifica distribution map was produced in 2004 using aerial photographs taken in 2000, and is a
refinement of a vegetation map identifying more general Island Chaparral and Island Woodland
communities (Knapp 2005). Mean polygon size is 3.2 ha (1353 polygons, range 0.08-318.5 ha). A Q. tomentella occurrence map was produced based on populations mapped in the field using a Trimble
GeoExplorer III Global Positioning Systems (GPS) unit (differentially corrected for sub-meter accuracy;
Trimble Navigation Limited, Sunnyvale, CA) as part of a full census and survey of this species on the island completed in 2005 (McCune 2005).
Quercus pacifica plots
The Q. pacifica survey was designed for the purposes of vegetation association classification, species
modeling, and mapping. Q. pacifica plots were surveyed using the relevé method (Mueller-Dombois &
Ellenberg 1974), which involves detailed plant and substrate information collected within a large plot. Plots were 40 x 10 m (400 m
2), a size found to be appropriate on other islands (K. McEachern, pers.
comm.), and which is near the larger end of the range (200-500m2) recommended for tree plots by
Mueller-Dombois and Ellenberg (1974). Care was taken to ensure that, with regard to environmental
variables, sampled plots were homogeneous and were laid out, where possible, with the long axis parallel to the slope. Data were collected between January and July of 2005. A sample of 200 point locations was
randomly generated within the mapped distribution of Q. pacifica recorded in a GIS (Geographic
Information Systems). Points that were accessible (roads to the hiking access point were drivable) were located using a GPS and surveyed first in the spring months; when the roads dried out, the number of
plots that could be feasibly completed were randomly chosen from among the remaining points. A total of
131 relevé plots were sampled.
Variables measured in the plots (Table 1) included species composition, relative and absolute plant cover
(determined visually using teams of two calibrated observers), substrate cover (e.g. soil, rock, litter),
vertical structure (cover in ground, low shrub, high shrub, and tree layers), number of dead shrub and tree individuals, and observed disturbance type (animal trailing, animal scat, human sign) and level (high,
medium, or low). Basal diameter was recorded for all trunks of up to eight selected tree individuals (up to
two within each quadrant) at each plot. In addition, soil samples were taken at both 0-10 cm (“A” horizon) and 10-20 cm (“B” horizon) depths, and were analyzed by the Soil Ecology and Restoration
Group (SERG, San Diego State University, California). Soil samples were taken under the largest
72
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Table 1. Field data available for modeling Quercus pacifica. Plots were 40 x 10 m (400 m2)
Type of Variable Name Description
Terrain Variables
TOPOPOSI Topographic position, determined in the field:
1=Bench (flat or nearly flat area on a slope); 2= Ridge
top; 3= Upper 1/3 of slope; 4= Middle 1/3 of slope; 5= Lower 1/3 of slope; 6=Slope bottom; 7= Stream or draw
bottom; 8= Stream terrace; 9= Other
SOILDEPT Soil Depth, estimated in the field when taking soil
samples: 1= 0 cm; 2=1-10 cm; 3=10+ cm
Disturbance Variables
TRAILING Trailing observed (low, medium, high, or none - 0-3)
RTSXPSD Roots exposed (low, medium, high, or none - 0-3)
ANTHDIS Anthropogenic disturbance evident, including fence, road, trail, cultivation, building site, corral, or other
(low, medium, high, or none - 0-3)
Cover Variables
%SOIL % cover of soil in releve plot
%ROCK % cover of rock in releve plot
%LITTR % cover of litter in releve plot
%CRUST % cover of crust in releve plot
%MOSS % cover of moss in releve plot
%LICHEN % cover of lichen in releve plot
%PLNTCOV % cover of plants in releve plot
%GRNDCOV % cover of ground-level (0-0.5 meter) plants in releve
plot
%LOWSHRU % cover of low shrubs (0.5-2 meters) in releve plot
%HISHRU % cover of high shrubs (2-4 meters) in releve plot
%TREE % cover of trees (4-6 meters) in releve plot
Soil Variables
A horizon (0-10 cm
depths); Analyzed by Soil
Ecology & Restoration
Group, SDSU.
B horizon data not used in
modeling.
PO43 (PO43-) from soil sample, ug/g sample
NH4 (NH4+) from soil sample, ug/g sample
NO3 (NO3-) from soil sample, ug/g sample
pH pH, from soil sample
ORGCON Organic content from soil sample, (g/g soil)
ORG % organic content from soil sample, by weight
CLAY % clay, from soil sample
SAND % sand, from soil sample
SILT % silt, from soil sample
CLASS classification, from soil sample
Quercus pacifica
Variables
QUPAcov cover of Quercus pacifica
50DEAD value of 1 if greater than half of the Quercus pacifica
trees are dead
DEAD number of dead Quercus pacifica in plot
SAPLS number of Quercus pacifica saplings in plot
SDLS number of Quercus pacifica seedlings in plot
REPRO sum of SAPLS & SDLS, above
AVELDIAM Average diameter of up to 8 live trees in the releve plot
AVEDDIAM Average diameter of up to 8 trees in the plot including
dead trees
73
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
measured living oak tree; if there were no live oaks in the plot, soil was collected from beneath the largest
measured dead tree. Soils data obtained include PO43-
, NH4+, NO3
-, pH, organic content, and texture.
Quercus tomentella survey
Field survey information for Q. tomentella was collected during a 2004-2005 census and survey of all groves on the island. Variables describing the abundance and vigor of Q. tomentella and used in modeling
included number of mature trees, area of each grove, density of mature individuals, maximum DBH,
average DBH, number of dead stems, health (by category) and reproduction (Table 2). Representatives of average-sized stems were assessed visually; if the stems varied too much in size to easily pick an
“average sized” tree, then the largest and the smallest stems were each measured and recorded, and the
average calculated. Size class was assessed as follows: mature = DBH 3 cm; sapling = basal diameter >1cm and <3cm; seedling = basal diameter <1cm (the “seedling” size class, as defined, includes more that
just first year recruits). In order to take into account the size of a grove when comparing the number of seedlings and saplings between groves, a Reproductive Index was calculated for each mature grove by
dividing the total number of saplings and/or seedlings within a grove by the total number of stems, then
multiplying by 100 to give the percentage of stems which have „replacement‟ seedlings or saplings.
