Content uploaded by Jesslyn F. Brown
Author content
All content in this area was uploaded by Jesslyn F. Brown on Dec 06, 2017
Content may be subject to copyright.
Measuring Phenological Variability from Satellite Imagery
Author(s): Bradley C. Reed, Jesslyn F. Brown, Darrel VanderZee, Thomas R. Loveland, James
W. Merchant, Donald O. Ohlen
Source:
Journal of Vegetation Science,
Vol. 5, No. 5, Applications of Remote Sensing and
Geographic Information Systems in Vegetation Science (Nov., 1994), pp. 703-714
Published by: Opulus Press
Stable URL: http://www.jstor.org/stable/3235884
Accessed: 22/10/2008 12:46
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=opulus.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the
scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that
promotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.
Opulus Press is collaborating with JSTOR to digitize, preserve and extend access to Journal of Vegetation
Science.
http://www.jstor.org
Journal
of Vegetation
Science 5: 703-714, 1994
? IAVS; Opulus
Press Uppsala.
Printed
in Sweden 703
Measuring phenological variability from satellite imagery
Reed, Bradley C.1', Brown, Jesslyn F.1, VanderZee, Darrel1, Loveland, Thomas R.2,
Merchant, James W.3 & Ohlen, Donald O.1
'Hughes STX Corporation, EROS Data Center, Sioux Falls, SD 57198, USA; 2U.S. Geological Survey, EROS Data
Center, Sioux Falls, SD 57198, USA; 3Center
for Advanced Land Management Information Technologies,
Conservation and Survey Division, University of Nebraska-Lincoln, Lincoln, NE 8588-0517, USA;
*Authorfor
correspondence;
Tel. +1 605 594 6012; Fax +1 605 594 6589; E-mail
REED@EDCSERVERl.CR.USGS.GOV
Abstract. Vegetation phenological phenomena are closely
related to seasonal
dynamics
of the lower atmosphere
and
are
therefore
important
elements
in global models and
vegetation
monitoring.
Normalized
difference
vegetation
index (NDVI)
data derived from the National Oceanic and Atmospheric
Administration's
Advanced Very High Resolution Radiom-
eter
(AVHRR)
satellite sensor
offer a means of efficiently and
objectively
evaluating phenological
characteristics over large
areas. Twelve metrics linked to key phenological
events were
computed based on time-series NDVI data collected from
1989 to 1992 over the conterminous
United States. These
measures include
the onset of greenness,
time of peak NDVI,
maximum NDVI, rate of greenup, rate of senescence, and
integrated
NDVI. Measures of central
tendency
and variabil-
ity of the measures were computed
and analyzed
for various
land cover types. Results from the analysis showed strong
coincidence between the satellite-derived
metrics and pre-
dicted phenological characteristics.
In particular,
the metrics
identified interannual
variability of spring wheat in North
Dakota,
characterized
the
phenology
of four
types
of grasslands,
and established
the
phenological
consistency
of deciduous and
coniferous
forests. These results have implications
for large-
area land cover mapping and monitoring.
The utility of re-
motely sensed data as input
to vegetation
mapping
is demon-
strated
by showing the distinct phenology of several land
cover types. More stable information contained
in ancillary
data
should be incorporated
into the
mapping process, particu-
larly
in areas with high phenological
variability.
In a regional
or global monitoring system, an increase in variability
in a
region may serve as a signal to perform
more detailed land
cover analysis with higher
resolution
imagery.
Keywords: GIS; Land cover; Remote sensing; Time-series
analysis; Vegetation
monitoring.
Abbreviations: AVHRR = Advanced
Very High Resolution
Radiometer;
GIS = Geographical
Information
Systems;
MSS
= Landsat
Multispectral
Scanner;
NDVI = Normalized
Differ-
ence Vegetation Index; NOAA = National Oceanic and At-
mospheric
Administration.
Introduction
Seasonal characteristics of plants, such as emer-
gence and senescence, are closely related to characteris-
tics of the lower atmosphere, including the annual cycle
of weather pattern shifts, and temperature and humidity
characteristics. Changes in phenological events may,
therefore, signal important year-to-year climatic vari-
ations or even global environmental change. The timing
and progression of plant development may provide in-
formation to help researchers make inferences about the
condition of plants and their environment (e.g. soil
moisture, soil temperature, illumination, temperature).
Although most phenological research has ecosystem
monitoring as an ultimate goal, the phenology of entire
ecosystems has rarely been studied. In fact, most work
on phenology has been conducted with individual or-
ganisms or individual species of plants. In recent years,
however, satellite imagery available globally with a
daily repeat cycle, has provided data to examine and
monitor phenological events over large regions. Nor-
malized difference vegetation index (NDVI) data de-
rived from the National Oceanic and Atmospheric Ad-
ministration's (NOAA) Advanced Very High Resolu-
tion Radiometer (AVHRR) offer a means for objec-
tively evaluating phenological characteristics of land
cover regions and assessing their variability over large
geographic areas.
In this study, a suite of 12 measures (metrics), de-
rived from 4 years of time-series NDVI data, were
developed to describe phenological phenomena of land
cover types in the conterminous United States. The
interannual variability of these metrics in selected land
cover regions was then assessed to provide information
concerning cover types and phenological characteristics
that are most consistent from year to year. The goal was
to assess the validity of the metrics and their potential
for monitoring global environmental change.
