Full text of DOI:10.1080/01431160117906

This article was downloaded by: [University of Reading]
On: 26 December 2014, At: 10:10
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number:
1072954 Registered office: Mortimer House, 37-41 Mortimer Street,
London W1T 3JH, UK
International Journal of
Remote Sensing
Publication details, including instructions for
authors and subscription information:
http://www.tandfonline.com/loi/tres20
Measurement of oak tree
density with Landsat TM
data for estimating biogenic
isoprene emissions in
Tennessee, USA
W. Baugh , L. Klinger , A. Guenther & C. Geron
Published online: 25 Nov 2010.
To cite this article: W. Baugh , L. Klinger , A. Guenther & C. Geron (2001)
Measurement of oak tree density with Landsat TM data for estimating biogenic
isoprene emissions in Tennessee, USA, International Journal of Remote Sensing,
22:14, 2793-2810
To link to this article: http://dx.doi.org/10.1080/01431160117906
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all
the information (the “Content”) contained in the publications on our
platform. However, Taylor & Francis, our agents, and our licensors
make no representations or warranties whatsoever as to the accuracy,
completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of
the authors, and are not the views of or endorsed by Taylor & Francis.
The accuracy of the Content should not be relied upon and should be
independently verified with primary sources of information. Taylor and
Francis shall not be liable for any losses, actions, claims, proceedings,
demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in
relation to or arising out of the use of the Content.

This article may be used for research, teaching, and private study
purposes. Any substantial or systematic reproduction, redistribution,
reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access
and use can be found at http://www.tandfonline.com/page/terms-and-
conditions

int. j. remote sensing, 2001, vol. 22, no. 14, 2793–2810
Measurement of oak tree density with Landsat TM data for
estimating biogenic isoprene emissions in Tennessee, USA
W. BAUGH , L. KLINGER , A. GUENTHER and C. GERON
2
1
Atmospheric Chemistry Division, National Center for Atmospheric Research,
1
1
1
P.O. Box 3000, Boulder, CO 80307, USA
2
US Environmental Protection Agency, OYce of Research and Development,
National Risk Management Research Laboratory, Research Triangle Park,
NC 27711, USA
(Received 20 September 1999, in Ž nal form 5 June 2000)
Abstract. Isoprene emissions from oak trees in the eastern USA play an import-
ant role in tropospheric ozone pollution. Oak trees (Quercus) emit an order of
magnitude more isoprene than most other emitting tree species, and are by far
the largest source of biogenic isoprene in the eastern US. We used Landsat TM
data to measure oak tree abundance near Oak Ridge, Tennessee, to estimate
 uxes of isoprene. The Landsat classiŽ cation was performed using multi-date
data, supervised classiŽ cation techniques, and an iterative approach. Training
sites were selected based on transect data, and ten vegetation classes were mapped.
A supervised classiŽ cation algorithm called the Spectral Angle Mapper was used
to classify the data. Empirical vegetation emission data were used to estimate the
isoprene  ux from each of the vegetation classes. The resultant isoprene  ux maps
were compared with concentrations measured in the Ž eld, and a good correspond-
ence was observed. We also compare the Landsat classiŽ cation with three other
landcover schemes including the USGS’s Global Landcover ClassiŽ cation, which
is based on AVHRR data. Results from these landcover classiŽ cations are used
as input for models that predict tropospheric ozone production and are used to
investigate ozone control strategies.
1. Introduction
One of the most persistent air pollution problems in the USA is tropospheric
ozone, a photochemical oxidant and an important component of smog. Elevated
ozone concentrations in urban and rural air were recognized in the early 1950s
(Haagen-Smit and Fox 1954); and massive and costly cleanup eVorts have been
implemented since the 1960s. Controlling ozone pollution is diYcult because ozone
is not emitted directly into the atmosphere. Instead, ozone is the product of precursor
gas emissions and atmospheric chemical reactions. Under certain meteorological
conditions, ambient ozone concentrations in and downwind of some urban areas
can reach levels higher than the concentration considered acceptable for public
health by the US Environmental Protection Agency (EPA) (National Research
Council 1991).
Ozone (O ) is a reactive oxidant gas that is produced naturally in the Earth’s
atmosphere. Most of the atmospheric ozone occurs in the stratosphere (Ž gure 1).
3
International Journal of Remote Sensing
ISSN 0143-1161 print/ISSN 1366-5901 online © 2001 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/01431160010006944

2794
W. Baugh et al.
Figure 1. Generalized atmospheric ozone proŽ le (after National Research Council 1991).
Stratospheric ozone plays a critical role in absorbing damaging ultraviolet solar
radiation. A smaller portion of naturally occurring ozone is found in the troposphere
(Ž gure 1). Tropospheric ozone accounts for about 10–15% of the atmospheric total.
Its concentration is measured in orders of 10 parts per billion (ppb), while strato-
spheric ozone concentrations can exceed 10 000 ppb (National Research Council
1991 ).
Ozone is formed by a series of atmospheric photochemical reactions. The
reactions require two types of gaseous compounds: nitrogen oxides (NO ) and
x
volatile organic compounds (VOCs). NO and VOC sources are both anthropogenic
(of human origin) and natural. Most NO is produced from atmospheric nitrogen
x
x
during high temperature combustion. Examples of NO sources are automobiles,
x
fossil-fuel power plants, lightning, forest Ž res, and natural soil processes. Anthro-
pogenic VOC sources include fossil fuel combustion products, and evaporation of
fuels, solvents, and paints. Natural (biogenic) VOCs are emitted primarily by forests.
The presence of naturally emitted biogenic VOCs increases the complexity of
ozone control. On regional scales, natural sources of these precursor components
are often greater than anthropogenic sources (Lamb et al. 1993, Guenther et al.
1995). Yet, until recently, ozone control strategies were based on modelling studies
that did not consider biogenic emissions.
Biogenic VOCs include many organic compounds, and the amount emitted varies
considerably among plant species (table 1). One of the compounds emitted by certain
deciduous trees is isoprene (C H ); conifers commonly emit monoterpenes such as
5 8
a- and b-pinene. These few compounds account for about 50% of the biogenic VOC
emissions; with the remainder, a wide variety of other compounds (Guenther et al.
1994). In the USA, forests are believed to emit about 90% of the biogenic VOCs,
with agricultural and scrub lands contributing the rest (Lamb et al. 1987).
In the south-eastern USA (where ozone pollution problems are particularly
persistent), most of the biogenic VOC emissions are from relatively few tree species.
Regional models that incorporate tree emission rates and total biomass of each

