PEER.REVIEWED
ARIICLE
Forest
Mapping
Volcanic
at Lassen
National
Park,Galifornian
UsingLandsat
TMDataanda Geographical
Information
System
JosephD. White,GlennC. Kroh,and JohnE. Pinderlll
Abstract
Knowledge of forest species composition is on integral part
of designing and implementing resource manangement policies in a national park. Managers musf rely on cost-effective
methods of vegetotion mopping, namely, use.of rentotely
sensed data coupled with digital geographic data, to help
them meet their management goals. In this study, we demonstrate that genus-level maps can be generated from unsupervised classifications of Landsat rM data at an accuracy level
of 73 percent. SpeciesJevel maps can be created to on occuracy level of 58 percent by post-stratification of the spectral
classification with topographic data in a geographic information systent (cts). This modification method is a rule-based
system whereby spectral forest classes are sorted based on elevation and soil-moisture gradients established for each species through ecdogicd reseorch. Our obseruotions illustrote
that spectral classification is optimized by using all six reflective rM bands and that classification accurocy is affected
by conopy cover ond understory vegetation. Modifying spectral classifications by environmental data in a GIS is a useful
way of defining species composition of forests in an area
where access to forests is limited but need for map information is great.
Introduction
Reduction of forest and woodland habitats in North America
has stimulated development of resource management tools to
assist future forest conservation. Preservation of these forests
is vital because of their role in global carbon cycling and
species diversity (Bormann, 1985). Anthropogenic disturbance such as air pollution, acid rain, and global climate
change could affect the functioning, distribution, and survival of plant and animal species in their present ranges. National parks, national forest wilderness areas, and private
reserves such as those held by the Nature Conservancy provide refr.rgesfor these species and their habitats. Automated
phytocartography from remotely sensed data coupled with
J. D. White and G. C. Kroh are in the Department of biology,
Texas Christian University, P.O. Box 32916, Ft. Worth, TX
761,29.
f. D. White is presently with the Numerical Terradynamic
Simulation Group, School of Forestry, University of Montana. Missoula. MT 59812.
J. E. Pinder III is at the Savannah River Ecology Laboratory,
Drawer E. Aiken. SC 29081.
PE&RS
geographic
databases
can play an importantrole in docupresentplant speciesdistributionin theseprotected
il:lJl"t
Computer-aided classification of remote sensing information has become a standard tool for mapping vegetation
across large areas (Iverson et al.,19Bg). Data from Landsat
and airborne radiometers have been used to identifv the species composition of various deciduous and coniferous forest
ecosystems (Likens et ol.,],982; Khorram and Katibah, t98+;
Benson and DeGloria, 1985; Hughes et al., tg96; Skidmore,
1989). Mayer and Fox (1981) discriminated lodgepole pine,
ponderosa pine, and fir at a rate of 91 percent using Multispectral Scanner (vss) data at the McCloud Ranger District
near Mt. Shasta, California. Walsh (19s0) mapped forests
composed of lodgepole pine, ponderosa pine, white fir,
shasta red fir, and mountain hemlock at Crater Lake National
Park using classification of I'ass data with an overall accuracy
Ievel of BB.Bpercent. Moore and Bauer (1000) reported accuracies of 63 to 67 percent mapping various Minnesota conifer
species such as red pine, jack pine, black spruce, white pine,
and tamarack using the Landsat Thematic Mapper (rv). Shen
et al., (198s) mapped the same species using TM simulator at
68 percent. The higher mapping performance of the studies
using MSS over TM sensors is notable considering the refined
spectral and spatial resolution of the rrrl sensors; however,
this may be function of both the environmental conditions and spectral characteristics of those oarticular forest species.
Typically, previous mapping projects have utilized supervised classification techniques. Supervised classification
can be costly in terms of time and money spent selecting and
sampling training sites to represent the range of morphological and topographical variation of a particular mapping category (Walsh, 1980; Mayer and Fox, 1981). Unsupervised
classification of spectral data may be preferable when mapping large areas with a rugged terrain (Fleming et al., 1.s75).
Terrain effects, expressed as topographic data, may be used
to modify remote sensing classifications to map forested areas in which species are not easily distinguished spectrally
(Skidmore, 1989; Richards et o/., 1982; Cibula and Nyquist,
Photogrammetric Engineering & Remote Sensing,
Vol. 61, No. 3, March 1995, pp. 299-305.
0 0 9 s - 11 1 2 l 9 5 / 6 1 0 3 - 2 9 9 $ 3 . 0 0 / 0
O 1995 American Society for Photogrammetry
and Remote Sensine
PEER
REVIEWED
80
2
o ..]
r..U
f
\
501
a
Q
LlJ
z
r
cc
.'l\
\
;:La-
P jef lrcy
+
P.c.,ttttu
\
57
TM Band
-
A.twillil
-1
/1 ,ktrililiu
Figure1. Spectralresponseof four dominanttree species of LassenVolcanicNationalParkacrossthe reflect i v eT M b a n d s .
1987).Bolstadand Lillesand (1992)mapped forestspecies
between 73 and 89 percentby coupling tM data with soil
and topographic data.