Reed, B.C. et al.
Background
The topic of phenology
is most often
associated
with
agricultural
events such as planting,
emergence, fruit-
ing, and harvest. Lieth (1974) noted however, that the
U.S. International
Biological Program
Phenology
Com-
mittee
suggested
a broader definition
of phenology:
"the
study of the timing of recurring
biological events, the
causes of their timing with regard
to biotic and abiotic
forces, and the interrelation
among phases of the same
or different species." Seasonality was defined as "the
occurrence of certain obvious biotic and abiotic events
or groups of events within a definite limited period or
periods
of the astronomic
(solar,
calendar)
year"
(Lieth
1974). Factors
influencing
phenology vary by species,
but include
photoperiod,
soil moisture,
soil temperature,
air temperature,
and solar
illumination.
Phenological characteristics of physiognomic types
Satellite
analysis
of phenology
is fundamentally
dif-
ferent from traditional
ground-based
observations.
The
capacity of satellite sensors to detect important
phenological events such as budding and flowering is
limited due to the ground
resolution of the sensors and
the effects of other vegetation and soil background
characteristics. While limited
in some capacities,
satel-
lite sensors are still capable of measuring
broad-scale
changes
in the landscape
that
may
not be associated with
phenological
events of specific plants,
but are descrip-
tive of ecosystem condition. Some of the seasonal char-
acteristics that
can be measured
by satellite and factors
influencing their interannual
variability
are discussed
below by physiognomic
type.
Agricultural crops. Phenological phenomena in non-
irrigated
agricultural
areas are
dependent
on short-term
weather events. Changnon & Kunkel (1992) pointed
out, for example, that 1990 was an anomalous weather
year
in the Midwest,
being
both the warmest and wettest
to date. Excessive precipitation
from January
through
June
in the Midwest
delayed
planting
for several weeks.
The cool, wet spring increased
pest development,
and
then higher than normal winds in spring and summer
minimized opportunities
to spray crops when needed.
These events led to changes in the timing of major
phenological events (emergence, senescence), particu-
larly in areas
of dryland
agriculture.
Irrigated
areas,
on
the other
hand, are relatively more phenologically sta-
ble, assuming
that the same crops are
planted.
Grasslands. Grasses can be considered as mainly
peren-
nials that annually shed photosynthetic
tissue and re-
spond rapidly to environmental factors such as day
length, soil water and soil temperature.
In the tallgrass
prairie,
biomass reaches its maximum in early summer
(Barnes
et al. 1983), while shortgrass prairies
normally
reach their
maximum
later
in the growing season. An-
nual
grasses, such as in the California and
desert
grass-
lands,
have less predictable
phenological
patterns
since
they are
dependent
on less reliable
rainfall events. Vari-
ations
in the precipitation
regime during
any particular
season
are
readily
reflected
in productivity,
especially
at
drier sites (Lewis 1971). The shortgrass
prairie and
annual
grasslands
act as pulsed systems,
with
vegetative
growth
responding
to rainfall
events (Rauzi
& Dobrenz
1970). Phenological events in the tallgrass
prairie
are
expected to be relatively stable in comparison
to the
other
grasslands
since precipitation
is not as limiting.
Desert shrublands. The temperature
regime in warm
deserts is usually
not limiting;
instead,
nearly
all pheno-
logical events and photosynthetic
activity are moisture
related.
Cold deserts
have a more limited growing sea-
son in which
photosynthetic
activity
occurs,
but
vegeta-
tive activity
is still moisture
dependent.
Water
availabil-
ity in early spring
seems to be more reliable
in both the
warm and cold deserts of the United States because
moisture
is stored over the winter with minimal
evapo-
ration,
resulting in a relatively regular
surge of photo-
synthetic
activity in the spring. Photosynthetic
activity
during the summer months is less reliable, as it is
dependent
on occasional summer
precipitation.
Forests. Photosynthetic
activity in deciduous
forests is
dependent
on photoperiod,
moisture and temperature
with specific adaptations
ranging between species. A
typical pattern
of photosynthetic
activity in deciduous
forests is a winter minimum followed by a rapid in-
crease to a maximum
by late spring, persisting
at that
level for 2 or 3 months, then decreasing
steadily until
defoliation
in autumn.
Dormancy
appears
to be control-
led
primarily by photoperiod,
but the
critical
photoperiod
varies among and within species (Flint 1974). Ever-
green trees differ primarily
in that a low rate of photo-
synthesis persists in winter during periods with tem-
peratures
above 0 ?C.
Because these patterns
of photo-
synthetic activity are mostly functions of temperature
and photoperiod,
both of which are relatively stable
from
year
to year,
the interannual
variability
of satellite-
derived
phenological
measures over both
deciduous and
coniferous forests is expected
to be relatively
low.
Remote sensing and phenology
Remote sensing of vegetation involves an implicit
link to vegetation
phenology.
The seasonal behavior of
vegetation is a fundamental
component of successful
704
- Measuring phenological
variability
from satellite
imagery
-
image interpretation
(Morain
1974; Lillesand & Kiefer
1987; Lloyd 1990). For land cover assessment, the
timing of image acquisition can be critical. Further-
more, knowledge of crop calendars
and phenology is
often a crucial
element
in vegetation
interpretation.