Table 2. Field data available for modeling Quercus tomentella, collected in stands or groves
Variable Description
MATUR # of mature individuals (distinct groups of stems, although some are likely the same genet).
Area Area of the grove in square meters
DNSITY Density: # of mature individuals divided by area of grove
MaxDBH Diameter at breast height for the largest stem observed
AveDBH Average diameter at breast height, either by averaging maxDBH and minDBH or by
measuring an average-sized individual in the field
STSD % stems dead, not resprouting
Hlth Health by subjective categories, based on observed impacts of the noted threats: poor=1,
fair=2, good=3, very good=4, excellent=5. Poor = “the demise of the grove appears
imminent,” while Excellent = “no or almost no imperfections or impacts could be found.”
Repro Reproductive index, calculated by taking the number of seedlings + saplings, dividing by the
number of mature stems, and multiplying by 100
All Q. tomentella groves on the island were surveyed and sampled, and all mature stems (ramets) were
counted and assessed at each grove (because these trees are clonal, the number of true individuals is
difficult to determine). A total of 95 groves were surveyed (for the purposes of this survey, a grove is defined as a group of trees whose canopies were not separated by more than 15 meters), clustered in
seven occurrences on the island. Occurrences are defined by the California Native Plant Society as
populations occurring more than one quarter mile from each other (Bittman 2001).
Mapped explanatory variables
Environmental variables representing factors important to oak distributions and abundances (discussed in the introduction) were used as explanatory variables. Although weather stations were recently installed on
the island, sufficient data had not yet been collected to use in this analysis. Fire distribution data were also
not available, although no oak plots were placed in areas known to have burned within the last 20 years, and fires larger than two ha have been uncommon on Catalina Island over the last century (DK, unpubl.
fire history data). GIS-derived variables used as predictors were the same for both species (Table 3), and
74
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Table 3. GIS data and derived mapped variables available for modeling Quercus pacifica and Q
tomentella
Variable Description
ELEV Elevation in meters, derived from a 5-m DEM (maximum island elevation: 670 m)
SLOPE Slope in degrees, generated from the 5-m DEM using ArcView
ASP Aspect in degrees, generated from the 5-m DEM using ArcView, converted to “northness,”
an aspect index (=cos(ASP+45)), for modeling. CURV Curvature, generated from 5-m DEM using ArcView (negative for convex, positive for
concave slopes) calculated as the average of planform and profile curvature
were primarily derived from a five-meter gridded Digital Elevation Model (DEM) developed photogrammetrically for the Conservancy in 2004 by Vexcel Corporation (Boulder, Colorado). This
spatial resolution of the DEM correlates well with the scale of mature oak canopies on Catalina, which
typically are five to ten meters in diameter. At the scale of the DEM, the Topographic Moisture Index (TMI) showed consistent fine-scale patterns of variability throughout the island, and thus represents
microtopographic moisture changes rather than broad landscape-scale changes. TMI is calculated by
scaling the flow accumulation (number of uphill grid cells) by the slope of each cell, if subsurface flow rate is assumed to be constant (Franklin et al. 2000).
Island-wide soil data were obtained from Soil Conservation Service maps produced in 1955, digitized by
ESRI (Redlands, CA) for the Conservancy. Within this dataset, however, permeability and soil depth had very low variability among the oak plots. General geologic variables were derived by GIS consultant
Mike Klinefelter with geologists at University of California Riverside, starting with the Geology/bedrock
GIS layer produced by the Center for Natural Areas (based on work by E. Bailey in 1940's), and combining these types into five coarser categories of geologic substrate (with 2.6% of the island mapped
75
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
as Unknown); mean polygon size of the resulting map was 294 ha. Flow length (distance from point to
ocean along stream channel) was explored as a surrogate for coastal fog, in the absence of fog distribution data, as fog is often observed to travel up drainage basins (Figure 1). Presumably, areas farther from the
coast would have less of the moderating influence of fog, and would thus be drier.
Mean values of gridded predictor variables for each relevé were obtained using the Grid Statistics
function within the Surface Areas and Ratios from Elevation Grid (v. 1.2) extension (Jenness Enterprises,
www.jennessent.com) in ArcView 3.2 (ESRI, Redlands, CA). Relevé values for categorical (polygon) variables such as geology and soils were determined using the XTools extension in ArcView 3.2
(www.arcscripts.esri.com).
Although grazing is hypothesized to have a negative effect on recruitment, no spatially-explicit measure
of grazing intensity or history was available for this modeling study (although see Stratton and Manuwal
& Sweitzer, this volume).
Data analysis
An overview of the data analysis and statistical modeling steps is given in Table 4. Q. pacifica plot location data were used to analyze factors correlated with overall abundance, size, recruitment, mortality,
and occurrence. Q. tomentella survey data were used to analyze factors correlated with density,
regeneration and occurrence. Exploratory analyses were conducted for both species using simple (single predictor) GAMs, CTs, scatterplots, and Spearman‟s rank correlations. On the basis of these preliminary
analyses, variables were selected for use in two multiple-predictor tree models, GAMs, and GLMs.
Variables were screened for multicolinearity using pairwise scatterplots and linear correlation coefficients. In order to derive uncorrelated (orthogonal) composite variables that capture the main
variance in the soil data, a principal components analysis was conducted of soil “A”-horizon variables
using the covariance matrix, in the package labdsv in the R software. The resulting composite variables are weighted sums of the original variables. Of the composite soil variables derived from principal
components analysis, PC1 is positively correlated with A horizon NO3-, PC2 is positively correlated with
silt and negatively with sand, and PC3 is negatively correlated with A horizon NH4+; these three principle
components explained 92% of the variance in the 9-variable soils dataset. Soil did not extend beyond 10
cm depth for many of the plots, therefore only the A horizon data were used for analysis. The same
Figure 1. Fog often travels up drainage basins to a level around
400-500 m. Photo shows Blackjack
Peak in the distance. (Photo by D.