The purposeof this study is to demonstratethat (1) genus-levelmapping of mixed conifer forestscan be accomplished using simple, unsupervisedclustering of Landsat Ttr,t
data and (2) forestspeciescan be identifiedby modifying
spectral classificationwith topographic information through
a geographicinformation system (crs). The importance of
this study is to show that valid forestcompositionmaps can
be producedusing readily availablespectraland imageprocessing/ctssoftware.Use of this technologyby resourcemanagersin other protected areaswill increaseour knowledge
about forestsat present and can ultimately be used to monitor the changein forest composition over time (Vande Castle,
1991).Also, compilationand integrationof vegetationmaps
into a geographicdatabasefrom many public and private reserveswould expand our understandingof forestspeciesdistribr.rtionsfrom local to resional scales.
Methods
The Earth Resource Data Analysis System fnnnas) software
was used to perform an unsupervised classification of Landsat TM data covering the forests of Lassen Volcanic National
Park. Two techniques were tested for their usefulness in
identifying conifer tree species. In addition, a GISmodel
based on topographic data was developed to differentiate
species that could not be differentiated using only spectral
data. The format of the mapping process consisted of (1)
classification of the spectral data, (2) interpretation, (3) modification of the spectral classification using habitat variables
in the GIS environment, and [+) accuracv assessment.
Site Descilption
Lassen Volcanic National Park covers approximately 500 km,
of the southernmost peaks of the Cascade Mountain range.
Elevation in the park varies from 1616 m at Warner Valley to
300
ARTICTE
3187 m on Lassen Peak. Most of the park below 2400 m is
forested, with the distribution of forest species being affected
by altitude (Gillette et a1., 1961). Abies magniltco (red fir)
and Pinus conlorlo var. murroyona (lodgepirle pine) dominate upper elevations (2100 to 2400 m), whereas A. concolor
(white fir) and P. ieffreyi (|effrey pine) are most abundant at
Iower elevations (<2100 m). Limited stands of Tsuga tnertensiono (mountain hemlock) occur along the treeline>2400 m.
Other vegetation communities occurring in the park include (1) montane chaparral dominated bv Arctostaphvlos
potula. Coslonopsis sempervirens,and Cetnot hus velitinus:
and (2) seasonally wet habitats located in valley meadows
and along streams and lake margins. Alnus tenuifolio and
Cnrex spp. are common constituents of these saturated areas
(Bailey,1963).
DataProcessing
Ttradata were acquired for 12 June 1987 at approximately 9:
30 atrawith cloud-free conditions. This date was selected to
optimize the photosynthetic potential of the tree vegetation
(as well as possible spectral characteristics) while minimizing snow cover at higher elevations. A subset of 935 by 6bB
pixels containing Lassen Park was extracted from the original
Landsat data for classification. This subset was linearly georectified using forty ground control points selected from
USGS1:24,000-scalequadrangle maps to a root-mean-square
error (Rvsn) of one pixel in the horizontal plane.
Errors related to reflectance path and rindom sensor
malfunction result in radiometric noise in the data (fensen,
1986; Conese ef o.1.,19BB). To reduce the confoundine effects
of this noise, a 3- by 3-pixel low pass filter was applGd to
the data.
SatelliteDataClassification
andInterpretation
The TM data were classified using a clustering algorithm off e r e d i n t h e E R D A Ss o f t w a r e .B a n d s 1 , 2 , 3 , 4 , 5 , a n d Z w e r e
classified into 20 clusters using the CLTJSTR
program. CLUSTR
is a sequential classifier in which cluster means are calculated from the data just as it occurs in the data file, beginning with data in row one, column one and continuing to a
user defined limit. Classes, therefore, mav be affected bv the
spectral information contained within the first few rows of
data.
Interpretation of the 20 classeswas based on expert
knowledge of the forested areas supplemented with ierial
photography interpretation. Classeswere placed into Pinus,
Abies, non-forest, and non-vegetation categories. Non-forest
classes consisted of shrub and meadow domintated areas and
non-vegetation classes were bare ground, lava, water, and
snow. In cases where classes could not be resolved by this
method, a normalized difference vegetation index (rurivr) was
calculated based on the mean cluster values for bands 3 and
4 lbr each class. Classes with low NDVI values were assigned
as non-vegetation, medium values as forest, and high values
as non-forest vegetation. The range of NDvIs was established
from NDVIs of known classes.
Average DN values for ten homogeneous species stands
were collected for the six reflective bands to test the spectral
definition of each species. These data were plotted (Figure 1)
and Student's f test values were calculated to determine significant differences in spectral values between species and
g e n e r a .S p e c t r a l c l a s s e s - w e r eo n l y i n t e r p r e t a b l ea t t h e g e n u s
level based on this analysis.
PE&RS
PEER.REVIEWED
TnaLe1.
or Spectnnr Bnseo TRre Crnsses av ErevlrtoN.
SEPARATIoN
Elevation [m)
SpectralClass
Pinus
< 1830
1830 to 1920
1980 to 2180
2220 lo 2350
> 2350
> 2560
> 2620
P. jeffrevi
pine (non-specific)
pine Inon-specific)
pine (non-specific)
P. contortl
I. merfensrona*
P. albir:aulis*
Abies
A. concolor
A. concolor
fir (non-specific)
A. ntagnifico
A. ntagnifica
T. nrcrtensiano*
P. albicaulis"
*Identified but no accuracy assessed.