Since
the 1970s, many
researchers
have
recognized
the poten-
tial of multitemporal
satellite observations to provide
information
about the
phenological
development
of natu-
ral
vegetation
and
crops.
One
early
effort
applied
ERTS-
1 (Landsat)
multispectral
scanner
(MSS) data to identify
and analyze phenologic events over the central Great
Plains (Rouse et al. 1973). This research was one of the
first to use a 'vegetation
index' derived from the red and
near-infrared
channel data
to monitor the 'green
wave'
and 'brown
wave' of rangelands.
Vegetation
indices,
computed
from combinations
of
visible red and near-infrared
spectral measurements,
were developed primarily for vegetation study. The
advantages
of using these numerical
transforms rather
than the original spectral
observations
include the fol-
lowing: minimizing soil and other
background
effects,
reducing data dimensionality, providing a degree of
standardization
for comparison,
and
enhancing
the veg-
etation
signal (Curran
1981;
Malingreau
1989; Goward
1989). One of the more commonly used vegetation
indices, the NDVI, is based on Equation
1:
NDVI = (NIR - RED) / (NIR + RED) (1)
where
NIR are data from
the near-infrared,
and
RED are
data from the visible red part of the electromagnetic
spectrum.
The normalization
in the formula partially
compensates
for changing
illumination conditions and
surface terrain effects (Lillesand
& Kiefer
1987;
Kidwell
1991).
NDVI data
have been
frequently
used in research
on vegetation 'greenness'
or productivity.
In addition
to identifying
the occurrence
of signifi-
cant phenological
events, spectral-temporal
profiles of
vegetation index values represented
by a vegetation
index graphed
over time have been used for the last
decade
to classify different
vegetation
types and to map
their spatial distributions (Badhwar et al. 1982).
Odenweller
& Johnson
(1984) compared
observed
spec-
tral-temporal
profiles
derived
from Landsat 79 m x 79 m
resolution
MSS data
to expected
phenological
patterns
of several different
crop types to assist in crop identifi-
cation. They pointed out that due to Landsat's
revisit
period (16 days), key observations may be missing,
altering
the form of the spectral-temporal
profile.
In another
study, multitemporal
data were used for
agricultural
land cover classification (Lo et al. 1986).
The authors
tested several classification methods and
concluded that multidate Landsat
MSS data using an
unsupervised
classification method provided the best
results.
Knowledge
of crop phenology
was used to label
each cluster according
to an analysis of the temporal
ratio profile of the cluster. The absence of data during
the most critical part
of the growing season (late July
and August) was cited as a key reason
for results
in the
84-86 % accuracy
range.
Many
initial
experiments
incorporated
Landsat 30 m
x 30 m resolution Thematic Mapper (TM) and MSS
data, however, the 16-day repeat cycle (18 days with
Landsats 1 - 3) of the satellite restricted the ability to
track the seasonal characteristics
of vegetation. The
launch of daily orbiting
weather
satellites provided
re-
searchers with the ability to track
phenological events
more closely.
The AVHRR sensors are carried
on weather satel-
lites operated by NOAA. The AVHRR records data
with 1 km x 1 km nominal spatial resolution and has
daily global coverage. The primary
advantage of the
AVHRR is its frequent temporal
coverage over large
geographic areas, which allows a greater
opportunity
for obtaining cloud-free coverage during important
phenological
stages of the land cover. The sensor
speci-
fications for the AVHRR in comparison
to the Landsat
MSS and
TM sensors are given in Table 1.
Some of the first attempts
to study the phenologic
patterns
of large
areas
incorporated
resampled
AVHRR
products
distributed
by NOAA, (global area
coverage
at
4 km
x 4 km nominal resolution and global vegetation
index at 16 km x 16 km nominal resolution) (NOAA/
National Environmental Satellite
Data and
Information
Service (NESDIS) 1990;
Kidwell 1991). These data are
available
over large areas and provide
researchers
with
Table
1. AVHRR,
MSS,
and
TM sensor
specifications.
Sensor Repeat Channel bandwiths Spatial
cycle (um) resolution
AVHRR 12 hours 0.58-0.68 1.1 km
0.72-1.10
3.55-3.93
10.3-11.3
11.5-12.5
Landsat MSS 16 days 0.5-0.6 79 m
0.6-0.7
0.7-0.8
0.8-1.1
Landsat
TM 16 days 0.45-0.52 30 m
0.52-0.60
0.63-0.69
0.76-0.90
1.55-1.75
2.08-2.35
10.4-12.5
705
Reed, B.C. et al.
smaller,
more manageable
data
volumes. A number
of
studies directed
toward
analysis of global or continen-
tal-scale vegetation patterns
using AVHRR data were
initiated
in the mid-1980s (Goward
et al. 1985; Justice
et al. 1985; Tucker et al. 1985; Townshend et al. 1987).
Many
of these
investigations
incorporated
multitemporal
profiles (the NDVI plotted against time), frequently
using
these
graphs
as indicators
of vegetation
phenology.