Measurement of oak tree density with L andsat T M data
2795
Table 1. Isoprene and terpene emission rates for the dominant tree genera in Oak Ridge,
TN, study area (from Guenther et al. 1994).
% Leaf
biomass
Isoprene emission Terpene emission
1
(mg C gÕ hÕ )
1
1
(mg C gÕ hÕ )
1
Genus
Common name
Quercus
Pinus
Acer
Oak
Pine
Maple
Elm
Sourwood
Tulip Poplar
Juniper
Hickory
Blackgum
Sweetgum
Beech
29.9
12.6
6.1
5.3
5.1
4.4
4.0
3.8
2.8
2.5
2.4
70
0.2
3.0
1.6
<
<
<
<
<
<
<
0.1
0.1
0.1
0.1
0.1
0.1
0.1
<
Ulmus
0.1
0.6
0.2
0.6
1.6
0.6
3.0
0.6
Oxydendrum
Liriodendron
Juniperus
Carya
Nyssa
Liquidambar
Fagus
14
70
<
0.1
species have shown that isoprene emitted by oak trees dominates the daytime
biogenic VOC  ux in the eastern USA (Geron et al. 1994, Guenther et al. 1995).
In response to a need for accurate distribution maps of oak trees, the use of
Landsat Thematic Mapper (TM) data for mapping oak tree densities in the Oak
Ridge, TN, area was investigated. Species level emissions of biogenic VOCs are well
known in this area, but distributions of oak trees are not well known. This classiŽ ca-
tion was compared with the US Geological Survey’s (USGS) Seasonal Land Cover
Regions (SLCR) landcover classiŽ cation, which is based on Advanced Very High
Resolution Radiometer (AVHRR) data; the US Forest Service’s (USFS) Forest
Inventory and Analysis (FIA) database; and the EPA’s Biogenic Emission Landcover
Database (BELD). Both satellite landcover maps were converted to isoprene emission
maps by assigning Ž eld-measured leaf-level emission rates to the landcover classes.
Field-measured isoprene concentrations were compared with the isoprene emissions
predicted by the remotely sensed maps. These emission maps are used as input to
regional ozone production models and therefore represent a critical link between
leaf-level measurements and regional isoprene  ux estimates.
2. Field study methods
2.1. Study area
This study was conducted in and around the Walker Branch Watershed
84 17 30 ), which is located within the Oak Ridge National
° ê
(35 57 30 ,
°
ê
-
Laboratory (ORNL), a US Department of Energy reservation near Oak Ridge, TN
(Ž gure 2). The Oak Ridge area (and the south-eastern USA) commonly exceeds EPA
ozone standards during the summer months. Field studies were conducted in the
summers of 1992 and 1995; the latter was timed to coincide with the 1995 Southern
Oxidant Study (SOS) Ž eld campaign. During this campaign other institutions
conducted experiments ranging from soil sampling to aircraft air sampling.
The Ž eld area lies on the western edge of the Appalachian Mountains, and is
characterized by forest-covered, northeast trending ridges and valleys with a relief
of about 125 m. The forests are primarily hardwood with some conifer stands. The
dominant tree genera are: oak (Quercus), maple (Acer), hickory (Carya), pine (Pinus),
sweetgum (L iquidambar), sourwood (Oxydendrum), and blackgum (Nyssa) (table 1).
A 44 m walkup tower provided a platform for leaf and branch enclosure experiments,

2796
W. Baugh et al.
Figure 2. Index map of Oak Ridge, TN. Land classes are based on Landsat TM data from
this study. ‘Oak forest’ includes both high and medium oak classes; ‘pine forest’
includes all conifer classes. The circles show the fetch areas for various sampling
methods and are described in §4.3.
and for above-canopy atmospheric measurements. A tethered balloon atmospheric
sampling system was deployed from a clearing about 400 m north of the tower.
Further information about the site can be found in Johnson and Van Hook (1989).
2.2. Ground transects
In 1992 three transects with a total of thirty-two 10 m 10m (100 m ) individual
×
plots were sampled. Each transect was approximately 1 km long. Transects began at
2
benchmarks with known locations and were measured with tape and compass. In

Measurement of oak tree density with L andsat T M data
2797
1995 eight 200 m long belt transects were sampled radially outward from the 44 m
tower in the following directions: N, NE, E, SE, S, SW, W, and NW (Ž gure 3). Each
belt transect consists of adjacent 10 m 10 m plots. In all transects each plot was
×
sampled for: slope, aspect, species composition, diameter at breast height (dbh)
[
1.5 m], tree height, sapling count, and percentage cover of the dominant growth
~
forms. Trees are deŽ ned as individuals greater than 4 cm dbh; saplings are individuals
less than 4 cm dbh and taller than 1.5 m.
Transect data were processed to estimate the crown-area percentage of each tree
species in the canopy (Guenther et al. 1996a). The plots were classiŽ ed using these
percentages and the same classiŽ cation scheme that was applied to the TM data
(§3.2). A subset of the classiŽ ed TM image surrounding the tower and the classiŽ ed
transect map are compared in Ž gure 3.
Figure 3. Comparison of tower transects and Landsat classiŽ cation. Larger background
pixels are from the Landsat classiŽ cation; smaller squares are tower transect plots.
The transect plots are shown as 5 m squares for clarity; the actual plots are 10 m
squares. See discussion in §2.2 and §4.1.