Itstt 2. SepnRertonor TRre SpectRnr/ElevnrtouCrnsses BY PorENrlAL
RD
o HT
artl TopocRepHtc
T o p o c R n p H tM
c o t s r u R eG R n o t e n t s( G E N E R A L I zr E
Motsrunr Cusses ro Tunee).
Tree/Elev. Im)
Classes
meslc
pine/1830to 192o
pine/1980to 2180
pinel2220 to 2350
fir/1980to 2180
P.
P.
P.
A.
XCTIC
contot'tl
contortI
contorta
concolor
P
P
P
A
iefJreyi
contoftl
contorta
tttuE;nifica
P. ieffreyi
P. jeffreyi
P. ieffreyi
A. ntagnifica
Usinga GISModel
Modification
Classification
A GIS model based on elevation, slope, aspect, and proximity
to a water source was used to subdivide the Pinus and Ables
classes from the spectral classifications to the species level.
Elevation and soif moisture value ranges which defined the
habitats of tree species were obtained from the literature (GiIl e t t e e f o 1 . ,r g o r i V a n k a t , 1 9 8 2 ; H o l l a n d , 1 9 8 6 ; R u n d e l e t a l . ,
1988).
Elevation for the GIS model was derived from 1:24'000scale elevation data (nau) obtained from the U.S. Geological
Survey. These contained errors of horizontal striping and
empty areas caused by scanning procedure-sused in their
geneiation (USGS, 1987). To reduce the effects of theses errors, a 3 by 3 row-wise filter,
0-1 0
0-1 0
0-1 0
was applied to the DrM to reduce striping' Missing data pointi were then identified in the data and replaced with the
irr".ug". of the surrounding elevation values (Stitt, 1990)' Elevatidns were grouped into 30 classes smoothed with a 5- by
5-pixel median value filter to remove errors persisting after
filierine the DEM data,
Poiential soil moisture was evaluated using a topographic moisture gradient (ruc) model suggestedby-Vankat
irs-Bz), Though other models exist for evaluating soil moisture (Beven and Wood, 1983; O'Loughlin, 1986), the rMG
model was chosen because of the (a) specificity of the model
to the Lassen forest species and (b) simplicity in construction-based topographic analysis' In this model, potential soil
moisture is expressed as a range of values from 1 (wettest) to
10 (driest) and is a function of slope, aspect, and position on
slopes. Stream vallevs are considered wettest and ridgetops
dri6st. The intermediate range is created as a combination of
aspect and slope. Slopes and aspects were computed using
the topographit module in sRnRs. Ridgetop and valley posi-
PE&RS
ARTICTE
tiorrs were clerived by applying a filter to the DEM data which
returned the average-heighi of-the peripheral pixels to the
center,
0.125
o.1,25
o.1.25
0.1.25
o
0.1.25
0.'125
0,125
0.125
termed D.,. Values of D,, were subtracted from the original
nsM data (D.) and scaled to a mid-range of t00 to obtain a
- D,, + 100). For D,
topographic position value D, (2, : D"
<'10"0, ihe pbsition is a lower slope or valley depending on
the differenie from 100, D, : 100, positions are midslopes'
and for rr > 100, positions are upper slopes and topographic
ridses. The file was recoded so that the resultant GIS file had
thr6e classes:valley, midslope, and ridge. Stream channels
were established by correlating low sloping areas with digitized perennial and lntermittent streams from the 7'5-minute
topographic maps. Streams flowing.through valleys with shilLow slopes would tend to greatly affect soil moisture beyond the confines of the stream channel (Gregory et o1',
1991J.
A species composition map was created from a poststratificition of the spectral classified images using the elevatiorr data and the TMG map. In a decision tree process, Pinus
and , bies categories of the spectral based classifications
were partially separated into species based on elevation
.ung"i. As species could not be distinguishe-d solely by.elevation due to broad overlap of ranges at mid-elevational leve l s ( G i l e t t e e t a \ . , 1 . 9 6 L ;H o l l a n d , l s 8 o ) ( T a b l e 1 ) , t h e r M G
variable was used to differentiate species at elevations where
separation rvas incomplete (Table 2).
Assessment
Accuracy
Map accuracy was assessedfor genus-level, spectral classification and the species-level, Gls-stratified map. A^ccuracy was
checked againstio field samples ranging in size from 240 by
240 m to iso to rso m for a total of 1580 pixels sampled.
Samples were subdivided into 30- by 30-m plots for plot-tooixei comparison. The variation in sample size was a function of sampling technique used. Generally, size had no
relation to variation in map accuracy at this scale' Species
and estimated coverage fof three vegetation layers in the forest canopy and other site variables including sloP.e,aspect,
and general site moisture conditions were assesedfor each
pixeiwithin sampled areas' Forest types were determined by
in the upper canopy in each 30- by 30-m plot.for .