Goward
et al. (1987) studied selected biomes in North
and South
America,
observing
that
temporal
variations
of similar
vegetation
types
were related to known
clima-
tology of the continents. These authors
demonstrated
that a time integral
of the NDVI measurements
over an
annual
period (the area under
the temporal
curve) pro-
duced a measurement
related to published
net primary
productivity
values of different
biomes. The explained
variance
(r2)
of this relation was 0.90.
Malingreau
(1986) used weekly AVHRR
NDVI ob-
servations from 1982 to 1985 over Asia to derive infor-
mation about vegetation dynamics and cropping pat-
terns. He identified the timing of phenological events
such as emergence and end of growing season using
multitemporal
profiles. He also showed that climate
anomalies
were exhibited
in the satellite NDVI tempo-
ral
curves.
Malingreau
created a physiognomic
classifi-
cation of vegetation of southeast Asia using the time
integral
of growing season NDVI values.
There is a growing body of research
indicating
that
the phenologic behavior of different broad
vegetation
types can be observed,
analyzed,
and mapped
using the
NDVI multitemporal profiles.
Lloyd (1990) employed
a
phenologic approach
to global vegetation
cover classifi-
cation. He used a supervised
binary
decision-tree
classi-
fier incorporating
three variables
extracted from NDVI
multitemporal
profiles: time of maximum photosyn-
thetic activity, length of growing season, and annual
mean daily maximum potential photosynthetic rate
(which is equivalent
to the annual time integral
of the
NDVI mentioned
previously).
Methodology
An automated,
quantitative
approach
was developed
to derive phenological measures from multitemporal
AVHRR
NDVI observations
in a consistent,
systematic
manner. In order to explore and analyze the variability
of phenologic parameters
for certain
land cover types, a
per-pixel derivation of the phenological metrics was
created and the variability
of the metrics
was evaluated.
Input
data
from AVHRR satellite imagery, and land cover data
also derived from AVHRR data. The U.S. Geological
Survey's
Earth
Resources
Observation
System (EROS)
Data Center
has created 14-day
maximum
NDVI com-
posite data sets for the conterminous
United States for
each
year
since 1989. The
yearly
data
sets are
comprised
of 21 14-day maximum NDVI composites, including
each 2-week period from March
through
October
and
one 2-week period for each month from November
through
February
(Eidenshink
1992). A sequence of 4
years
of NDVI data
(1989 through
1992) are used in this
study.
There are several characteristics
of the biweekly
composited
NDVI data that
should be considered.
Posi-
tive characteristics
include:
(1) it is a consistently
proc-
essed data set; (2) there is dense temporal
coverage;
(3) there is a relatively
low data volume covering
large
areas;
and (4) cloud contamination
is greatly
reduced.
Negative aspects of the biweekly compositing
process
include;
(1) atmospheric
corrections are not included
in
the processing, (2) there is a bias toward high view
angles at some locations (Allen et al. 1994), (3) the
biweekly period may be too long to determine
some
phenological
events in sufficient
detail for some appli-
cations, and (4) even with the maximum value NDVI
compositing process, not all cloud contamination
is
removed. The approach
for characterizing
phenology
that is used in this study compensates
for many of the
factors
that
affect
image
quality by parameterizing
curve
characteristics instead
of only using the biweekly
NDVI
values.
Land cover data were derived from the land cover
characteristics data base
created at the
EROS
Data Center
(Loveland
et al. 1993). The data
base portrays
regions
composed of similar
land cover mosaics, as defined
by
multitemporal
AVHRR NDVI data collected in 1990,
and
attributes of these regions
including
terrain,
climate
and ecoregions (Loveland et al. 1991; Brown et al.
1993). There are 159 land cover types defined in the
data
base. Because the classification was originally per-
formed on time series NDVI data derived from the
AVHRR in 1990, there is a potential
for a bias toward
less variability
within land cover classes for this year.
However, when analyzing the results, there was not a
clear trend toward lower variability exhibited in the
1990 data.
Because of high computational
demands,
a system-
atic 1 % sample (every tenth line and tenth sample)
of
the conterminous
United States NDVI and land cover
characteristics data base was used in this study. This
resampling
resulted in a total of 92,923 land pixels for
study.
Two primary
data sets were used: NDVI data derived
706
- Measuring phenological
variability
from satellite
imagery
-
.8- ...... Raw Data
.7 - - Smoothed Data
.6 - ' ,
.5 -
~.4
.2
0
TIME
Fig. 1. Noise in cloud contaminated
pixel for a single pixel in
area
of irrigated
agriculture
and results
of data
smoothing.
Data preprocessing
In order to extract the phenological metrics, some
preprocessing
of the data was necessary.
An underlying
assumption
in many
time-series
analyses
is that the data
observations
are regularly
spaced
in time (Hoff 1983).
This is not the case with the NDVI composites
which are
generated
every 14 days between March and October,
but for only one 14-day period per month between
November and February.
As a result, only 21 data
values (unevenly spaced in time) were available for
each of the 4 years. To correct this, additional data
points were added
to each year's observations
by inter-
polating between adjacent
observations
for November
through
February.
Another
problem
was caused by cloud contamina-
tion for a substantial
number of pixels. To adjust
for
this,
a line-smoothing algorithm
was applied
to the
time-series
NDVI data.