Knapp.)
76
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Preprocess soil variables Principal Components Analysis
Examine predictors for multicollinearity Bivariate scatter plots
Exploratory analysis of potential predictors Correlation, simple regression (GLM) or GAM, or
simple classification trees (CT) for categorical
predictors, to select variables to be included in multiple
predictor models for each
Generate absence data for presence/absence models
Random sample of points falling outside respective species maps
Multiple predictor models of each response
variable, each species
GAMs and GLMs, with AIC used for variable
selection; Classification trees (CT) for categorical
responses and regression trees (RT) for continuous
responses.
Model evaluation R2 and p-values used to assess models of continuous
response variables;
AUC used to assess predictive performance of binary
presence/absence models
Spatial prediction using presence/absence
models
GLMs implemented in GIS as weighted sums of
predictor maps; two alternative models compared
variables tended to be correlated for the two soil depths, therefore A horizon data alone is felt to be
representative of soil characteristics.
In order to develop robust models of those factors associated with likelihood of oak species‟ occurrence,
we generated samples of absences (797 for Q. pacifica, 1000 for Q. tomentella) by randomly sampling the island but excluding those points falling within the species maps. The total number of points used for
modeling was 928 for Q. pacifica (131 present, 797 absent; sample prevalence is 14% presences) and
1095 for Q. tomentella (95 present, 1000 absent, sample prevalence is 9% presences). Models using
multiple predictors were developed using GAMs, GLMs, and RTs (for continuous response variables describing abundance) or CTs (for species presence/absence, a binary, categorical response). The Akaike
Information Criterion (AIC) was used for model selection. Manual variable selection was compared to
stepwise (automated) backward elimination.
For models of continuous responses, the explained variance and significance of the models were used for
evaluation. For the models of species presence/absence the Area Under the Curve (AUC) of the ROC
(receiver operating characteristic) plot was used as a measure of the model‟s ability to correctly discriminate between presences and absences. In this study we did not have independent data on species
presence to evaluate the models, so AUC for each presence/absence model was calculated from the same
data used to develop the models; this likely gave an overly-optimistic estimate of prediction accuracy for new data. Therefore, for the best model for each species, we also calculated a bootstrapped AUC to
estimate what that lower (penalized) AUC would probably be if new data were available.
Using the GLMs, syntax was then developed for use in the Arc View Map Calculator (ESRI, Redlands,
CA) to predict the log-likelihood ratio of species presence/absence from a weighted sum of all of GIS
layers representing the predictor variables used in each model, and to transform this to a predicted
77
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
probability of presence, on a scale of 0-1. Predictive raster layers were converted to binary habitat
suitability maps, and then to point layers, and overlain with a vegetation map for the island (Knapp 2005) to determine the distribution of existing vegetation in areas predicted to be suitable habitat.
RESULTS
Quercus pacifica cover
Plots varied greatly in percent cover of island scrub oaks, a measure of species abundance in the plot, from one to 80 percent (mean 34%, median and mode 33%; Figure 2). Exploratory analyses of predictors
of abundance using Q. pacifica cover as the dependent variable (scatterplots and boxplots) suggested that
cover was negatively related to A-horizon pH (r=-0.25) and soil depth (r=-0.17). Cover was positively related to flow length (distance from the coast; r=0.20) and soil PC2 (high silt, low sand; r=0.22). Q.
pacifica trees occurred somewhat more often on igneous intrusive substrates than expected given their
extent on the island (27% occurrence, 16% available). No relationship was found with the remaining
variables. Based on this exploratory analysis, candidate variables were selected and models were fit using multiple predictors, described below.
The RT model explained 38% of the variance in cover, and included the following predictor variables:
ELEV, A-horizon pH, PC2, FLWLEN, the “northness” aspect index (=cos(ASP+45)), and GEOL (refer
to Table 1 and 3 for definitions of these variables). Higher cover values are found at elevations >418 m, or lower elevations if A-horizon pH<5.49 (especially at higher flow length [farther into the island‟s
interior]), OR if A-horizon pH is higher, then at higher values of PC2 (silty, not sandy).
The GAM of cover suggested non-linear relationships between cover and PC2, elevation, flow length and
northness (Figure 3), although none of these non-linear responses had significantly greater fit than linear
responses (details not shown). Higher cover values were associated with low values of soil depth and A-horizon pH, moderately high values of PC2 (silty textured soils) and northness, high values of flow
length, and extreme (high and low) values of elevation.
In the GLM of cover, only ELEV (2nd
order polynomial; p=0.003), A-horizon pH (p=0.095) and soil PC2 (p=0.041) were significant or nearly so; the model was significant (p=0.005) but explained a modest
amount of variance (R2 =
0.237). Higher cover values were associated with high values of soil PC2 (silty
textured soils), but low values of A-horizon pH, low values of soil depth, high northness, extreme values of elevation and high flow length, as in the GAM (but with less variance explained, as noted above).
Figure 2. Histogram of Q.
pacifica cover values in the
131 plots.
78
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Figure 3. Smoothing function fit for Q. pacifica GAMs. X-axes are the explanatory variables, y-axis represents the predicted value of cover based on
the smoothing function.
Quercus pacifica size
Average size may indicate conditions appropriate for growth and survival. Exploratory GAMs, scatterplots and Spearman‟s rank correlations suggested that size (average diameter) is slightly positively
related to elevation (r=0.17), flow accumulation (r=0.12) and slope (r=0.09) (the nonlinear terms not
significant), and slightly negatively related to northness, flow length, and cover of rocky substrate (r=-0.13). Size was also slightly positively related to soil depth and animal scat, and unimodally related to A-
horizon NH4+ and organic matter. Based on this exploratory analysis, RT and GLM models were fit using
multiple predictors, described below.