"or.".ug"
gJnus and/oripecies' Most forests sampled were domieach
nated"by one forest iypg by at least 50 percent cover and
forests.
t y p e lorests.
p e c i e stype
i n g l esspecles
as
I d
e s i g n a t ea
d s ssingle
were
designated
on
r r geographically
uasuu u
*"re
ere u
chosen
l r u s e r r based
locauons w
Sample
S
a m D l e locations
SEUUrdPurLdrrJ
identifiable
trail, stream, or road crossings, and aerial photo-
graphs. Remote sampled areas which could not be identified
6n z.s-minute topographic maps were located using the Global Positioning Syste* (crs). Though a random sampling
regime is preferred to reduce the incidence of spatial autocoirelation associated with rv data, the steep, volcanic terrain of Lassen coupled with dense forests and underbrush
necessitated location of sampling areas in accessible areas'
Number maps were produced from the gelus- and-species-level mapt iot correiation with the field data' Pixel by
pixel counts were made to assess map to field aggreement
ind disasreement and were collected in a contingency table
format fo"r each map (Tables 3 and 4). Percent correct, com-
PEER.REVIEWED
fnars 3.
AccuRncyAssEssMENT
or rlE SpecrRirrCLnsstrrcrcrior.i.
TM SpectralClassification
Field Classes
LandsatClasses
Pinus
Abies
Non-forestINF)
Non-vegetation(NV)
PercentCorrect(%)
CommissionError (%)
Pinus
Abies
632
196
52
15
'lo4
^12
361
1B
7
I
3
2
4
5
66
7'1.
2S
74
26
81
19
B6
14
NF
NV
Map Statistics
n = 1580
PercentCorrect = 737o
CommissionEnor : 27o/o
ConfidenceInterval(95% ConfidenceLevel] : O.71to O.TE
K = 0.5850
dr = 0.0017
mission error, standard deviation, confidence interval (95
percent confidence level)(Thomas and Allcock, 1984), and
the Kappa statistic (r) and variance (o,") (Congalton ef o,1.,
1.983) were calculated for each table.
Results
andDiscussion
An analysis of the average spectral response of some of the
most heavily forested and species-homogeneoussample areas
demonstrates that marginal differences exist between- forest
species {Figure 1). For bands 1, z, and 3, the brightness values among species were nearly identical. The maximum differences among species occurred for bands 5 and 7 where
the pine species had greater brightness values than the fir
species. There was little interspecific difference between reflectance patterns between P. jiffreyi and P. contorta,bul
greater differences between the fir species, especially for
bands 5 and z, though not statistically signifi-ant. The higher
reflection in band 5 of pines over firs may be indicative of
the lower projected leaf area (mll of pines (Kaufmann ef a1.,
1982). The Lassen forest species could not be classified to
the species-level based on spectral clustering alone due to
the small spectral variation among species. Au*orre., genuslevel classification was possible based on the slight spectral
characteristic differences at this hieher taxonomic level.
Spectral information alone wai suitable for discriminating between forest genera. The spectral classification predicted Pinus and.4bres forest classesbv 71 percent and Z+
percent, respectively (Table 3). The claisificition u"".r.u"y
for the map on the whole was 73 percent, including non-iorested and non-vegetated areas. The Kappa value for this map
was lower at 0.59. The commission errors, or errors of inclusion, differed between Pinus and lbies classes from 29 oercent to 26 percent. Of these errors, 7 percent of the Pjnus
class were contributed from non-forest and non-vegetated
classes versus b percent in the Abies class. This strtht ditterence may be explained by prominence of understory vegetation in the pine forests. Analysis of understory vegetation
shows that understory cover is significantlv greater (P<0.05)
in Pinus forests than ,Abies forests.
The ability to discriminate forest species is positively
correlated with canopy closure (Franklin, 1986) and negatively correlated with understory vegetation prominence
(Spanner et a1., 1gS0) and reflection from soil (Ezra ef a1.,
1984). Sub-pixel heterogeneity, or variation in the land cover
302
ARTICI.E
at a scale less than sensor detection, may also be important.
Mean sample canopy and variance in understory cover are
plotted along an axis of decreasing classification accuracy
(percent) for forest genera from 20 selected samples based on
confidence in the cover values (Figure Z). This plot lttustrates two major trends in the classified data: (1) forest classification accuracy decreases with canopy cover and (2)
variance in-understory increases with decreasing canopy
cover and decreasing forest classification accuracy. One major exception to this is the NoBLEs-or site which had hieh
canopy cover, low understory variance, and low classification accuracy. This site was dominated by a dense, moderately aged, P. contorta forest intermingled with small
riparian shrubs. The health and general vigor of the stand
was poor, probably a result of the high density of trees. Low
canopy foliage coupled with dense, exposed slemwood pres u m a b l y r e s u l t e d i n a c o n f u s i n g s p e c t r a l s i g n a t u r e .T h e v a r i ance in the understory may increase on a simple basis due
gaps in the canopy which provide growing spacefor forest
lo
tloor_vegetation. On the whole, understory cover may vary
greatly as the size and spatial distribution of the gapi vary
across a sample area.-From a mapping standpoint, these gap
communities are smaller than the sensor resolution and mav
be ommitted in a field survey; however, their presencu on u
sr.rb-pixel basis may greatly confound the speciral response of
a torest type.
The species mapping accuracies were less than those for
_
iorest genera. Mapping accuracies were 55 percent for p. jefJreyi,44 percent for P. contorfa, 84 percent for A. concolor,
and 50 percent for A. magnifico (Table 4). The overall map
accuracy was 58 percent, commission error of a) noraont
and Kappa'ut.,"iio.le. Clearly,somespecies-;J';;;ii;.
definedby the stratification
p.oiessthanbthers,especially
'
A. concolor. The elevational gradient for the fir forestsis
fairly restricted as opposed to the broad elevational intergrading of the pine forests.In addition, the moisture tolerances of the fir speciesare better defined in the literature.