2798
W. Baugh et al.
2.3. Field observations
With only one day available to work outside the training area, reconnaissance-
style observations were made to help assess the accuracy of the TM classiŽ cation.
Walking transects were conducted at two sites in the ORNL, and at two sites on
the Cumberland Plateau, about 20 km to the north-east of the tower site. Walking
transects consist of careful map reading; locating plots about every 30 m and record-
ing tree species, diameter, and canopy cover; and comparison with the TM image.
2.4. Gas  ux measurements
Biogenic isoprene  uxes were measured at three scales in the Walker Branch
Watershed (Guenther et al. 1996a). At the leaf-level scale, leaf and branch emissions
were studied with enclosure systems. Clear Te on bags were used for branch enclosure
measurements, while leaf cuvettes were used to sample foliar emissions in a controlled
environment. Emissions are estimated with a mass balance approach, in which
constituents of the air  owing to and from the sample chamber are compared. On
the canopy level, a Relaxed Eddy Accumulation (REA) system and a micrometeorol-
ogical vertical gradient system were used to estimate  uxes in the atmospheric surface
layer above the forest (Guenther et al. 1996a). These systems were mounted on the
44 m tower and use ambient isoprene concentrations to estimate the  ux. At the
landscape level,  uxes were estimated with data collected from a tethered balloon
proŽ ling system at elevations up to 2 km above ground level. These sampling and
 ux estimation methods are described in more detail in Guenther et al. (1996a, b),
and the results are compared with TM-based  ux maps in §4.3.
3. Image processing methods
3.1. General
The desired product of the image classiŽ cation was a landscape characterization
showing the relative density of oak trees in the forest. The classiŽ cation was performed
using multi-date Landsat TM data, supervised classiŽ cation techniques, and a multi-
layered (iterative) approach. An existing 1 km landcover classiŽ cation by the USGS
was also evaluated as a lower cost alternative to Landsat TM data. The two
classiŽ cations were co-registered and compared for correspondence of similar classes.
Both classiŽ cations were also converted to isoprene emission maps using empirical
Ž eld measurements. Isoprene  ux estimates based on the remotely sensed data were
compared with ambient isoprene concentrations measured during the 1992 Ž eld
campaign. Image processing was performed with the ‘Environment for Visualizing
Images’ (ENVI) [RSI, Boulder] software on a Power Macintosh 8100/80 computer.
Additional data manipulation and iterative classiŽ cation were done with programs
written in the ‘Interactive Data Language’ (IDL) [RSI, Boulder].
3.2. L andsat T M processing
The Oak Ridge data set consists of two standard seven-band digital Landsat TM
images of path 019, row 035 that were acquired on 17 December 1991 (winter)
and 30 September 1992 (early fall). From these images 10 land cover classes were
identiŽ ed: high oak/other deciduous (greater than 50% oak in deciduous forest);
medium oak/other deciduous (25–50% oak); low oak/other deciduous (less than
25% oak); conifer; mixed conifer/deciduous; conifer plantation; shrubs, grasses,
and agriculture (non-forest vegetation); water; bare soil; and urban/concrete.
ClassiŽ cation parameters for the Landsat TM data are summarized in table 2.

Measurement of oak tree density with L andsat T M data
2799
Table 2. Landsat TM image classiŽ cation parameters.
Number of
training
pixels
ClassiŽ cation
priority
Class
Subclass
Method of plot location
Season
Conifers
Loblolly
Mixed
Plantation
High oak
Medium oak
Low oak
195
129
151
11
10
9
130
Field observations
Field observations
Aerial photographs
Transect plots
Transect plots
Transect plots
Field observations
Highest
Highest
Highest
2nd
2nd
2nd
Winter
Winter
Winter
Autumn
Autumn
Autumn
Autumn
Deciduous
/
Shrub grass
Nonforest
vegetation
3rd
Agriculture
149
196
259
241
Field observations
Field observations
TM image
3rd
3rd
Lowest
Lowest
Autumn
Autumn
Winter
Winter
Soil
Urban
Water
TM image
The following processing was applied to both TM images before classiŽ cation.
TM band six (thermal infrared) was removed from the data. A regression adjustment
(dark object subtraction) was used to correct for atmospheric and sensor eVects
(Jensen 1986, Chavez 1988). Subsets (1024 pixels by 1024 pixels) covering the ORNL
and Oak Ridge city areas were extracted from the larger image to accelerate iterative
processing on a slow computer. The subsets were registered to the Universal
Transverse Mercator (UTM) coordinate system (zone 16) with a pixel size of 30 m
using nearest neighbour resampling and second order polynomial equations. A root
mean squared (rms) error of 0.49 pixel was achieved for the Ž t between each image
and the 1:24 000 scale USGS topographic maps using 24 evenly distributed ground
control points.
Field measurements made in 1992 were used to choose training sites for each
class. The Ž eld data consist of detailed vegetation inventories in transects with well-
located 10 m 10 m plots. Transect plot locations were measured by tape and com-
×
pass, and linked to ORNL benchmarks. The transects, unfortunately, were not
measured for the purpose of Landsat TM training. The 10 m square plots are
signiŽ cantly smaller than the 30 m TM pixels, and we must assume that the transect
plots represent the vegetation in the corresponding TM pixel. The risks due to this
assumption are described in §4.1.
Averaged reference spectra were extracted from the training site pixels. The
reference spectra were used as input for the Spectral Angle Mapper (SAM) classiŽ ca-
tion algorithm, and the results were combined manually with an iterative analysis.
As pixels were classiŽ ed in the original Landsat data, the corresponding pixels on a
‘blank slate’ image were assigned a value indicating the class (e.g. pine 10). Once
=
a pixel was assigned to a class, it was considered permanent and not re-assigned
(Wolter et al. 1995 ).
The SAM is a multi-spectral supervised classiŽ cation algorithm developed at the
University of Colorado (Boardman unpublished data, Kruse et al. 1993). This method
treats image spectra as n-dimensional vectors, where n is the number of bands. A
test spectrum (from a pixel) is compared with a reference spectrum (from the desired