The nonlinear
running
median line-smoother
implemented in the 'S' statistical software packagel/
was chosen as the algorithm
for reducing
the effects of
cloud contamination
(Becker
et al. 1988; Tukey 1977).
This technique
was effective at preserving
the essence
of the NDVI time-series,
while eliminating
much of the
contaminated
data (Fig. 1). This method reduced the
NDVI peaks in the curves which are assumed to be
valid,
uncontaminated
NDVI values,
therefore
the origi-
nal, unsmoothed
data values were used to compute as
many of the metrics as possible.
Phenological metrics
12 seasonal metrics derived from AVHRR NDVI
data were defined, many of which were adapted
from
Lloyd's (1990) and Malingreau's
(1986) lists of sea-
sonal
parameters.
The metrics
can be divided
into three
types;
(1) temporal
(based
on the
timing
of an
event), (2)
Table 2. 12 seasonal NDVI
metrics and their
phenological
interpretation.
Metric Phenological interpretation
Temporal NDVI metrics
Time of onset of greenness Beginning of measureable photosynthesis
Time of end of greenness Cessation of measureable photosynthesis
Duration of greenness Duration of photosynthetic activity
Time of maximum NDVI Time of maximum measureable photo-
synthesis
NDVI-value metrics
Value of onset of greenness Level of photosynthetic activity at
beginning of growing season
Value of end of greenness Level of photosynthetic activity at end of
growing season
Value of maximum NDVI Maximum measureable level of photo-
synthetic activity
Range of NDVI Range of measureable photosynthetic
activity
Derived metrics
Time-integrated NDVI Net primary production
Rate of greenup Acceleration of photosynthesis
Rate of senescence Deceleration of photosynthesis
Modality Periodicity of photosynthetic activity
NDVI-based (the NDVI value at which events occur),
and (3) metrics derived
from time-series characteristics
(Table 2). These metrics may not necessarily corre-
spond directly to conventional, ground-based
pheno-
logical events, but provide
indicators of ecosystem dy-
namics. For example, the moment of a crop's emer-
gence may not be distinguishable
from the soil back-
ground,
especially at the AVHRR sensor's 1 km x 1 km
spatial
resolution. But these
distinctive 'events' as meas-
ured from satellite (such as a rapid
increase in NDVI)
signal a measurable
change in ecosystem characteris-
tics.
There are considerable difficulties
in using
time-series
NDVI data for identifying
one of the key phenological
events: the beginning
of the growing
season (or onset of
greenness).
Past efforts
have
included
assigning
a thresh-
old NDVI value (e.g. 0.099; Lloyd 1990) at which
vegetative activity is assumed
to begin. This threshold,
however, varies with vegetation
type, soil background,
and illumination
conditions.
Therefore,
it is not
possible
to establish
a single, meaningful
threshold
that
signifies
the onset (or end) of vegetative activity for the wide
variety of cover types that occur in the conterminous
United States. The approach
taken
in this
study
identifies
when the NDVI exhibits a sudden increase that may
signal the onset of significant
photosynthetic
activity.
Once the onset of the green period
is identified
(and
the
end of the green period, using similar methods), the
other
phenological
metrics
are
relatively easily
computed.
The method chosen for extracting
features
from the
707
Reed, B.C. et al.
.8- ........
Moving Average
.7 - Smoothed Data
.5
.4
z
.2-
.1 -1
TIME
Fig. 2. Smoothed NDVI time-series and moving average
curves for a single pixel in an area of deciduous
forest.
NDVI time-series
uses a comparison
to a moving aver-
age of the time-series to identify departures
from an
established trend.
This approach
is adapted
from auto-
regressive moving
average
(ARMA)
models
(Hoff 1983;
Granger
1989). The calculation
of the moving average
is given in Equation
2:
Yt = (Xt+Xt-I_ + X,_t-2+.Xt-(w-)) /w (2)
where Yt
is the moving average
value for time t, Xt
is the
smoothed
NDVI value for time t, and w is the moving
average time interval (number of 14-day composite
periods,
in this case).
To identify trend
changes in the NDVI time-series
data
(e.g. a rapid
increase)
the moving average
is used as
a 'predictor' against which the actual observed data
values are
compared.
When
the
moving
average
is calcu-
lated for the smoothed
NDVI time-series,
a new series
is
produced
that
has an introduced
time lag (Fig. 2).
To identify the onset of the growing seasons, an
average
of the nine previous
smoothed
NDVI biweekly
composite
values was compared
to the smoothed
NDVI
values. Selecting the moving average
time interval
(the
number of NDVI composite periods used to calculate
the moving average) is a critical issue; a large time
interval results in a less sensitive trend
detector and a
small interval
may pick up insignificant
trend
changes
(Maxwell 1976). With NDVI data,
the wavelength
of a
typical cycle (the seasonality) should be considered
when establishing
the moving average
time interval
in
order to avoid the influence of the previous cycle (Hoff
1983). After experimenting
with several moving aver-
age intervals
varying from 3 to 15 composite periods
and examining the results for several cover types, a
moving average interval of 9 composite periods was
selected.
q,
0 5 10 15 20 25
14 Day Periods
Fig. 3. Derivation of phenological
metrics from actual
tempo-
ral NDVI profile. OnP = Time of onset of greenness;
OnV =
NDVI value at onset;
EndP
= Time of end of greenness;
EndV
= NDVI value at end;
DurP
= Duration of greenness;
MaxP
=
Time of maximum
NDVI; MaxV = Maximum NDVI value;
RanV = Range of NDVI values; RtUp = Rate of greenup;
RtDn
= Rate of senescence;
TINDVI
= Time-integrated
NDVI.