The lower accuracyof the A. magnifica class may be a function of the sparsercanopy and greaterunderstory prominence in A. magnifica forests.The misclassificationof the ,4.
nagnifica forestsmay, therefore,be due to classficationerror
in the spe^ctralclustering. The iow accuracyfor the pines is
mostly a function of the broad environmental limits of these
forests.T. mertensianocould not be delineated in the soecTABLE4.
AccuRAcy AssESSMENToF THE GIS-STRATIFtED
MAp.
Field Classes
GIS Classes
P. jeffreyi (Pj)
P. contorta (Pc)
A. concolor (Ac)
A. magnifica (Am)
Non-forest (NF)
Non-vegetation (NV)
PercentAccr.rracy
CommissionError
Pj
Pc
'167
170
35
260
43
94
26
33
23294t4
9607
55
45
44
56
Ac
Att't
NF
NV
13
11.
208
1,1,
41
39
20
1.22
1
L"I
4
4
95
3
2
0
4
0
5
66
B4
16
50
50
81
19
86
'14
Map Statistics
n - 1580
Percent Correct = 5B7n
Commission Error : 42,,/n
Confidence Interval (95u1,Confidence Level) = 0.56 to 0.60
K : 0.4769
c^ - 0.0009
PE&RS
PEER REVIEWED ARTICTE
tral classifications due to low tree densities and high snow
cover. However, T. mertensiana was identified preliminarily
in the GtS model by designating any non-forest vegetation
above 2560 m as ?. mertensiana. Non-forest vegetation comprises shrub, pasture, and areas of scattered trees' However,
it altit.,des > 2560 m shrubs and pastures do not occur and
only scattered trees are present. These groves of trees are primaiily T. mertensiana or rarely P. albicaulis. Accuracy values were not calculated except in the agreement with
personal
experience or field notes.
Although more advanced spectral classifications employing maximum-likelihood classifiers with signature analysis
are available, they are unlikely to improve the species classification unless they are able to differentiate species on spectral properties alone. The simple genus classifications
employed here were )70 percent accurate, whereas the species classifications were as low as 44 percent. Because GIS
modifications were used to refine the relatively higher accurate genera classes to less accurate species classes,the inacoccur in the cts process. More involved spectral
".,.uJi"t
analyses that only result in marginal increases in genera classifications cannot markedly improve the species accuracies,
This study demonstrates that conifer species in wildland
habitats can be accurately identified to the genus level solely
using Landsat data, but that identification of species using
only spectral data may be difficult. Species maPs can be produced using ancillary data in a GIS model to modify spectrallv based information.
The success of the cIS modification technique is related
to definition of species habitat, quality of ancillary data, and
software capabilities. Results of the Gls modified maps indicate that forest species identification is possible using a combination of soectral and ancillarv data. Use of terrain data to
augment the^mapping process ^ilo*t modification of landcover classes in cases where spectral data are insufficient
(Richards et al., tg8z; Likens et aI., tgaz). Accuracy may be
enhanced by (1) accurately defining the environmental habitat of each species, (2) improving the quality of ancillary
data input into the GIS, and (3) changing the method in
which data layers are integrated in the GIS software.
Ultimately, the accuracy is limited by the degree of environmental heterogeneity in an area and the ability to describe the habitats of the species. The accuracy for GIS
modified maps is based on the available habitat data reoorted in the literature on these species. Most of the data reported on topographically-induced zonation of these species
ire from the sierra Nevada (Vankat, 1982; Holland' 1986;
Rr-rndelet o/., 19BB).As some disparities exist between the
forest zonation in the Cascadesand the Sierra Nevada (Parker, 1991), it is difficult to extrapolate the optimum habitats
for the Lassen Volcanic National Park forest species from the
Sierra Nevada zonation models. Descriptions of vegetation
distribution at Lassen (Gillette et al., 1962: Heath, 1967;
Cooke, 1962; Griffin, 1967; Taylor, 1990; Parker, 1991) are
generally too limited in scope and emphasize the floristic
iather tlian the quantitative viewpoint required by this study'
At present, digital elevation models are the primary
source for topographic data at the 1:24,000 scale. Information
extracted from the nEvs such as slope, aspect, elevation, and
potential soil moisture are affected by errors associated with
ihese data (Walsh et o/., 1sB7) which can cause spurious results in the cIS modification process and reduced mapping
accuracv of forest species. These errors include horizontal
striping and missing data. Local errors such as missing.data
o. pitr may be corrected by interpolation with surrounding
PE&RS
'o"".''
.-
;rr
..]
\4
*l
\
(r
U
o
()
A
/\
r
4 ,1
t
\'
v-/
"]
\Iv^A
^ /
^^y^
..]