2800
W. Baugh et al.
class); a smaller angle between the spectral vectors indicates a better match. A
threshold angle (chosen by inspection) is used to determine whether a spectrum is
classiŽ ed as a speciŽ c class. The power of the SAM method is that it is very sensitive
to small spectral variations, yet it is resistant to illumination diVerences.
We began the classiŽ cation process by mapping the coniferous trees. These
evergreen trees are distinct in the winter image because the surrounding deciduous
trees have lost their leaves. Individual stands of loblolly pine and mixed deciduous/
pine forest were identiŽ ed during the 1992 Ž eld study and indicated on aerial
photographs. The corresponding areas were located on the winter Landsat image,
and training spectra were extracted and averaged (table 2). The averaged spectra
were used by the SAM to classify the TM winter subset.
Inspection of the results showed that some obvious evergreen stands were not
classiŽ ed as loblolly or mixed pine. Interpretation of stereo aerial photographs
showed that young conifer trees planted in regular rows densely cover these areas.
These areas are assumed to be cultivated conifer trees (plantation). Training pixels
were selected from these unclassiŽ ed evergreen areas, and the winter image was
reclassiŽ ed with the three conifer classes.
Because the conifer trees are easily detected with the winter image, the three
conifer classes were added to the Ž nal result image with the highest priority. These
classes would not be changed during further manipulations. The SAM output consists
of three bands (images) of  oating point data, one for each conifer class. The value
of each pixel is the spectral angle, in radians, between the class’ training spectrum
and the pixel’s winter TM spectrum. Pixels with lower values have a better match.
A program was written to select the band with the lowest value for each pixel. If
this value was below an arbitrary threshold of 0.1 radians then that pixel was
assigned to the corresponding class. The ENVI software can be used to automatically
evaluate the SAM results and return a Ž nal classiŽ ed image. In this case, however,
the evaluation was done manually so that the individual classes could be added one
layer at a time, with weighted priorities, to the Ž nal classiŽ cation image.
The second iteration of classiŽ cation focused on the deciduous and oak classes,
and used the early fall TM image. Oak trees commonly lose their leaves (senesce) 1
to 2 weeks later than other deciduous trees. Using this brief period of time when
oak trees are the only deciduous trees with signiŽ cant remaining leaves has been
quite eVective in mapping oak trees in the eastern USA (Baugh unpublished data,
Wolter, et al. 1995). The early fall image used in this study, however, is about 1
month too early to utilize this eVect and the oak mapping relies only on apparent
spectral diVerences.
Training spectra for high, medium, and low oak classes were derived from pixels
that correspond with speciŽ c ground transect plots from the 1992 survey. Detailed
vegetation characteristics were measured at 32 plots in three transects. The plots
were classiŽ ed as high, medium, or low-oak deciduous based on the vegetation
surveys. All plots with more than 30% pine were excluded (mixed pine class already
classiŽ ed). Remaining forest plots with less than 25% oak were considered low oak
deciduous; 25–50% oak were medium oak deciduous; and greater than 50% oak
were high oak deciduous.
Plot locations were mapped with tape and compass, and marked on stereo aerial
photos. The points were carefully transferred to 7.5 minute topographic maps and
the UTM coordinates measured. Based on the UTM coordinates, the corresponding
TM pixels were identiŽ ed, and average TM spectra were extracted (table 2). The

Measurement of oak tree density with L andsat T M data
2801
image was classiŽ ed with the SAM, and the results were added to the Ž nal classiŽ ca-
tion image with the same method that was applied to the pine classes. The relatively
small number of pixels available for deciduous class training is of some concern, and
we had to assume that the detailed transects provided representative training spectra.
Areas of shrubs, grasses, agriculture, and bare soil were identiŽ ed based on aerial
photographs that were annotated during the 1992 Ž eld work. The early fall image
showed the best diVerentiation for these classes. ClassiŽ cation was done with the
method described above. The SAM results were added to the Ž nal classiŽ cation
image with a lower priority than the coniferous or deciduous classes.
Finally, average spectra for the remaining classes (urban and water) were derived
from regions that were identiŽ ed by inspection of the Landsat images, and classiŽ ca-
tion was done using the winter image. These results were added (with the lowest
priority) to the Ž nal classiŽ cation image. To aid interpretation, the classiŽ cation
image was colour coded and roads and highways were added from USGS 1:100 000-
scale digital line graphs. When we were comfortable with the results of the classiŽ ca-
tion for the subset image we repeated the appropriate steps to classify the complete
90 km 90 km Landsat TM quad (Ž gure 4).
×
3.3. Seasonal L and Cover Regions (SL CR) processing
The USGS’s Seasonal Land Cover Regions (SLCR) satellite classiŽ cation, a part
of the global land cover characterization data base, was compared with the Landsat
TM classiŽ cation results and evaluated as a low cost alternative to classifying
Landsat data. The SLCR classiŽ cation is a vegetation classiŽ cation with 1 km
pixels that covers most of the world’s continents. It is based on seasonal vegetation
changes in 10-day averages of AVHRR Normalized DiVerence Vegetation Index
(NDVI) data spanning April 1992–March 1993. The classiŽ cation contains 205
vegetation and land cover classes (varies regionally) and is available globally.
These data are distributed on the Internet at the EROS Data Center website:
http://edcwww.cr.usgs.gov/landdaac/glcc/glcc.html.
According to the USGS web site, SLCR classiŽ cation units are not based on
precisely deŽ ned mapping units in a predeŽ ned land cover classiŽ cation scheme.
Instead the classes serve as summary units for both descriptive and quantitative
attributes, and delineate areas of similar characteristics. The regions (classes)
are composed of relatively homogeneous land cover associations (e.g. similar  or-
istic and physiognomic characteristics) which exhibit distinctive phenology (e.g.
onset, peak, and seasonal duration of greenness), and have similar levels of pri-
mary production. Further information can be found at the USGS website:
http://edcwww.cr.usgs.gov/landdaac/glcc/globdoc1_2.html.
For biogenic emission land cover mapping, the SLCR database could provide a
useful data source if it is suitably accurate, or if it can be locally calibrated. Using
the TM classiŽ cation of the Oak Ridge area, we evaluated the usefulness of the
SLCR database for predicting biogenic emissions. To make the comparison between
the Landsat TM classiŽ cation and the SLCR database, the following steps were
taken. A subset of the North American SLCR image, slightly larger than the Oak
Ridge Landsat quad, was extracted. The subset was reprojected from the original
Lambert Azimuthal Equal Area projection to UTM coordinates using the original
projection units and equations from Snyder (1987). The 1 km SLCR pixels were
resampled to 30 m pixels using nearest neighbour resampling. A subset was extracted