Key variables selected for use in a RT model included A-horizon NH4+ (high diameter at high values),
slope (high diameter at higher values), PC1 (A horizon NO3-), and northness (Figure 4). Although the 12-
node model used explained 41% of the variance in size, cross-validation results suggest that the full model is overfit to the training data; only a 2- to 6-node model would be necessary to make robust
predictions (11-29% of variance explained).
79
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Figure 4. Regression tree for Q. pacifica size. Graphical representation of tree decision rules.
Values at terminal nodes are average values of AVELDIAM (stem diameter in cm) for observations at that node (see Tables 1 and 2 for variable descriptions).
Significant predictors of tree size in the GLM included quadratic terms for A-horizon NH4
+ (p=0.046) and
slope (largest size at intermediate values), and positive linear terms for flow accumulation (p=0.015) and
disturbance (p=0.062) (largest average tree size at higher values of these variables, although disturbance
is likely the effect, not the cause). This model, although significant, only explained 12% of the variance in size (based on adjusted R
2).
Quercus pacifica recruitment
Exploratory analyses indicated that there was a slight tendency for juvenile abundance (count of seedlings plus saplings) to be higher in intermediate canopy cover (20-40%), at low values of PC1 (low A-horizon
NO3-) and flow length (nearer to the coast), and in plots with fewer dead trees (Figure 5). However, in a
multiple linear regression model (GLM) only flow length and PC1 were significant, and the model, while significant at the 0.05 level, only explains 7% of the variance in recruitment among plots. These same two
variables were used in a RT model but with similarly low explanatory power.
Figure 5. Boxplot showing number of dead trees (x-axis)
versus number of juveniles in Q.
pacifica plot (y-axis).
80
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Quercus pacifica dieback
Correlates of extensive tree mortality (“dieback”) were difficult to identify using these data because dead trees only occurred in 32 of the 131 plots, 19 of those having only one dead tree. Fewer than ten plots
contained 20 or more dead trees. Raw number of dead trees was slightly correlated with PC1 (high A-
horizon NO3-;r=0.177) and there was a slight tendency for plots with dead trees to be found at higher
elevation (r=-.146) and lower flow lengths (r=0.134) or high on the hillslope (upper or middle third of slope).
Quercus pacifica presence/absence
A classification tree model, pruned to the best ten nodes, indicated that Q. pacifica is more likely to be
present if:
Winter radiation is low (steeper, north-facing slopes), flow accumulation >2.54 (not right on a
ridge where FLOWACC=0), slope is not extremely steep (<26) AND elevation >189 m (58% of 108 observations);
OR if {elevation <189 m AND north facing slope} (node 41, 34% of 29 observations);
OR if {winter radiation is low, flow accumulation >2.54 (not right on a ridge), slope is extremely
steep (>26) AND texture class is gravelly-medium} (70% of 10 observations);
OR if {winter radiation is higher (>1169), but flow accumulation is higher than 1.26 (not right on
ridge), Geology is either igneous extrusive, a mixture of metamorphic and unknown, a mixture of
young sedimentary and igneous extrusive, or unknown, AND flow accumulation is >25 (mid
slope or lower slope} (57% of 14 observations).
To summarize the CT model results, Quercus pacifica tends to be present on low-radiation, steep, north-
facing slopes, or in higher radiation environments if lower on the hillslope or on certain substrates. Bivariate GAMs suggested a unimodal response curve of this species to radiation, elevation, slope, flow
length and topographic moisture index (TMI) (Figure 6); in other words, the highest likelihood of
presence is at intermediate values and the lowest likelihood is at extreme values of these variables.
The data for Q. pacifica show spatial structure, with species presences appearing clustered relative to the
randomly-located pseudoabsences. Join-count statistics showed that, at a spatial lag distance of 400 m, the
errors resulting from the GLM predictions (probability threshold of 0.15) were positively associated. False positives (FP) are near other FPs more often than expected by chance, and the same is true for false
negative predictions (FN) – the z-values are greater than 1.96. This suggests that future work should use
one of the existing methods to incorporate spatial dependence into species distribution models in order to improve predictions.
Models using multiple predictors were developed with aggregated geology and texture classes, and TMI.
A pruned 9-node CT model used five variables and had an AUC of 0.85. All GAMs and GLMs had AUC values >0.9, meaning that, for a randomly selected pair of presence-absence observations, the model
prediction for presence will be greater than the prediction for absence over 90% of the time. Stepwise
selection tended to produce more accurate models than manual, and GAMs better than GLMs, although these differences were negligible (Table 5). GLMs were chosen to produce predictive maps because they
are easy and intuitive to use for prediction.
81
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Figure 6. Shape of response function between the log-likelihood ratio of Q. pacifica
presence (y-axis) and environmental variables (x-axis) shown by bivariate GAMs.
Table 5. Summary of generalized linear models of Q pacifica and Q. tomentella probability of
presence used for spatial prediction of habitat suitability shown in Figures 7, 8, 13 and 14. * - bootstrap AUC. Refer to Tables 1 and 3 for variable descriptions.
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Elevation, slope, curvature, TMI, solar radiation, and geology (igneous intrusive, sedimentary, and
unknown) were used in the ArcView Map Calculator to predict the log-likelihood ratio of species presence/absence using the GLMs based on both manual (Qupapred1; Figure 7) and stepwise
(Qupapred2; Figure 8) variable selection (Table 5). This likelihood ratio was transformed to a predicted
probability of occurrence. Note that the colors in Figures 7 and 8 are scaled so that positive predictions
(high likelihood of presence, shown in white to red) are those probabilities greater than 0.15 (approximately equal to prevalence of the species in the sample, which was 14% in this case). The
prevalence of species presence in the sample can be an appropriate threshold to use for a binary prediction
of suitable versus unsuitable habitat according to the literature, although other threshold criteria can be used, such as average predicted probability.