,.],,,',
,,,.'^
'.1
^-:-
\/^\/
,
r --r
PERCENT
*
^\/\,
\/
Canopv cover
,l\
CORRECT
(%) bY SAMPLE
Understory variance
covervarlFigure2. Analysisof canopyand understory
accuracy
gradient
sample
decreasing
of
a
ance across
for spectralclassification.
data. Striping affects entire datasets, making error amelioration especia[y difficult. In this study, we-used mutliple filters to verticaily interpolate across striped areas and to
smooth the resultant image. This may have reduced the overall accuracy of the topographic information; however, prod-ucts from our topographic hnalysis were affected significantly
bv' the presence of these errors'
Mapping could be improved by altering how ancillary
and speciral information are correlated in the GIS context'
Skidmore (198g) reported that class modification by environmental data is enhanced using Bayesian probabilities in an
expert system. This method assumes that species are distributed across a continuum of environmental gradients rather
than distinct, exhaustive habitats. Translation of this approach into popular GIS softwares will ultimately increase
ability.
mapping
'
iti"itty, usage of the image processing techniques and
GIS integration described in this study beyond the Lassen
Park eco"svstemshould be done with discretion. The methods
and resulis presented here represent our best.effort to use the
soectral and environmental data to describe the Lassen landsiape. A review of the literature on forest mapping using re-oi" ,"nto. sources demonstrates some site-specificity as to
which spectral wavelengths are most informative or which
classificition procedure-is most effective. Th-e variablity from
one site to another is great given the spectral response to heterogeneity at the sub- and super-pixel level. Equivalent mappine perfbrmance in other forest systems using the
^rr.o6"ir,."t
outlined in this study would be a function of
iimilarity in forest species compositiol' structure, age-,vi8or,
and other stand variibles, as well as the response of the forest species to topographic zonation as defined here'
Conclusions
Forest species could not be separated using-only spectral ^
data, wliereas genus level mapping is possibie at a level of
73 percent accuracy using an unsupervised classification of
Landsat rM data. Forest structure, especially canopy coverage
PEER.REVIEWED
and understory vegetation, seems to affect classification performance. This study also demonstrated that spectral based,
genus level maps can be modified with terrain data in a GIS
to produce forest species maps at an accuracy level of SB
percent. Maps may be improved by refining the informaiion
through a majority analysis to accentuate specific map characteristics over others, based on usage and need. cIS map accuracies are a function of confinement of forest soecies to
modeled habitats and reliable definition of these habitats,
data quality, and cls software capability.
Acknowledgments
The authors would like to thank the following people for assistance in field data collection and tecltnical advise during
course of this project: Lynne Oliver, Allan Cook, Scott Kroh,
and Sharon White. We would also like to thank the adninistrators and staff of Lassen Volcanic National Park, especially
AIan Dennison, chief of natural resoLrrcemanagement, and
Gilbert Blinn, superintendent, for their logisticil support during our field sampling in the park. This research was funded
in part by the Adkins Fellowship', the Texas Christian University Research Foundation Grant #5-23719, the Natural Resource Division of Lassen Volcanic National Park, and by the
U.S. Department of Energy contract #DE-ACO9-76SROO-819
with the University of Georgia. The authors would also like
to thank members of the Numerical Tenadvnamics Simulation Group, including Drs, Steve Running, Ramakrishna Nemani, and E. Ray Hunt, Mr. Lars Pierce, and Mr. Peter
Thornton at the University of Montana, for their comments
and suggestions during the compilation of this paper.
Refelences
B a i l e y , W . H . , " 1 9 6 3 .R e v e g e t a t i o n o f t h e 1 9 1 4 - 1 9 1 5 D e v a s t a t e d A r e a
of Lassen Volcanic National Pork, PhD Thesis, Oregon State
Univ., Corvallis, Oregon, 195 p.
Benson, A.S., and S.D. DeGloria, 1985. Interpretation of Landsat-4
Thematic Mapper and Multispectral Scanner Data for Forest
Surveys, Photogranmetric Engineering & Remote Sensrng, 51(9):
'128"t-1.289.
ARIICTE
Fleming, M.D., J.S.Borkebile, and R.M. Hoffer, 1975. ConputerAided Analvsis of LANDSAT-1 l..fSS Data: A Contparison of
Three Approoches, Including a "Modified Clustering" Appiooch.
LARS Information Note 072475, Laboratory for Applications of
R e m o t e S e n s i n g , P u r d u e L J n i v e r s i t y ,W e s t L a f a y e t t e , I n d i a n a .
Franklin, I., 198tj. Thematic Mapper Analvsis of Coniferous Forest
Structrrre and Cornposition, lnternationol lournol of Rentote
Sensirrg, 7 ('1.
0\1287 -1.J0 1.
Gillette, G.W., J.T. Howell, and H. Leschke, 1961. A Flora of Lassen
Volcanic National Park, California, The Wasntann lournal of Biology. 191l):l-ltt5.
G r e g o r y , S . V . , F - . JS
. wanson, W.A. McKee, and K.W. Cummins, 1991.
An Ecosystem Perspective of Riparian Zones, Bioscience,41(g):
5 4 0 - 5 51 .
Griffin, J.R., 1967. Soil Moisture and Vepetation Potterns rn rVorthern
CaliJornia Fole.sts, U.S. Forest .Servir;,rResearch paper PSW-46,
22 p.
Heath, f.P., 1967. Primary Conifer Succession,Lassen Volcanic National Park, Ecology, 4BlZ):271-27 S.
Holland, R.F., 1S86. Preliminarv Descfiptions of the Terrestilal Natural Contntunities of Califoritio, State of California Department of
Fish and Gamc, 156 p.
