Ecological Archives E088-173-A2

D. Richard Cutler, Thomas C. Edwards, Jr., Karen H. Beard, Adele Cutler, Kyle T. Hess, Jacob Gibson, and Joshua J. Lawler. 2007. Random forests for classification in ecology. Ecology 88:2783–2792.

Appendix B. Data descriptions and details of data preprocessing.

This appendix contains tables giving descriptions of all the variables used in our three example analyses, and also information on the preprocessing of the data.

INVASIVE PLANTS IN LAVA BEDS NATIONAL MONUMENT, CALIFORNIA, USA

Data Sources

Lava Beds National Monument (NM) personnel provided us with data on detections and treatment in 2000–2005 for Verbasum thapsus (common mullein), Urtica dioica (nettle), Marrubium vulgare (white horehound), and Cirsium vulgare (bull thistle), as well as GIS layers for roads and trails in Lava Beds NM. For data analysis purposes, we imposed a 30-m grid over Lava Beds NM and a 500 m buffer outside the park. There were a total of 244,733 grid points in the park and buffer. Values of all the topographic and bioclimatic predictor variables were obtained for all points on the grid and minimum distances to roads and trails for all points on the grid were computed in a GIS and merged with the other predictor variables. Table B1 contains a list of names and descriptions of variables used in our analyses of the Lava Beds NM invasive plant data. Topographic variables (Elevation, Aspect, and PercentSlope) were obtained from the National Elevation Dataset (NED) (Gesch et al. 2002). Bioclimatic variables were obtained from the DAYMET 1 km grid daily weather surfaces (Thornton et al. 1997) by interpolation. Daily values for each variable were aggregated to create monthly variables. Thus for each bioclimatic predictor in Table B1 (except DegreeDays) there were originally 12 monthly values.

TABLE B1: Names and descriptions of predictor variables used in analyses of invasive plant data from Lava Beds National Monument, California, USA.

Variable
type

Variable
name

Variable
description

Units

Bioclimatic

DegreeDays

Degree days

°C days

 

EvapoTrans

Monthly potential evapotranspiration

mm

 

MoistIndex

Monthly moisture index

cm

 

Precip

Monthly precipitation

cm

 

RelHumid

Monthly relative humidity

%

 

PotGlobRad

Monthly potential global radiation

kJ

 

AveTemp

Monthly average temperature

°C

 

MinTemp

Monthly minimum temperature

°C

 

MaxTemp

Monthly maximum temperature

°C

 

DayTemp

Monthly average daytime temperature

°C

 

AmbVapPress

Monthly average ambient vapor pressure

Pa

 

SatVapPress

Monthly average saturated vapor pressure

Pa

 

VapPressDef

Monthly average vapor pressure deficit

Pa

       

Topographic

PercentSlope

Percent slope

%

 

Aspect

Aspect

°

 

Elevation

Elevation

m

       

Distances to
Roads or Trails

DistRoad

Distance to the nearest road

m

DistTrail

Distance to the nearest trail

m

 

DistRoadTrail

Distance to the nearest road or trail

m

Preliminary Data Analyses and Processing

All preliminary data summaries and statistical analyses were carried out in SAS version 9.1.3 (SAS Institute, Cary, North Carolina). The variable aspect is measured on a circular scale. Values of aspect near 0° and 360° both represent directions close to due north, yet are at opposites ends of the aspect scale. The usual remedy for this problem is to apply a trigonometric transformation to the raw aspect values. The transformation we used is given by the formula Transformed Aspect = [1 - cos(2piAspect/360)]/2. This formula is similar (but not identical) to that used by Roberts and Cooper (1989). The transformed aspect values lie in the interval from 0 to1. Values near 0 represent aspects close to due north, while values near 1 represent aspects close to due south. East and west are treated identically with transformed aspect values of 0.5.

Preliminary analyses showed that correlations among the monthly values for each of the 12 sets of bioclimatic predictor variables (excluding DegreeDays) were extremely high. Principal components analyses of the correlation matrices of the 12 sets of bioclimatic variables showed that, in each case, the first principal component was approximately an average of the 12 monthly measurements, and the second principal component was a contrast of values for the 6 “summer” months (April–September) to the 6 “winter” months (October–December and January–March). For each set of 12 monthly variables, these two principal components explained over 95% of the variability, and in most cases the first two principal components explained over 99% of the variability in the sets of variables. Accordingly, for each set of monthly bioclimatic predictors, we defined two new variables:

1. the average of the 12 monthly variables, and

2. the difference between the sum of the April–September monthly values and the October–December and January–March monthly values, divided by 12.