2802
W. Baugh et al.
Figure 4. Landsat TM classiŽ cation results (upper) and the corresponding AVHRR SLCR
image (lower). Oak Ridge and the study area are located approximately in the center
of the images. The high oak area in the upper part of the Landsat image is the
Cumberland Plateau, and the pine and high oak area in the lower right corner is the
edge of the Great Smoky Mountains.
from the resampled SLCR image that had the same dimensions and map locations
as the TM quad classiŽ cation image. The SLCR and TM images were compared
pixel by pixel to look for relationships between the classes (§4.2).

Measurement of oak tree density with L andsat T M data
3.4. Isoprene emission maps
2803
Isoprene emission maps were created from the TM and SLCR classiŽ cations.
Each TM class was converted to isoprene  ux using empirical emission factors that
are based on vegetation enclosure measurements and a simple canopy isoprene
emission model, table 1, (Guenther et al. 1996a). SLCR classes were converted to
isoprene  ux using emission factors recommended by Guenther et al. (1994). The
results are images where each pixel value consists of a biogenic emission capacity.
The emission predictions can be resampled to other scales and used as input for
ozone production models. A comparison between the two remotely sensed isoprene
 ux estimates and Ž eld isoprene  ux measurements is presented in §4.2.
4. Results and discussion
4.1. Error analysis
This study lacks a direct quantitative error analysis of the image classiŽ cation.
VeriŽ cation consists of three somewhat indirect methods that are not individually
conclusive, but together suggest good agreement. The three methods are: comparisons
with two existing land cover databases—the USFS’s Forest Inventory and Analysis
(FIA) database and the EPA’s Biogenic Emission Landcover Database (BELD);
comparison with tower transects; and qualitative Ž eld observations.
4.1.1. FIA and BEL D comparisons
FIA data used in this study were collected in 1988 and 1989. Results of the
comparisons are presented in table 3. FIA data are collected by the USFS on
approximately ten-year intervals to monitor forest conditions. The process consists
of air photo interpretation and Ž eld surveys (Hansen et al. 1992). FIA Ž eld plots
Table 3. Comparison of the Landsat TM classiŽ cation (described in this paper), the Forest
Inventory and Analysis (FIA) database (Hansen et al.1992), and the Biogenic
Emissions Landcover Database (BELD), (Kinnee et al. 1997).
FIA
(% of crown area)
Landsat
(% of forest classes)
Land use types
Oak trees
Non-oak trees
Conifer
Other evergreen
Total area (ha)
37.7
47.9
14.4
0.02
NA
30.0 (18.6–41.3)
46.7 (35.4–58.1)
23.3
NA
653 672
Land use type
BELD
(% coverage)
Landsat
(% of all classes)
Total forest
Oak forest
Non-forest vegetation
Urban
60.6
21.0
14.5
5.4
70.2
15.4 (9.7–21.1)
14.9
6.1
Barren
Water
NA
1.3
0.8
3.1
Other
Total area (ha)
18.2
997 266
4.9
997 265
Note that for the FIA comparison only the forested pixels of the Landsat image are
considered, and in the BELD comparison the oak forest category is a subset of the forest
category.

2804
W. Baugh et al.
are sampled primarily by prism cruising, with Ž xed radius subplots to characterize
seedling/sapling populations. Recorded observations include tree diameter at breast
height, species, and other measurements that enable the prediction of the trees’
volume.
FIA inventories are extensive inventories that provide reliable estimates for large
sampling areas. As data are subdivided into smaller areas, such as a county, the
sampling errors increase and the reliability decreases. According to Hansen et al.
(1992), the eastern FIA inventories are designed to have a maximum 3% error per
1 million acres of timberland at the 67% conŽ dence limit (one standard error) for
basal area and volume. The Landsat image (quad) covers about 2.4 million acres;
therefore, the corresponding FIA data should have a 1.9% maximum sampling error
at the 67% level (standard error of the mean).
About 240 FIA plots are evenly distributed throughout the Landsat image, and
it was originally hoped that these plots could be used to create a contingency table
of error for the TM classiŽ cation. But, because the plots’ map coordinates describe
generalized polygon centres rather than speciŽ c transect locations and because of
the error discussion above, the FIA data were summarized for the entire Landsat
image area. Tree basal area values from the FIA plot data were converted to crown
area using equations from Geron et al. (1994). Crown areas were summed for four
categories: oak trees, all other deciduous trees, all conifer trees, and other evergreen
trees, such as Magnolia and Ilex. Values are reported as percentages of total crown
area for the four categories.
Because the FIA data pertain only to forested areas, only the forest categories
from the TM classiŽ cation were compared with the FIA data. The TM forest
categories are: conifer; conifer plantation; mixed conifer/deciduous; and low, medium,
and high oak deciduous. From these categories the areas covered by oak, other
deciduous, and conifer trees were estimated. The parameters used to create the TM
classes (§3.2) resulted in the following equations for the average area coverages
(maximum and minimum values were also calculated for ‘oak’ and ‘other deciduous’
and are presented in table 3):
Oak (LO 0.125) (MO 0.375) (HO 0.75) 0.35(0.332 MF)
=
(1)
×
+
×
+
×
+
×
Other Deciduous (LO 0.875) (MO 0.625) (HO 0.25) 0.35(0.668 MF)
=
×
+
×
+
×
+
×
(2)
(3)
Conifer Conifer Conifer Plantation (0.65 MF)
=
+
+
×
LO are low oak pixels, MO are medium oak pixels, HO are high oak pixels, MF
are mixed conifer/deciduous forest pixels, and Conifer and Conifer Plantation are
the respective number of pixels. In the mixed forest component of the ‘oak’ and
‘other deciduous’ equations, 0.35 is the fraction of deciduous in the mixed forest.
Within this 35% fraction, 0.332 is the factor that describes the oak component and
0.668 is the factor that describes the other deciduous component. In the mixed forest
component of the ‘conifer’ equation, 0.65 is the fraction of conifer in the mixed forest.
The TM conifer value was held constant to simplify the table.
Since the FIA data are limited to forested areas only, BELD data were used to
evaluate the remaining categories. This database is a compilation of several data
sources that were assembled by Kinnee et al. (1997). Non-forest categories were
taken from the USGS Land Cover Characteristics database and the US Census
Bureau’s TIGER database; forest categories were taken from resampled FIA data.