The onset-of-greenness metrics (time of onset and
onset NDVI value) were defined as the period beginning
at the point where the smoothed time-series data crossed
the moving average in an upward direction and re-
mained above it for the greatest sustained increase in
NDVI (Fig. 3). The crossing
illustrates a departure
from
the value predicted
by the moving average and repre-
sents a significant trend change. The end-of-season
metrics
(end period and ending NDVI value) were de-
termined
in a similar manner,
but the moving average
was applied
in reverse
chronological
order.
The seasonal duration,
measured
in days, was de-
rived
by subtracting
the time of onset of greenness
from
the time of the end of season (Fig. 3). The original,
unsmoothed data were used to identify the maximum
NDVI composite
period
and the value of the maximum
NDVI for the growing season. The original data
were
used
for these metrics since the maximum
NDVI values
are likely to be uncontaminated,
with no smoothing
needed.
After
establishing
the metrics for the maximum NDVI
value, the range
in NDVI was computed
by subtracting
the NDVI value of either the onset or end of greenness,
whichever was lower, from the maximum NDVI value
(Fig. 3). The rates of greenup and senescence were
computed
as straight
line slopes from onset to the maxi-
mum and from the maximum to the end, respectively
(Fig. 3). To determine the time-integrated
NDVI, the
area under the smoothed curve was used because cloud
contamination biases data values downward
(Fig. 3).
Finally, modality was determined
in conjunction
with
the identification
of the end of season. Those seasons
for
708
- Measuring
phenological variability
from satellite
imagery
-
which there was more than one significant
decrease in
the smoothed NDVI were identified as multi-modal.
Analysis
Measures of central
tendency
and dispersion
of the
12 metrics were calculated for each land cover type.
Because there
is still a degree
of cloud and
atmospheric
contamination in the
data,
even after
smoothing,
anoma-
lous metric
values occasionally
occur.
To minimize the
effect of these outliers
on the statistics,
the median and
Median Absolute Deviation About the Median (or
MADAM; Huber 1981) were used as the measures
of
central
tendency
and
dispersion, respectively,
for all of
the metrics
except modality.
Modality
was summarized
by calculating
the percentage
of pixels within each class
that were multi-modal.
When analyzing
the results,
it is important
to realize
that these metrics
were calculated for each pixel within
the sample and then summarized
by land cover class.
For
each class, there is a degree
of inherent within-class
variability
due to varying soils, latitude, illumination,
and other factors, even under optimal growing condi-
tions. To identify
areas
of change
or areas
experiencing
stressful
conditions,
an observer
should
look for differ-
ences in the MADAM from one year to another rather
than
looking at the MADAM value in a single year.
Because the methodology
utilizes a moving average
time interval of nine composite
periods (requiring
data
from eight previous composite periods), results ob-
tained for the onset of greenness
are valid only for 1990,
1991, and 1992. Similarly,
the end of greenness
metric
is valid only for 1989, 1990, and 1991. Consequently,
many of the other metrics
that are dependent
upon the
onset and
end,
can
be calculated
only for 1990 and
1991.
Results are
presented
only for these two years.
Results and Discussion
The analysis was performed
on all 159 land cover
classes from the EROS land cover characteristics data
base, but for ease of presentation
and to make some
general observations,
we have selected several exam-
ples to represent major land cover types including:
agricultural
crops, grasslands,
desert
shrubs,
and forests
(Fig. 4). The median of the phenological metrics for
1990 and 1991 are
given in Table 3 for the various land
cover
types. The MADAMs
of the phenological
metrics
are shown in Table 4. Although the methodology de-
scribed
in the previous section uses the composite pe-
riod as the unit of measurement to derive the metrics,
they are translated into more commonly used units for
ease of presentation.
The
units for
the
time
period
metrics
are calendar
dates,
the NDVI value metrics
are
given in
NDVI units, which vary from - 1.0 to 1.0. Time-inte-
grated
NDVI is also given in NDVI units and the rates
of
greenup and senescence are given in NDVI units per
composite period.
Modality
is expressed
as the percent-
age of pixels within a land cover type exhibiting a
multimodal
NDVI. No measure
of dispersion
is given
for modality.
Agricultural crops
Agricultural crops analyzed were of three types:
small grains,
row crops, and irrigated
crops. The grain
crop analyzed was spring wheat located in North Da-
kota. The median onset of greenness is quite different
between the 2 yr (April 8 and
February
25 for 1990 and
1991,
respectively), possibly
due to the time of snowmelt
(Table 3). After the snow melts, there
is a rapid
increase
in NDVI since the wheat
is already
actively growing.