We use the suffix “Ave” to indicate the average of the 12 monthly values of each bioclimatic predictor and “Diff” to indicate the Summer–Winter difference. For example, PrecipAve is the average precipitation over all 12 months and PrecipDiff is the normalized difference between summer and winter precipitations, as described above. In all our classification analyses we used only the pair of derived variables for each bioclimatic predictor, not the 12 monthly values. Thus there were 12 pairs of bioclimatic predictors, DegreeDays, three topographic variables (using transformed Aspect instead of raw Aspect), and three variables containing distances to roads and trails, for a total of 31 predictor variables.

LICHENS IN THE PACIFIC NORTHWEST, USA

Data Sources

The lichen data sets were collected in seven national forests and adjoining BLM lands in Oregon and southern Washington USA, between 1993 and 2003. The Current Vegetation Survey (CVS) is a randomly started 5.4 km grid that covers all public lands in the Pacific Northwest. On all public lands except designated wilderness areas and national parks, the primary grid has been intensified with 3 additional grids spaced at 2.7 km from the primary grid. The primary purpose of the CVS grid is to generate estimates of forest resources (Max et al. 1996). The Lichen Air Quality (LAQ) data were collected as part of a study to evaluate air quality in the Pacific Northwest (Geiser 2004). The data used in our analyses is from 840 sites on the primary CVS grid. The pilot random grid surveys (EVALUATION) were conducted by the Survey and Manage Program as part of the Northwest Forest Plan, the conservation plan for the northern spotted owl (Strix occidentalis caurina). The EVALUATION surveys were conducted in three areas in the Pacific Northwest: Gifford-Pinchot National Forest in southern Washington, the Oregon Coast Range, and the Umpqua Basin, also in Oregon. At each location, a stratified random sample of 100 sites from the intensified CVS grid was drawn. The stratification criteria were Reserve Status (Reserve, Non-reserve) and Stand Age Class (< 80 years and 80+ years). The allocations of the sampled sites to the strata were (at each location): 60 to Reserve/80+, 20 to Reserve/< 80, and 10 to each of the Non-reserve strata. These allocations reflected the information priorities of the Survey and Manage program at the time of the surveys.

The four lichen species used in our analyses–Lobaria oregana, Lobaria pulmonaria, Pseudocyphellaria anomala, and Pseudocyphellaria anthraspis–were the four most common species observed in the LAQ surveys that were also searched for in the EVALUATION surveys. All four species are large, foliose, broadly distributed cyanolichens that can be found on tree trunks, live branches, and leaf litter of conifers in the Pacific Northwest. All achieve their largest biomass in riparian and late seral forests. Eye-level habitat and large size makes them relatively easy to find and identify. All sites in both surveys were surveyed by field botanists trained in the recognition and differentiation of regional epiphytic macrolichens, and specimens were obtained at all sites for later laboratory identification.

Table B2 contains the names and descriptions of the predictor variables used in our analyses. Topographic variables were obtained from the NED (Gesch et al. 2002). Aspect was transformed according to the formula Transformed Aspect = [1 - cos(2pi(Aspect – 30°)/360)]/2, following Roberts and Cooper (1989). Daily values of the DAYMET bioclimatic predictors (Thornton et al. 1997) were aggregated to monthly values. Correlations among the monthly bioclimatic predictors were very high and principal components analyses suggested that considerable dimension reduction could be carried out without loss of information. For the variables EvapoTrans, MoistIndex, Precip, RelHumid, and PotGlobRad, the 12 monthly values for each observation were replaced by an average of the 12 values, denoted with the suffix “Ave” on the variable name, and a difference of the average values for the six “summer” months (April–September) and the six winter months (January–March, October–December), denoted with the suffix “Diff.”. For the temperature and vapor pressure measurements, further dimension reduction was possible for the LAQ and EVALUATION data. The 24 vapor pressure measurements at each site were replaced with just two values: an average of all 24 values and a summer-to-winter difference. Similarly, the 48 temperature measurements were replaced by just two values, again an average and a summer-to-winter difference. In all our classification analyses we used only the derived bioclimatic variables. The total number of predictor variables used in our analyses was 24.