Measurement of oak tree density with L andsat T M data
2805
All are generalized to the county level. The BELD comparison used here (table 3)
was taken directly from Kinnee et al. (1997). Values are reported as percentages of
total area for consistency with the FIA comparison. Note that the larger total area
for the Landsat image in the BELD comparison (relative to the FIA comparison) is
due to all classes being considered—not just the forest classes.
Although the results of these comparisons are not a substitute for a contingency
table based on numerous Ž eld plots, they do provide feedback about the overall
accuracy of the Landsat classiŽ cation in the absence of such data. The comparisons
between the TM classiŽ cation and the FIA and BELD data are quite good (table 3).
In the FIA comparison, the oaks and other deciduous categories match very well.
Conifers appear to be over-predicted by the TM data. This could be due to other
evergreen trees and shrubs, such as Magnolia, Ilex, or Rhododendron, being misclassi-
Ž ed as conifer trees. The low occurrence of non-conifer evergreen trees in the FIA
data (0.02%) suggests that evergreen trees are not a signiŽ cant component of the
TM conifer category. But, Rhododendron shrubs, which are not included in the FIA
database, could be contributing to this over-classiŽ cation.
In the BELD comparison, the total forest, oak forest, non-forest vegetation, and
urban comparisons are very good. The TM classiŽ cation shows signiŽ cantly more
water. This is probably due to the small TM pixel identifying water bodies that the
coarser BELD classiŽ cation misses. There is no ‘barren’ category in the BELD, and
the ‘other’ category includes the unclassiŽ ed TM pixels.
4.1.2. T ower transects
The next method of veriŽ cation, a qualitative comparison between the TM
classiŽ cation and the 1995 radial tower transects, shows a general match of forest
patterns (Ž gure 3). Note that the high oak area surrounding the tower is well
characterized, as is the mixed deciduous area north of the tower. Other correspond-
ences can be found by inspection. Note that the transects were measured to character-
ize the forest near the tower, and not as a ground-truth check. The sampling strategy
was therefore not optimized for ground-truth purposes. The diVerence in scales
between the 900 m satellite pixels and 100 m transect plots hinders the comparison.
2
2
A feature that dominates a transect plot may be an insigniŽ cant part of a larger
satellite pixel. Also, the number of transect plots available (46) is too small for a
statistically valid comparison (Jensen 1986).
4.1.3. Field observations
The third method of veriŽ cation, qualitative Ž eld observations, is also not suitable
for statistical error analysis. It does, however, provide excellent feedback about the
accuracy of the classiŽ cation when it is combined with knowledge of the Ž eld area.
For example, it was known from experience that most oak trees grow on ridge tops
because they thrive on well-drained soils. The classiŽ cation showed high concentra-
tions of oak trees on the ridge tops, and walking transects in the Ž eld conŽ rmed this
observation.
Two sites in the Copper Ridge area of the ORNL, comprising high to medium
and high to low oak classes, were visited (Ž gure 2). At the Ž rst site, vegetation was
observed along a 200 m walking transect that crossed from medium to high oak
pixels. Observed vegetation corresponded well with the classiŽ ed image. In the
medium oak zone, oak trees were about 30–50 cm dbh and about 15 m tall. Hickory
(Carya) and maple (Acer) trees were intermixed, and the canopy was relatively dense.
About 75 m into the transect, the density and size of oaks notably increased. There

2806
W. Baugh et al.
were fewer hickory and maple trees, the canopy became nearly closed, and the oaks
were larger with a dbh of 50–90 cm and a height of about 25 m.
The second walking transect was a circle about 250 m in diameter and included
low, medium, and high oak pixels. Based on observations in this transect, the low-
oak pixels were under-classiŽ ed. Oak trees covered about 35–40% of the area in the
low-oak zone (a medium concentration). The trees were tall and spindly, however,
and likely exposed much of the highly re ective limestone bedrock to the satellite.
The bedrock exposure may account for the under-classiŽ cation.
The high and medium oak pixels on the second walking transect were classiŽ ed
more accurately than on the Ž rst. The canopy was mostly closed, and the bedrock
was covered by soil. Tree characteristics were similar to those described for the Ž rst
walking transect.
At the two Cumberland Plateau sites, trees were identiŽ ed with binoculars from
an automobile. The Cumberland Plateau is of particular interest because the classi-
Ž cation shows consistent, widespread high oak densities at the higher elevations
(NNW portion of Ž gure 4). The Ž rst site was about 11 km north of Oak Ridge at
the top of Walden Ridge. About 80% hardwood and 20% conifer cover the densely
forested ridge. Oak species comprise about 40–60% of the total canopy, with the
remaining hardwoods being maple (Acer), tulip poplar (L iriodendron), walnut
(Juglans), and hickory (Carya). The pine trees are concentrated in bands and patches
along the ridge. This is consistent with the classiŽ cation image, although the oaks
may be slightly over-classiŽ ed.
The second site is characterized by rural residential use, and the forest has been
heavily disturbed by human activity. Several hundred meters separate homes; and
yards contain lawns, gardens, non-oak hardwood trees, and pine trees. High concen-
trations of oak trees were observed, however, on hills behind the residential zone.
The classiŽ cation shows scattered high- to low-oak deciduous, and conifer pixels for
this area, consistent with the heterogeneous rural land use patterns.
The Ž nal site overlooks a ridge to the southwest of State Route 116. The ridge
is forested with a high density of oak trees and correctly classiŽ ed as homogeneous
high oak.
4.2. L andsat—SL CR comparison
We found that in the Oak Ridge area there is a poor relationship between the
SLCR class names and the corresponding TM classes. The SLCR classes also did a
poor job of distinguishing the land cover types (oak trees) that are important for
characterizing biogenic emissions. These results are to be expected, however, when
using 1 km AVHRR data to try to identify speciŽ c tree species in a mixed forest.
The high-oak tree stands are too small to be distinguished and therefore are lumped
into generalized deciduous and mixed deciduous SLCR classes. Despite these diVer-
ences, SLCR data were able to model isoprene emissions to within a factor of 3 of
measured values (§4.3). This is only a fair result for a well-studied site like Oak
Ridge, but it would be a good result for a more remote site with few previous studies.
To compare the TM and SLCR images, the fraction of each TM class that occurs
in each SLCR class was calculated as a percentage (i.e. SLCR classes were described
in terms of TM categories). The results are presented in table 4. The columns of
this table are SLCR classes. The rows are the composition of each class in terms of
TM (classiŽ cation) pixels. The numbers indicate the percentage of the TM class that
occurs in the SLCR class.