In
1990, the median rate
of greenup
was more
rapid
than in
1991 (0.08 vs. 0.05 NDVI per
composite period),
result-
ing in similar timing of the remaining phenological
metrics with the exception of duration
of growing sea-
son (224 vs. 266 days). The median time-integrated
NDVI is higher (3.13 vs. 2.76) and more variable as
measured
by the MADAM (0.285 vs. 0.409) in 1991,
the year
with an earlier
greenup
(Table
4). These results
are consistent with statements
by Changnon
& Kunkel
(1992) regarding
the characteristics
of 1990 weather
patterns.
For row crops, analyses were made of the corn and
soybean belt of Indiana,
Illinois, Iowa, South Dakota,
and Nebraska. These row crops are consistent in their
phenological
metrics
in the 2 yr (Table
3). The onset of
greenness
is 14 days apart
in the 2 yr, but since the 2-
week composite period is the unit of measurement
in
this methodology, 14 days is the minimum
measurable
difference. The other metrics and their variability
are
similar
(Table 4). The consistency of the phenological
metrics
is a result
of the more
reliable
moisture
regime
in regions where row crops are
normally
grown.
An example of irrigated
crops was selected from
southern Texas. Table 3 shows that
the median
time of
the end of growing season (July 15 vs. September
23)
and duration
of growing season (140 vs. 196 days) for
this area were quite
different
in the 2 yr, and
the within-
year variability (MADAM) of time-integrated
NDVI
was considerably higher in 1991 (0.378 vs. 0.281),
perhaps indicating
that at least a portion
of this region
may
have been planted
with a different
crop
(Table
4). If
the same crops are planted in subsequent
years, the
phenological metrics should be very consistent since
moisture is not a limiting factor. If the metrics are
significantly
different,
this probably
signals a change
in
709
Reed, B.C. et al.
; /I
' ' I . .. 1
-ji: -l 4 ^- ;
K
'k r r ; -!* " - Y-n -
, '.
, W..! . . i '*i..
.,4:'-:~'':~:':~:~'~:"'
?..'-~','..:,:,:'-h~:"~.~
.'7'
...
..... ,:,)??~;..
, .: -,;,
~ -. -...., .... ':-=r...
' ';/.~~~~~~~~~~~~~~~~~~~~~~~~vy
::?:::..:.:..:.:..?,,~.-.;r:....!
.:.~, . .':... ...'i5 ...., "...
:3 5 1I 1
11
16:l22 30 58 62 65 66E69 T9S 84 '192 g95197 98B 119
Fig. 4. Distribution of land cover classes used as representative
examples.
3 = spring
wheat; 11 = corn
and soybeans;
22 = irrigated
crops;
58 = short
grasslands;
62 = annual
grasses;
65 = tall grasslands;
69 = creosote
shrubland;
79 = desert
grassland;
84 = Chaparral;
95 = oak-hickory forest; 98 = southern coniferous forest; 119 = western coniferous forest.
Table 3. Median of phenological metrics for 1990 and 1991.
Median phenological metrics
Land cover type Year Onset End Duration Peak Onset End Range Peak Time-int. Rate of Rate of
value value of NDVI value NDVI greenup senescence Modality
Spring wheat 1990 4/8 11/18 224 7/15 0.10 0.09 0.49 0.58 2.76 0.08 0.05 0.0
1991 2/25 11/18 266 7/1 0.09 0.07 0.50 0.57 3.13 0.05 0.04 0.0
Corn and soybeans 1990 3/11 11/18 252 8/12 0.11 0.10 0.55 0.65 4.06 0.04 0.07 0.0
1991 2/25 11/4 252 7/29 0.10 0.09 0.56 0.65 4.12 0.04 0.06 0.0
Irrigated crops 1990 2/25 7/15 140 6/3 0.11 0.14 0.27 0.38 0.98 0.03 0.08 13.8
1991 3/11 9/23 196 5/20 0.13 0.15 0.25 0.38 1.10 0.04 0.03 13.8
Short grasses 1990 4/22 12/2 224 8/26 0.12 0.12 0.22 0.34 1.53 0.02 0.02 1.8
1991 3/25 11/18 238 7/29 0.12 0.12 0.19 0.31 1.38 0.02 0.02 0.7
Annual grasses 1990 11/5 10/21 352 3/25 0.20 0.16 0.32 0.48 2.06 0.03 0.02 0.0
1991 1/28 8/27 210 5/6 0.16 0.20 0.35 0.51 1.94 0.05 0.03 0.2
Tall grasses 1990 3/25 11/18 238 6/17 0.17 0.18 0.39 0.56 3.45 0.05 0.03 0.0
1991 3/11 11/18 252 6/17 0.17 0.14 0.43 0.57 3.15 0.05 0.03 0.2
Desert grasses 1990 7/29 12/2 126 10/7 0.13 0.16 0.23 0.36 1.07 0.04 0.02 18.7
1991 4/8 11/18 224 9/23 0.16 0.13 0.18 0.31 1.04 0.01 0.02 13.6
Creosote shrubland 1990 7/29 12/2 126 5/20 0.12 0.12 0.10 0.22 0.40 0.02 0.01 2.2
1991 3/25 11/18 238 9/9 0.13 0.12 0.10 0.22 0.46 0.01 0.01 1.9
Chaparral 1990 1/1 10/21 294 4/22 0.29 0.21 0.26 0.47 1.73 0.02 0.01 2.7
1991 2/11 9/9 210 5/6 0.21 0.25 0.28 0.49 1.74 0.04 0.02 1.4
Oak-Hickory forest 1990 3/11 11/18 252 6/17 0.26 0.27 0.43 0.69 4.39 0.06 0.03 2.4
1991 3/11 11/18 252 6/3 0.27 0.23 0.42 0.65 3.96 0.06 0.03 0.4
Southern pines 1990 3/11 11/18 252 5/6 0.36 0.38 0.19 0.55 1.39 0.04 0.01 38.7
1991 3/11 11/4 238 5/6 0.38 0.33 0.22 0.55 1.28 0.03 0.01 20.7
Western conifers 1990 3/11 11/4 238 6/3 0.38 0.35 0.20 0.55 1.63 0.02 0.01 4.5
1991 2/25 10/21 238 6/17 0.35 0.35 0.21 0.56 1.75 0.02 0.01 1.1
710
- Measuring phenological variability
from satellite
imagery
crop type. This region has a relatively
large percentage
of pixels that are multi-modal
(Table 3). Fig. 5 shows
the consistent, multimodal
temporal
profile of a pixel
within this region. This multimodality
is likely related
to double-cropping
in this region.