A cross-tabulation of the thresholded binary maps of predicted suitable habitat based on these two models with the map of existing vegetation is presented in Table 6. The majority of the area of predicted suitable
habitat is found in locations currently mapped as Island chaparral; less area is located in grassland, and
still less in coastal sage scrub. These areas of predicted suitable habitat are not located randomly with
respect to existing vegetation (Chi-square, p<0.001), and disproportionately occur in Island chaparral and grassland. There are differences between the manual and stepwise GLMs, including the variables selected
and the form of the response (linear versus unimodal). The manual selection method predicted a larger
area of potential habitat (Qupapred1; 17%) than the stepwise method (Qupapred2; 4%, which is suspiciously low), however suitable habitat is predicted to occur in roughly the same area.
Figure 7. Spatial predictions of Q. pacifica based on the best GLM using manual variable selection (Qupapred1). Sampled oak points are also shown in blue (Qupapoint.shp).
83
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Figure 8. Spatial predictions of Q. pacifica based on the best GLM using stepwise (backward
elimination) variable selection (Qupapred2). Sampled oak points are shown in blue (Qupapoint.shp).
Table 6. Existing vegetation (from vegetation map) found in areas of predicted suitable habitat for Q.
pacifica. Only values of 0.5% or higher are presented.
Vegetation Type (percent
area of island)
Percent of predicted area Area predicted (ha)
Qupapred11 Qupapred22 Qupapred1 Qupapred2
Island Chaparral (29.4%) 42.0% 48.8% 1,369 388
Grassland (19.5%) 29.8% 28.2% 972 224
Coastal Sage Scrub (38.1%) 12.0% 17.1% 746 136
Bare (9.4%) 3.7% 3.2% 121 25
Island Woodland (0.5%) 0.6% 1.5% 22 12
1 GLM manual selection, probability of presence 0.15 or higher 2 GLM stepwise selection, probability of presence 0.15 or higher
Quercus tomentella density
Tree density shows strong spatial structure, with high values in the most northwest cluster of groves (in the Mount Orizaba area) and low values in the southwest cluster (Figure 9). Preliminary screening of Q.
tomentella density data using scatterplots indicated that solar radiation, flow accumulation, elevation,
geology, soil texture, slope, and curvature appeared to be related to abundance (density). The relationship
with elevation appeared to be unimodal with a peak at about 350 m, while the relationship with slope, solar radiation, and curvature appeared to be nonlinear but monotonic-increasing (positive) (Figure 10).
84
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Figure 9. Q. tomentella density proportional to size of circles in the 95 groves, showing high values in the northwest cluster of groves and low values in the southwest cluster.
Figure 10. Shape of response function between Q.
tomentella density (y-axis) and environmental variables (x-
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
A multiple regression tree (RT) model with four nodes explained 24% of the variability in density.
Density was predicted to be higher at lower values of flow accumulation (higher on the slope or in the watershed), where values of radiation were relatively higher but slope curvature was low, and also where
radiation was low (north-facing slopes). Cross-validation of this model suggested that it was not
particularly robust.
A GAM using six variables explained 31% of the variance in density and predicted higher density at
higher winter solar radiation, lower values of flow accumulation, higher values of curvature, and
intermediate values of elevation (Figure 11). Only the non-linear term for radiation was significant.
Quercus tomentella reproductive index
Exploratory variable selection using bivariate GAMs indicated a significant and positive relationship
between Q. tomentella reproductive index and elevation (D2 = 0.06) and curvature (unimodal with
optimum at 20 or moderately concave; D2 = 0.19). A four-node regression tree (RT) model predicted
higher reproduction on more concave slopes or, if on less concave slopes, at elevations around 400 m. While the full model explained 33% of the variance in reproduction, cross-validation suggests that
curvature yields the only significant split. Similarly, a multiple GAM (adjusted D2 = 0.25) revealed
higher reproduction at curvature values around 20 and elevation around 400 m, as well as lower winter radiation, with a slight preference for certain substrates (igneous) and soil texture classes (gravelly,
medium) (Figure 12). The relationship with flow accumulation was complex. A GLM (Gaussian link
function) using the same variables explained 17% of the variance in reproduction, however the only significant coefficient was for curvature.
Quercus tomentella presence/absence
Both single- and multi-variable classification trees were used to identify groups of categorical data classes
associated with presence or absence of Q. tomentella. Geologic substrate (igneous extrusive and intrusive,
sedimentary) and soil texture (gravelly, medium) had a positive association with species presence, and were aggregated for use in further analysis. Bivariate GAMs were used for variable selection and to
examine response curve shapes for continuous variables. Flow length, radiation, and slope showed
unimodal responses with highest likelihood of presence at intermediate values (Figure 13). High
prevalence was also found on the most extreme (concave and convex) slope curvatures, which does not make ecological sense.
A multivariate classification tree, pruned to the best 12 nodes (not shown), used geology, slope, flow length, TMI, and solar radiation as predictors, and had an AUC of 0.998. This model indicated that the
probability of finding Q. tomentella is high if:
Geology is igneous extrusive, sedimentary, or sedimentary young AND texture class is
gravelly/medium AND flow length (distance to coast) is intermediate (7519 > FLWLEN > 4843)
AND radiation in winter is low (< 5548) (94% of 72 points where Q. tomentella is present);
OR if geology is igneous intrusive, metamorphic, or unknown (98.4% of 625 points where Q.
tomentella is absent);
BUT exceptions are found where TMI is high (>4.28), texture is gravelly/medium or medium,
slope is greater than 10 degrees and flow length is high (> 1744) (100% of 7 points where Q. tomentella is present).
86
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Figure 11. Shape of response function between Q. tomentella density (y-
axis) and environmental variables (x-axis) shown by a multivariate GAM.
Figure 12. Shape of response function between reproduction (y-axis)
and environmental variables (x-axis) shown by a multivariate GAM.