TABLE B2: Names and descriptions of predictor variables used in analyses of lichen data from the Pacific Northwest, USA.

Variable
type

Variable
name

Variable
description

Units

Bioclimatic

EvapoTrans

Monthly potential evapotranspiration

mm

MoistIndex

Monthly moisture index

cm

Precip

Monthly precipitation

cm

RelHumid

Monthly relative humidity

%

PotGlobRad

Monthly potential global radiation

kJ

AveTemp

Monthly average temperature

°C

MinTemp

Monthly minimum temperature

°C

MaxTemp

Monthly maximum temperature

°C

DayTemp

Monthly average daytime temperature

°C

AmbVapPress

Monthly average ambient vapor pressure

Pa

SatVapPress

Monthly average saturated vapor pressure

Pa

Topographic

PercentSlope

Percent slope

%

Aspect

Aspect

°

Elevation

Elevation

m

Stratification

ReserveStatus

Reserve Status (Reserve, Non-reserve)

StandAgeClass

Stand Age Class (< 80 years, 80+ years)

Vegetation

AgeDomConif

Age of the dominant conifer

Years

PctVegCov

Percent vegetation cover

%

PctConifCov

Percent conifer cover

%

PctBrdLfCov

Percent broadleaf cover

%

ForBiomass

Live tree (> 1inch DBH) biomass,
Above ground, dry weight.

tons/acre

CAVITY NESTING BIRDS IN THE UINTA MOUNTAINS, UTAH, USA

Data Source

Data was collected on nest sites for three cavity nesting birds in the Uinta Mountains, Utah, USA. The species Sphyrapicus nuchalis (Red-naped Sapsucker)(n = 42), Parus gambeli (Mountain Chickadee)(n = 42), and Colaptes auratus (Northern Flicker)(n = 23). Within the spatial extent of the birds nest sites for the three species, 106 non-nest sites were randomly selected, and the same information as for the nest sites was collected. The plot size for the data collection was 0.04 ha. The variables used in our analyses are given below.

TABLE B3: Stand characteristics for nest and non-nest sites for three cavity nesting bird species. Data collected in the Uintah Mountains, Utah, USA.

Variable
name

Variable
description

NumTreelt1in

Number of trees less than 2.5 cm diameter at breast height

NumTreel1to3in

Number of trees 2.5 cm to 7.5 cm diameter at breast height

NumTree3to6in

Number of trees 7.5 cm to 15 cm diameter at breast height

NumTree6to9in

Number of trees 15 cm to 22.5 cm diameter at breast height

NumTreelt9to15in

Number of trees 22.5 cm to 37.5 cm diameter at breast height

NumTreelgt15in

Number of trees greater than 37.5 cm diameter at breast height

NumSnags

Number of snags

NumDownSnags

Number of downed snags

NumConifer

Number of conifers

PctShrubCover

Percent shrub cover

StandType

Stand Type (0 = pure aspen; 1 = mixed aspen conifer)

LITERATURE CITED

Geiser, L. 2004. Manual for monitoring air quality using lichens in national forests of the Pacific Northwest. USDA Forest Service, Pacific Northwest Region, Technical Paper R6-NR-AQ-TP-1-04. URL:http://www.fs.fed.us/r6/aq.

Gesch, D., M. Oimoen, S. Greenlee, C. Nelson, M. Steuck, and D. Tyler. 2002. The National Elevation Dataset. Photogrammetric Engineering and Remote Sensing 68:5–12.

Max, T. A., H. T. Schreuder, J. W. Hazard, J. Teply, and J. Alegria. 1996. The Region 6 vegetation inventory monitoring system. General Technical Report PNW-RP-493, USDA Forest Service, Pacific Northwest Research Station, Portland, Oregon, USA.

Roberts, D.W., and S. V. Cooper. 1989. Concepts and techniques in vegetation mapping. In Land classifications based on vegetation: applications for resource management. D. Ferguson, P. Morgan, and F. D. Johnson, editors. USDA Forest Service General Technical Report INT-257, Ogden, Utah, USA.

Thornton, P. E., S.W. Running, and M. A. White. 1997. Generating surfaces of daily meteorological variables over large regions of complex terrain. Journal of Hydrology 190:214–251.



[Back to E088-173]