Measurement of oak tree density with L andsat T M data
2807

2808
W. Baugh et al.
Inspection of the table shows that there is very little correlation between the two
classiŽ cation methods. General patterns do emerge, however. The crop and grassland
SLCR classes primarily correspond with urban and suburban areas. The mixed
deciduous forest (SLCR 104) class is well correlated with the TM high oak areas.
There are relatively few SLCR 104 pixels in the image, however, and the bulk of the
high oak areas are accounted for by SLCR class 105 (mixed deciduous). SLCR 105
also accounts for most of the medium- and low-oak areas as well as much of the
mixed deciduous/conifer areas. The general SLCR mixed forest classes do correspond
well with the TM mixed forest classes.
4.3. Emission modelling
Biogenic isoprene  ux was modelled using oak tree (and other vegetation) distri-
butions from the TM and SLCR classiŽ cations. The TM results match Ž eld-measured
isoprene  uxes to within a factor of 2. We consider this to be reasonably good
agreement since a factor of 2 random diVerence due to heterogeneity of sources
within this region can be expected (Guenther et al. 1996c). The  ux modelling
technique is described in Guenther et al. (1996a), and uses estimated foliar densities
and area-averaged estimates of leaf-level parameters. In this case foliar density was
not used in the  ux estimate. It could be incorporated to a limited degree, however,
using a derivative of NDVI from the remotely sensed data sets to estimate relative
foliar density.
Fluxes were calculated for circular fetch areas with radii of 0.5, 3, and 14 km
from the tower/balloon sites. The 0.5 and 3 km footprints correspond with diVerent
methods of tower measurements, and the 14 km footprint corresponds with balloon
measurements (Guenther et al. 1996a). The footprint is the upwind ground area
responsible for emissions received at the sampling site. The estimates presented here
could be further reŽ ned by using the wind direction recorded at the time of sampling
and modelling the footprint as a wedge-shaped area upwind from the sampling site
(rather than a complete circle around the sampling site). Table 5 compares the
modelled  uxes with the measured  uxes. The agreement is surprisingly good and
supports the validity of estimating biogenic VOC  uxes with remotely sensed data.
The SLCR-based  ux estimates are about two to three times greater than the
measured  uxes. This is still a fair agreement, and it is encouraging for the future
use of SLCR data in isoprene emission studies. Also, the correspondence between
the TM and SLCR  ux estimates begins to converge over larger areas (TM quad
scale). It is likely that over larger regions the diVerences between TM and SLCR
classiŽ cations will diminish.
Table 5. Comparison between measured and modelled isoprene  ux (after Guenther et al.
1996a).
Measured  ux
TM modelled  ux SLCR modelled  ux
2
(mg C mÕ hÕ )
1
2
(mg C mÕ hÕ )
1
2
(mg C mÕ hÕ )
1
Footprint radius
±
0.5 km
3 km
14 km
Landsat ‘Quad’
4.8 0.6
±
5.0 1.7
4.6
3.4
2.5
6.1
10.5
10.4
9.7
±
2.4 0.3
±
4.8 0.6
8.6

Measurement of oak tree density with L andsat T M data
4.4. Forest classiŽ cation and ozone modelling
2809
The overall utility of this method, of estimating isoprene  uxes from satellite
vegetation maps, is strongly supported by the agreement between the estimated
isoprene  uxes and the measured  uxes, even though the accuracy of this TM
classiŽ cation is not validated on a pixel-per-pixel basis. This evaluation has signiŽ cant
uncertainties due to the limited available transect data. Future eVorts should use
transects designed speciŽ cally for the training of the intended remotely sensed data
set.
Regional ozone models use landcover data at a much coarser resolution than
that of Landsat TM. They use grid cells that range from hundreds to thousands of
square kilometres. The purpose of using higher resolution satellite data is to maximize
the accuracy of the oak tree density mapping. With TM data, the pixel size
approaches the crown size of large oak trees or small tree clusters. With fairly pure
pixels (unmixed), the spectral mapping is able to separate oak trees from other
deciduous trees. The 1 km pixels of the SLCR database are too large to eVectively
separate oak trees from the deciduous forest. Yet coarser data, such as SLCR or the
anticipated MODIS, are much more convenient for mapping at the global scales of
ozone models. A balance must be found between the accuracy of small pixel data
and larger pixel coverage. This balance will probably take the form of locally
calibrating coarse data with higher resolution imagery and Ž eld work.
Future work with this method will include a study of the eVect of changing scales
and classiŽ cation accuracy. The scale of classiŽ cation necessary to provide meaningful
biogenic input for local and regional ozone models has yet to be determined. We
will continue to work on locally calibrating the SLCR classes with higher resolution
satellite imagery and Ž eld studies. We also eagerly await the availability of MODIS
data. The higher spectral resolution, more precise bandwidths, and (some) higher
spatial resolution channels (250, 500 m) may prove very useful for this application.
5. Conclusions
The multi-temporal Landsat TM classiŽ cation was successful in measuring oak
density in the Oak Ridge, Tennessee, area. The classiŽ cation was achieved using
detailed ground transects to provide training sites for the supervised classiŽ cation.
Information was extracted from the early fall and winter TM scenes and added to
the Ž nal classiŽ cation as layers. Field observations and comparisons with the FIA
and BELD databases support the accuracy of the classiŽ cation. The isoprene  ux
maps created from the TM classiŽ cation predicted the isoprene  uxes observed in
the Ž eld to within a factor of 2.
Existing landcover data (SLCR) with 1 km pixels are available from the USGS
via the World Wide Web. The correspondence between the TM and SLCR classiŽ ca-
tions is fairly poor. Isoprene emission estimates from the SLCR data are two to
three times greater than the measured (and TM estimated)  uxes. This diVerence is
small enough, however, to encourage continued work in calibrating SLCR data with
Ž eld and other satellite landcover measurements. Due to its lower cost and smaller
volume, SLCR data should prove to be more practical than TM data for compiling
isoprene  ux estimates required by regional and global chemistry models.
Acknowledgments
This research was partially supported by the US Environmental Protection
Agency, Research Triangle Park, NC, under Interagency Agreement Grant No.