87
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Figure 13. Shape of response function for Q. tomentella between log-likelihood of species
presence (y-axis) and environmental variables (x-axis) shown by bivariate GAMs.
Stepwise selection tended to produce more accurate models than manual (Table 5), and GAMs better than GLMs, although these differences were negligible. GLMs were chosen to produce predictive maps
because they are easy and intuitive to use for prediction. The best GLM based on manual variable
selection (Qutopred2) included geology (igneous intrusive, metamorphic), TMI, summer solar radiation
(p=0.033), and flow length as significant predictors (p < 0.001 unless otherwise noted), the last two as second-order polynomials approximating a gaussian response function. Stepwise variable selection
yielded the final GAM and GLM (Qutopred3) models, which included elevation (p=0.017), slope
(p=0.002), TMI (p=3.0e-07), flow length (p=1.75e-05), winter solar radiation (p=0.025), and aggregated geology (p=9.77e-13) as significant predictors (TMI and geology as second-order polynomials).
Binary geology (igneous intrusive, metamorphic, sedimentary, and sedimentary young) maps were
created, and from these flow length, TMI, and solar radiation at equinox were used in ArcView Map Calculator to predict the log-likelihood ratio of species presence/absence. This was transform to a
predicted probability of occurrence. The first prediction map, Qutopred2, derived from manual variable
selection (Figure 14; 14%), produced a smaller area of high habitat suitability than Qutopred3 (Figure 15; 29%), which was based on stepwise variable selection. However, suitable habitat is predicted to occur in
roughly the same locations.
A cross-tabulation of the thresholded binary maps of predicted suitable habitat based on these two models
with the map of existing vegetation is presented in Table 7. Raster cells with probabilities greater than the
threshold value of 0.10 (approximately equal to prevalence of the species in the sample which was 9%)
were assumed to be suitable habitat. Fourteen and twenty-nine percent of the island was found to be suitable using the manual and stepwise results, respectively, at this threshold. The majority of the area
predicted to be suitable for Q. tomentella currently supports coastal sage scrub, fewer are in Island
88
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Figure 14. Spatial prediction for Q. tomentella from the GLM based on manual variable selection. Mapped oak stands (polygons) are also shown in blue (Qutopoly.shp).
Figure 15. Spatial prediction for Q. tomentella from the GLM based on stepwise variable
selection. Mapped oak stands (polygons) are also shown in blue (Qutopoly.shp).
89
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Table 7 Existing vegetation (from vegetation map) found in areas of predicted suitable habitat for Q.
tomentella. Only values of 0.5% or higher are presented.
Vegetation Type (percent area of
island)
Percent of predicted
area
Area predicted (ha)
Qutopred21 Qutopred32 Qutopred2 Qutopred3
Coastal Sage Scrub (38.1%) 35.4% 38.8% 990 2,197
Island Chaparral (29.4%) 30.3% 30.0% 846 1,695
Grassland (19.5%) 25.8% 19.1% 721 1,082
Bare (9.4%) 5.4% 7.9% 152 445
Non-native herbaceous (0.5%) 0.8% 0.9% 24 52
Developed (1.1%) 0.0% 0.8% 0 47
Island Woodland (0.5%) 0.7% 0.7% 20 38
Southern Riparian Woodland (0.3%) 0.6% 0.6% 18 34
1 GLM manual selection, probability of presence .15 or higher
2 GLM stepwise selection, probability of presence .15 or higher
chaparral, and still fewer are in grassland, according to the map of existing vegetation. The areas of
predicted suitable habitat appear to be located randomly (proportionally) with respect to existing
vegetation. Both occur in similar proportions and a chi-squared test shows the difference being non-
significant (p=0.904 and .994; quotopred2 and qutopred3, respectively).
DISCUSSION
The fine resolution of the DEM used to generate explanatory variables, the mapping accuracy of the plots
and distributions, and the random location of the plots makes this an ideal data set with which to perform
modeling (Franklin et al. 2000). Presence/absence modeling was found to be much more robust (optimistic AUC between 0.85 and 0.998) than modeling of species parameters such as cover, density,
and size (7% to 38% variance explained). In other words, multiple regression models of the species
abundance and fitness variables, although they identified significant predictors, did not explain much of
the variability in those variables, as is often the case with ecological datasets. However, predictive models of species occurrence performed well in discriminating known locations of species presence and
absence. The two oak species are discussed separately below.
Quercus pacifica models
The two most explanatory models for Quercus pacifica were for presence/absence and, to a lesser extent,
cover. Greater abundance (cover) was associated with moderately high elevations (above 400 m), low A-horizon pH, shallow and silty soils, north-facing slopes, and locations in the island interior (greater
distance to the coast). Larger average tree size within plots was associated with intermediate values of A
horizon NH4+, steeper slopes, and high flow accumulation (lower drainage basin position), but very little
variation in size was explained.
Classification tree models of Q. pacifica presence/absence data (AUC 0.85) indicate that this species tends to be present on low-radiation north-facing slopes, or sometimes in higher radiation environments if
lower on the hillslope or if on certain geological substrates. GLMs of suitable Q. pacifica habitat based on
multiple variables (AUC 0.89 and 0.92 for manual and stepwise variable selection) predict the highest
probability of species presence at intermediate values of elevation, slope steepness, TMI, radiation, and
90
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
distance from the coast (flow length). Less suitable habitat is associated with igneous intrusive and
sedimentary substrates.
These results, taken together, indicate that Q. pacifica dominates at intermediate levels of moisture
(represented by intermediate TMI, radiation, and soil texture), and prefers north-facing slopes. This is
supported by the correlation of larger size at higher flow accumulation and northness, and is consistent with the findings of other researchers (Bauer 1936, Jones et al. 1993, Westman 1991). Yet, Q. pacifica
also reaches high cover at higher elevations and greater distances from the coast; this higher (400-600 m)
elevation zone may be particularly dry because it is above the fog layer, but not high enough to benefit from orographic precipitation (Miller 1979).