H u g h e s , J . S . ,D . L . E v a n s , P . Y . B u r n s , a n d J . M . H i l l , 1 9 8 6 , I d e n t i f i c a tion of Two Southern Pine Species in High-Resolution Aerial
MSS Data, Photogrammetric Engineering & Renrcte Sensing,
5 2 ( 8 ) : 1 1 75 - " l " l B O.
I v e r s o n , L . R . , R . L . G r a h a m , a n d E . A . C o o k , t 9 8 9 , A p'Landscctpe
plications of Satellite Remote Sensing to Forested Ecosystems,
Ecology, 312):13'l-142.
lensen, J.R., 1S86. Introductoty Digital Inage Processing, PrenticeH a l l , N e w J e r s e y ,3 7 9 p .
Kaufmann, M.R., C.B. Edminster, and C.A. Troendlc, 1982. Leaf Area
Deterntinotions for Subolpine Tree Spedes in the Centrol ilocky
Mountctins, Rest:arch Paper RM-238 Ror;kv Mountain Forest and
R a n g eE x p e r i m e n t S t a t i o n .
Khonam, S., and E.F. Katibah, 1984. Vegetation/Land Cover Mapping of the Middle Fork of the Feather River Watershed Using
Landsat Digital Data, Forest Science, 3o(1):z4B-ZSB.
Likens, W. K. Maw, and D. Sinnott, 1982. Final Report: Landsat
Land Cover Analysis Conpleted for CIRSS/San Bernadino
County Project, NASA Technical Mentorandun Bt24t,2S p.
Beven, K.J.,and E.F. Wood, 1993. Catchment Geomorphology and
the Dynamics of Runoff Contributing Areas, /ournal of Hydrology.65:139-158.
Mayer, K.E., and L. Fox III, 1981. Identification of Conifer Specres
Groupings frotn Landsat Digitol Classifications, Photogrontntetric
Engineering & Renote Sen.sing,4B("1"t):1607-1614.
Bolstad, P.V., and T.M. Lillesand, 1992. Improved Classification of
Forest Vegetation in Northern Wisconsin Through a Rule-Based
Combination of Soils, Terrain, and Landsat Thematic Maoper
D a t a , F o r e s l S c i e n c e , 3 8 [ 1) : 5 - 2 O .
Bormann, F.H., 1985. Air Pollution and Forests:An Ecosvstem Perspective, BioScience, 35(7):434-44'1.
Moore, M.M., and M.E. Bauer, 1990. Classification of Forest Vegetation in North-Central Minnesota psing Landsat Muttispectral
Scanner and Thematic Mapper Data, Fores/ Science,36(Z):330342.
Cibula, W. G., and M.O. Nyquist, 1982. Use of Topographic and Climatological Models in a Geographical Data Base to Improve
Landsat MSS Classification for Oiympic National Park, Photogrommetric Engineering & Rcmote Sensing, 53(1):67-75.
Conese,C., G. Maracchi, F. Miglietta, F. Maselli, and V.M. Sacco,
1988. Forest Classification by Principal Component Analyses of
TM Data, Internotional lournol of Remote Sen.sing, 9[10-11):
1.597-1.61.2.
Congalton, R.G., R.G. Oderwald, and R.A. Mead, 1983. Assessing
Landsat Classification Using Discrete Multivariate Analysis Statistical Techniques, Photogrammetric Engineering & Renote
Sen si n g, 49(12) :167 1-"1678.
C o o k e , W . B . , 1 9 6 2 . O n t h e F l o r a o f t h e C a s r ; a d eM o u n t a i n s , T f i e
Wasmann lournal of Biology, 20(1):1.-67.
E z r a , C . E . , L . R . T i n n e y , a n d R . D . J a c k s o n ,1 9 8 4 . E f f e c t o f S o i l B a c k ground on Vegetation Discrimination Using Landsat Data, Remote Sensing of the Environment, "t6:233-t42.
304
O'Loughlin, E.M., 1986. Prediction of Surface Saturation Zones in
Natural Catchments by Topographic Analysis, \4/oter Resources
R es e ar c h , 2 2 ( s ) : 79 4 - 8 o 4 .
Parker, A.f., 1991. Forest environment relationshios in Lassen Volcanic Naitonal Park, California. U.S.A., fournol of Biogeography,
1 8 (5 l : 5 4 3 - 5 5 2 .
Richards, ].A., D.A. Landgrebe, and P.H. Swain, 1982. A Means for
Utilizing Ancillary Information in Multispectral Classification,
Remote Sensing of the Environment, 72:463-477.
R u n d e l , P . W . , D . ] . P a r s o n s . a n d D . T . G o r d o n , 1 g B B .M o n t a n e a n d
S u b a l p i n e V e g e t a t i o n o f t h e S i e r r a N e v a d a a n d C a s c a d eR a n q e s ,
Terrestrial Vegetation of California (M.G. Barbour and J. Maior,
editors), California Native Plant Societv Special publication
Number g.
Shen, S.S., G.D.'Baclhwar,and f.G. Carnes, 1985. Separabilitv of Boreal Forest Species in the Lake JennetteArea, Minnesoti, photogrammetric Engi neering & Remote Sensrng, 51(11):17 7 S-.1783.