2810
Measurement of oak tree density with L andsat T M data
DW49934973-01-0, and by the Southern Oxidants Study (SOS) under Assistance
Agreement No. CR 817766 to North Carolina State University. The National Center
for Atmospheric Research is sponsored by the National Science Foundation.
References
Chavez, P. S., 1988, An improved dark-object subtraction technique for atmospheric scattering
correction of multispectral data. Remote Sensing of Environment, 24, 459–479.
Geron, C., Guenther, A., and Pierce, T., 1994, An improved model for estimating emissions
of volatile organic compounds from forests in the eastern United States. Journal of
Geophysical Research, 99, 12 773–12 791.
Guenther, A., Zimmerman, P., and Wildermuth, M., 1994, Natural volatile organic com-
pound emission rate estimates for US woodland landscapes. Atmospheric Environment,
28, 1197–1210.
Guenther, A., Hewitt, C., Erickson, D., Fall, D., Geron, C., Graedel, P., Harley, L.,
Klinger, L., Lerdau, M., McKay, W., Pierce, T., Scholes, B., Steinbrecher, R.,
Tallamraju, R., Taylor, J., and Zimmerman, P., 1995, A global model of natural
volatile organic compound emissions. Journal of Geophysical Research, 100, 8873–8892.
Guenther, A., Baugh, W., Davis, K., Hampton, G., Harley, P., Klinger, L., Vierling,
L., Zimmerman, P., Allwine, E., Dilts, S., Lamb, B., Westberg, H., Baldocchi,
D., Geron, C., and Pierce, T., 1996a, Isoprene  uxes measured by enclosure, relaxed
eddy accumulation, surface layer gradient, mixed layer gradient, and mixed layer mass
balance techniques. Journal of Geophysical Research, 101, 18 555–18 567.
Guenther, A., Greenberg, J., Harley, P., Helmig, D., Klinger, L., Vierling, L.,
Zimmerman, P., and Geron, C., 1996b, Leaf, branch, stand and landscape scale
measurements of volatile organic compound  uxes from US woodlands. T ree
Physiology, 16, 17–24.
Guenther, A., Zimmerman, P., Klinger, L., Greenberg, J., Ennis, C., Davis, K., Pollock,
W., Westberg, H., Allwine, G., and Geron, C., 1996c, Estimates of regional natural
volatile organic compound  uxes from enclosure and ambient measurements. Journal
of Geophysical Research, 101, 1345–1359.
Haagen-Smit, A. J., and Fox, M. M., 1954, Photochemical ozone formation with hydrocarbons
and automobile exhaust. Journal of the Air Pollution Control Association, 4, 105–109.
Hansen, M. H., Frieswyck, T., Glover, J. F., and Kelly, J. F., 1992, T he Eastwide Forest
Inventory Database: Users Manual, General Technical Report, NC-1512 (St. Paul, MN:
Forest Service, Department of Agriculture).
Jensen, J. R., 1986, Introductory digital image processing, a remote sensing perspective
(Englewood CliVs, NJ: Prentice Hall).
Johnson, D. W., and Van Hook, R. I., (editors), 1989, Analysis of biogeochemical cycling
processes in Walker Branch Watershed (New York, NY: Springer-Verlag).
Kinnee, E., Geron, C., and Pierce, T., 1997, United States land use inventory for estimating
biogenic ozone precursor emissions. Ecological Applications, 7, 46–58.
Kruse, F. A., Lefkoff, A. B., Boardman, J. W., Heidebrecht, K. B., Shapiro, A. T.,
Barloon, P. J., and Goetz, A. F. H., 1993, The spectral image processing system
(SIPS)—interactive visualization and analysis of imaging spectrometer data. Remote
Sensing of Environment, 44, 145–163.
Lamb, B., Guenther, A., Gay, D., and Westberg, H., 1987, A national inventory of biogenic
hydrocarbon emissions. Atmospheric Environment, 21, 1695–1705.
Lamb, B., Gay, D., Westberg, H., and Pierce, T., 1993, A biogenic hydrocarbon emission
inventory for the U.S. using a simple forest canopy model. Atmospheric Environment,
27, 1673–1690.
National Research Council, 1991, Rethinking the Ozone Problem in Urban and Regional Air
Pollution (Washington, D.C.: National Academy Press).
Snyder, J., 1987, Map projections; a working manual. US Geological Survey professional paper
1395, Washington D.C.
Wolter, P. T., Mladenoff, D. J., Host, G. E., and Crow, T. R., 1995, Improved forest
classiŽ cation in the northern lake states using multi-temporal Landsat imagery.
Photogrammetric Engineering and Remote Sensing, 61, 129–143.