Abundance of juveniles in Q. pacifica plots was positively associated with intermediate canopy cover, low A-horizon NO3
-, and low flow length (nearer to the coast) and negatively correlated with number of
dead trees, but very little variance was explained. Dead trees were more abundant in stands with higher
A-horizon NO3-, but this correlation was weak. The associations with higher nutrients may be indirectly
linked with competition from invasive annual grasses, which thrive in high nutrient environments (Brooks 2003), and have been implicated in poor oak recruitment in California (Gordon et al. 1989, Danielson and
Halvorson 1991, Gordon & Rice 2000). Additionally, higher recruitment at intermediate canopy cover
levels has been found with blue oak (Quercus douglasii) saplings (Swiecki et al. 1997), and may reflect a trade-off between greater success in open areas (Stratton, this volume) and the presence of nearby
propagule sources.
The predictive map produced using a GLM based on stepwise variable selection covered only four
percent of the island; this is suspicious given that a recent, refined Quercus pacifica map (D. Knapp,
unpubl. data) shows it covering 23% of the island. Manual variable selection produced a closer, more
reasonable estimate (17%), therefore Qupapred1 should be used for management planning rather than Qupapred2.
Quercus tomentella models
Models of Quercus tomentella presence/absence had high discriminatory power (bootstrapped AUCs >
0.95). They indicate that this species tends to be present on igneous and sedimentary-derived substrates
with a gravelly and medium texture, at high values of TMI, intermediate elevations, slope angles and flow lengths (distance to coast) and low to intermediate levels of solar radiation. The preference of low-
radiation (north-facing) slopes corresponds well with the findings of other researchers (Jones et al. 1993;
Kindsvater 2006).
Q. tomentella tree density was highest at moderately high elevations (300-350 m), high values of winter
radiation, steeper slopes, and lower flow accumulation. However, explained variance was only 20-30% in the best (non-linear and non-parametric) models. All multiple predictor models for Q. tomentella
reproduction suggest that regeneration is highest on concave slopes, and at higher elevations. Model
results showing a trend for higher reproduction on gravelly, medium-textured soils correlates well with
results of Kindsvater (2006), who found a higher percent sand content in sites with Q. tomentella seedlings, along with higher pH, higher exchangeable potassium, and lower phosphorous contents.
Explained variance was 17-34% and was higher for nonlinear models. The strong spatial structure of the
oak populations on the island, with two metapopulations much larger in size and with greater reproduction than all others, has likely confounded these relationships, however.
Predicted locations for Q. tomentella were found to be currently dominated by coastal sage scrub, grassland, and Island chaparral, in descending order (Table 7). This is in contrast with the results of a
study on Santa Rosa Island, where the three habitat types most frequently overlapping with a predictive
91
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
GIS model of core Q. tomentella habitat were closed-cone pine woodland, grassland, and bare ground
(Kindsvater 2006). These differences may reflect differences in current vegetation cover, island geology and a higher prevalence of bare and degraded habitat due to overgrazing on Santa Rosa Island. This is a
relict species, more widely distributed in a cooler, wetter periods (Muller 1967). Its current distribution,
as well as that of the other vegetation types, is likely the complex result of changing climate, severe
overgrazing, impacts of introduced species, and soil loss.
Suggestions for future study
Q. pacifica dieback remains an unexplained phenomenon, although hypothesized factors include: reduced
moisture (exacerbated by groundwater reduction, and expressed in drier microtopographic areas); smog
drift from the mainland; a pathogen; or senescence (or any combination of these). If moisture relations are significant, presence/absence modeling of dieback areas should indicate this; such an analysis would be
helpful in elucidating potential causes and solutions.
Incorporating mapped temperature and precipitation variables, which have been found to be important in analyses of other chaparral species, would likely improve the models. The final weather stations were
installed in 2007, in an array that should be adequate to interpolate climate variables for the entire island
once sufficient data is collected. Models can then be developed using more direct or proximal bioclimatic variables instead of their topographic proxies, elevation and distance to the coast. Maps of feral and
managed animal disturbance intensity would also be useful, which could be produced using grid layers of
numbers of feral animals removed, as well as management history data for various fenced zones on the island.
It would be particularly informative to conduct an experimental outplanting using predicted locations for
Quercus tomentella, distributed among the coastal sage scrub, grassland, and Island chaparral vegetation types. Differential success among these sites would contribute to our understanding of the environmental
preferences of this species.
ACKNOWLEDGMENTS
We would like to thank the Seaver Institute for funding this work and supporting the environmental
variable generation, Jenny McCune and Lauren Danner, who tackled the dense oak plots (and Q. tomentella survey, by Jenny) with attention to detail and good humor. Mike Klinefelter was invaluable
for help in the environmental variable GIS-layer generation, and UCR geologists aided him with
generalization of the island‟s geology map. All participants in the workshop provided useful feedback on this project. Comments by Kathryn McEachern greatly improved the clarity of the manuscript.
LITERATURE CITED
Bittman, R. 2001. The California Natural Diversity Database: A natural heritage program for rare species
and vegetation. Fremontia 29:57-62.
Brooks, M.L. 2003. Effects of increased soil nitrogen on the dominance of alien annual plants in the Mojave Desert. Journal of Applied Ecology 40:344-353.
Danielsen, K.C. and W.L. Halvorson. 1991. Valley oak seedling growth associated with selected grass
species. Pp. 9-13 in: R.B. Standiford, tech. coord., Proceedings of the symposium on oak woodlands and hardwood rangeland management. Pacific Southwest Research Station Gen. Tech. Report PSW-
12, Berkeley, CA.
Davis, F.W., S. Goetz, et al. 1986. The use of digital satellite and elevation data in chaparral ecosystems research. Pp. 19-27 in: J.J. DeVries, ed. Proceedings of the chaparral ecosystems research conference.
California Water Resources Center, University of California, Davis.