S k i d m o r e , A . K . , 1 9 8 9 . A n E x p e r t S y s t e m C l a s s i f i e sE u c a l y p t u s F o r e s t
Types Using Thematic Mapper Data and a Digital Terrain
PE&RS
PEER.REVIEWED
Model, Phofogrctntntetric Engineering, & Rentote Sensing, 55[10):
1.449-"1464.
S o a n n e r , M . A . , L . L . P i e r c e , D . L . P e t c ' r s o n ,a n d . S . W .R u n n i n g , 1 9 9 0 '
Remoto Sensing of Coniferous Ftlrest Leaf Area Index The Influence of Canopy Closure Understory Vegetation and Background
R e f l e c t a n c e ,i n t e r n a t i o n a l l o u r n a l o f R e n t o t e S e n s i n g . 1 1 ( 1 ) : 9 5 "1"11.
Stitt, S., 1990. ProcessingDEM data, Grassclrpprngs,4[1):1-8'
Tavlor, A.H., 1990. Tret' Invasion in Meadows of Lassen Volcanic
National Park, Calilbrnia, Professional Geographer' a2l4):457470.
Thomas, I.L., and C. McK. Allcock, 1984. Determining the Confidence Level for a Classificalion, Photograntfietric Eng'ineering &
R e n o t e S e n s i r r g ,5 0 ( 1 0 ) : 1 4 9 1 - 14 9 6 .
(/ser'^
U.S. Geologit;al Survey, 1987. Digitol Elevatiort Models Dotn
Guide, Department of the Interior, 38 p.
V a n c l e C a s t l e , 1 . R . ,1 9 s 1 . R e m o t e s e n s i n g a n d n l o d e l i n g a t ; t i v i t e s f o r
long-ternr ecological reseatch, ASPRS-GIS/LIS'91 Prot:eedings.
Atlanta, Georgia, PP. 544-550
Vankat, 1.L.,1982. A Gradient Perspt'ctive on the Vegetation of Sequoia National Park, California, Madro-no, 29(3):2oo-21'4.
W a l s h , S . J . ,1 9 8 0 . C o n i f e r o u s T r e e S p e c i e s M a p p i n g U s i n g L a n d s a t
Dala, Renote Sensing of the Environn7il1t, 9:'11-26'
W a l s h , S . J . ,D . R . L i g h t f o o t , D . R . B u t l e r , 1 9 8 7 . R e c o g n i t i o n a n d A s s e s s n e n t o f E r r o r i n G e o g r a p h i < :I n f o r m a t i o n S v s t e m s '
Photograntnetric Engineering & Renote Sensing, 53(10J:142314 3 0 .
(Received 18 Mav 1992; revised and accepted 20 April 1993: revised
15 June 1993)
|oseph White
inr"ptr White is a doctoralstudentand researchassistantin
ARIICTE
the School of Forestry at the University of Montana at Missoula, Montana. He received his Master's degree in Biology
from Texas Christian University in 1991. His research interests include vegetation mapping using expert systems based
on ecolosicai information. He is cunently conducting research in"ecological modeling using satellite, climate, and
topographic driven models and remote sensing of pre- ^and
port-"ti." burn as function of communitv structure and func-
Glenn Kroh
Glenn Kroh is an AssociateProfessorof Biology at Texas
ChristianUniversity in Fort Worth, Texas.He was awarded
his doctoraldegree'inPlant Ecologyin 1975 from Michigan
StateUniverist!. Researchhas includedplant-physiological
pathmodeling, plani-insect relationships,and suc^cession
wavs of ;a;tern deciduousand westernconiferousforests'
Hii current researchinterestsare vegetationmapping using
remote seningand comparisonof historicalremotely sensed
ciatawith obs"ervedbiomass accumulation in developing forests.
fohn Pinder
john Pinder is currently an AssociateResearchEcologist
with the SavannahRiver Ecology Laboratorythrough the
University of Georgia.He receivedhis doctoraldegreein Zoology through the Lniversity of Ggorgjl i,n 1977 and has pur,.,""d ."."u.Jh topics including old-field plant successionand
transportof radibactivematerialsthrough terrestrialecosystems.'Currentresearchincludesremotesensingof landscapes
to determinedistributionand patternsof plant compositio-n
and biomassin relation to geomorphicand anthropomorphic
influence.
HOW to LIE with MAPS
by: Mark Monmonier
Members,$]2' Stock# 50]4.
1991. 184pp, sofrcover.$15,.,4.'PR"'
us how to evaluatemapscritically
A lively, cleverlyillustratedessayon the useandabuseof mapsthatteaches
modelsof reality. This bookshowsthatmaps
andpromotesa healthyskepticismabouttheseeasy-to-manipulate
cannot only point the wayand provideinformation,mapslie. In fact, they haveto lieChaptersinclude:
o Elementsof the Map
I Map Generalization:
Little WhiteLies andLotsof
Them
o Blundersthat Mislead
o Mapsfor PoliticalProPaganda
o Maps,Defense,andDisinformation:Fool Thine
Enemy
of the Census
a DataMaps: Making Nonsense
bad
Fromthe peoplewho makemapsanddo not wantto misleadreadersor clients to thepotentialvictimsof andall the restof us
travelers,students,
managers,
planners,
business
voters,politicians,
maps: scientist,
everyonewill benefitfrom readingthis book'
For ordering information, see the ASPRS Store in the back of this journal'
PE&RS