92
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Davis, F.W. and S. Goetz. 1990. Modeling vegetation pattern using digital terrain data. Landscape
Ecology 4: 69-80. Davis, S.D., R.B. Pratt, et al. 2007. Freezing tolerance impacts chaparral species distribution in the Santa
Monica Mountains. Pp. 159-172 in: D.A. Knapp, ed. Flora and ecology of the Santa Monica
Mountains. Southern California Botanists Special Publication No. 4, Fullerton, CA.
Franklin, J. 1998. Predicting the distribution of shrub species in southern California from climate and terrain-derived variables. Journal of Vegetation Science 9:733-748.
Franklin, J., P. McCullough, et al. 2000. Terrain variables used for predictive mapping of vegetation
communities in southern California. Pp. 331-353 in: J. Gallant and J. Wilson, eds. Terrain analysis: principles and applications. John Wiley & Sons, New York, NY.
Gordon, D.R., J.M. Welker, et al. 1989. Competition for soil water between annual plants and blue oak
(Quercus douglasii) seedlings. Oecologia 79:533-541. Gordon, D.R. and K.J. Rice. 2000. Competitive suppression of Quercus douglasii (Fagaceae) seedling
emergence and growth. American Journal of Botany 87:986-994.
Griffin, J.R. 1988. Oak Woodland. Pp. 383-415 in: M.G. Barbour and J. Major, eds. Terrestrial vegetation
of California (2nd
edition). California Native Plant Society Special Publication No. 9, Sacramento, CA.
Harrison, A.T., E. Small, et al. 1971. Drought relationships and distribution of two Mediterranean-climate
California plant communities. Ecology 52: 869-875. Jones, J.A., S.A. Junak, et al. 1993. Progress in mapping vegetation on Santa Cruz Island and a
preliminary analysis of relationships with environmental factors. Pp. 97-104 in: F.G. Hochberg, ed.
Proceedings of the third California Islands symposium. Santa Barbara Museum of Natural History, Santa Barbara, CA.
Keeley, J.E. and F.W. Davis. 2007. Chaparral. Pp. 339-366 in: M.J. Barbour et al., eds. Terrestrial
vegetation of California (Third edition). University of California Press, Berkeley, CA.
Keeley, J.E., C.J. Fotheringham, et al. 2005. Factors affecting plant diversity during post-fire recovery and succession of Mediterranean-climate shrublands in California, USA. Diversity and Distributions
11:525-537.
Kindsvater, L.C. 2006. Conservation and restoration of the endemic Island oak, Quercus tomentella, in Channel Islands National Park using a habitat approach. Ph.D. dissertation. University of California,
Davis.
Kirkpatrick, J.B. and C.F. Hutchinson. 1980. The environmental relationships of Californian coastal sage
scrub and some of its component communities and species. Journal of Biogeography 7:23-38. Knapp, D.A. 2005a. Vegetation mapping on Santa Catalina Island using orthorectification and GIS. Pp.
193-203 in: D.K. Garcelon and C.A. Schwemm, eds. Proceedings of the sixth California islands
symposium. National Park Service Technical Publication CHIS-0501, Institute for Wildlife Studies, Arcata, CA.
Kummerow, J. and R. Mangan. 1981. Root systems in Quercus dumosa Nutt. dominated chaparral in
southern California. Acta Oecologica 2: 177-188. McCune, J. 2005. Report on the census and survey of Island oak (Quercus tomentella Engelm.) and
canyon live oak (Quercus chrysolepis Liebm.) groves on Catalina Island, 2004 and 2005.
Unpublished report prepared for the Catalina Island Conservancy, Avalon, CA.
Miller, P.C. 1979. Quantitative plant ecology. Pp. 179-232 in: D. Horn, G.R. Stairs et al., eds. Analysis of ecosystems. Ohio State University Press, Columbus, OH.
Miller, P.C., D.K. Poole, et al. 1983. The influence of annual precipitation, topography, and vegetative
cover on soil moisture and summer drought in southern California. Oecologia 56:385-391. Mueller-Dombois, D. and H. Ellenberg. 1974. Aims and methods of vegetation ecology. John Wiley &
Sons, New York, NY.
Muller, C.H. 1967. Relictual origins of insular endemics in Quercus. Pp. 73-77 in: R.N. Philbrick, ed. Proceedings of a symposium on the biology of the California islands. Santa Barbara Botanical
Garden, Santa Barbara, CA.
93
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.
Ng, E. and P.C. Miller. 1980. Soil moisture relations in the southern California chaparral. Ecology 61: 98-
107. Nixon, K. and C.H. Muller. 1994. New names in California oaks. Novon 4: 391-393.
Poole, D.K. and P.C. Miller. 1981. The distribution of plant water stress and vegetation characteristics in
southern California chaparral. American Midland Naturalist 105:32-43.
Swiecki, T.J., E.A. Bernhardt, and C. Drake. 1997. Factors affecting blue oak sapling recruitment. Pp. 157-167 in: Proceedings of a symposium on oak woodlands: Ecology, management, and urban
interface issues. USDA Forest Service Gen. Tech. Rep. PSW-GTR-160, Berkeley, CA.
Wells, P.V. 1962. Vegetation in relation to geological substratum and fire in the San Luis Obispo Quadrangle, California. Ecological Monographs 22: 79-103.
Wells, P.V. 1968. New taxa, combinations, and chromosome numbers in Arctostaphylos (Ericaceae).
Madroño 19:193-224. Westman, W.E. 1981. Diversity relations and succession in Californian coastal sage scrub. Ecology 62:
170-184.
Westman, W.E. 1991. Measuring realized niche spaces: Climatic response of chaparral and coastal sage
scrub. Ecology 72: 1678-1684.
94
Oak ecosystem restoration on Santa Catalina Island, California: Proceedings of an on-island workshop, February 2-4, 2007. Edited by D.A. Knapp. 2010. Catalina Island Conservancy, Avalon, CA.