Grasslands
Four grassland types were analyzed; short grass-
lands,
tall grasslands,
desert
grasslands,
and annual Cali-
fornia grasslands.
The short grasslands
and tall grass-
lands are quite consistent in their
phenological param-
eters over the 2 yr
of observation
(Table 3). As expected,
the
tall
grasslands
have an
earlier median onset of green-
ness and
higher
median
time-integrated
NDVI than
the
short
grasses. The annual
grasses of California
and the
desert southwest are inconsistent in their
median time of
onset (November
5 and January
28 for annual
grasses;
July 29 and April 8 for desert grassland)
and duration
(352 and 210 days for annual
grasses;
126 and
224 days
for desert
grassland)
in the 2 yr (Table 3). This is likely
due to the lack of water storage in the soil and the
dependence
of the grasses on seasonally
unpredictable
rainfall events. The within-class variability of these
events is higher in the California
grasslands
(Table 4).
The desert
grasslands
exhibit
high multimodal
phenolo-
gical profiles since even a small
rainfall
event can cause
a surge
of growth
in these areas
(Table 3).
.5
0 .4
z
.3
.2
.1
0
TIME
Fig. 5. Multimodality
of irrigated crop.
Shrublands
The shrublands,
a creosote shrubland
in the desert
Southwest and California
chaparral,
show the highest
phenological variability
of the physiognomic
types (Ta-
bles 3 and 4). The median onset and end of greenness
and the time of maximum NDVI are highly variable
within and between
years
for both shrublands
(Table 4).
Table 4. Median absolute
deviation about the median
of phenological
metrics for 1990 and 1991.
MADAM phenological metrics
Land cover type Year Onset End Duration Peak Onset End Range Peak Time-int. Rate of Rate of
value value of NDVI value NDVI greenup senescence
Spring wheat 1990 28 14 28 14 0.022 0.016 0.044 0.03 0.285 0.027 0.005
1991 14 28 14 14 0.019 0.014 0.045 0.03 0.409 0.010 0.006
Corn and soybeans 1990 14 0 14 14 0.019 0.024 0.033 0.02 0.297 0.005 0.009
1991 0 14 14 28 0.025 0.016 0.045 0.03 0.380 0.004 0.015
Irrigated crops 1990 28 14 42 14 0.019 0.119 0.069 0.06 0.281 0.009 0.02
1991 14 42 56 14 0.019 0.043 0.067 0.05 0.378 0.013 0.008
Short grasses 1990 14 14 14 14 0.024 0.018 0.035 0.03 0.364 0.005 0.005
1991 14 14 14 28 0.018 0.018 0.047 0.04 0.393 0.007 0.007
Annual grasses 1990 42 28 56 0 0.038 0.026 0.056 0.05 0.453 0.006 0.006
1991 14 28 28 28 0.025 0.022 0.070 0.05 0.433 0.009 0.009
Tall grasses 1990 14 0 14 0 0.025 0.016 0.026 0.02 0.207 0.003 0.003
1991 14 14 14 0 0.015 0.014 0.040 0.03 0.306 0.004 0.004
Desert grasses 1990 14 0 14 0 0.022 0.019 0.030 0.02 0.193 0.005 0.005
1991 14 14 14 14 0.020 0.015 0.031 0.03 0.370 0.009 0.009
Creosote shrubland 1990 14 14 28 0 0.013 0.013 0.024 0.02 0.141 0.011 0.005
1991 112 14 70 28 0.014 0.016 0.024 0.03 0.232 0.009 0.008
Chaparral 1990 56 14 48 14 0.036 0.021 0.044 0.03 0.445 0.010 0.004
1991 14 28 28 14 0.023 0.021 0.048 0.045 0.423 0.011 0.006
Oak-Hickory forest 1990 14 0 14 0 0.028 0.029 0.034 0.02 0.405 0.009 0.004
1991 0 14 14 14 0.029 0.028 0.036 0.02 0.514 0.015 0.004
Southern pines 1990 14 14 14 14 0.045 0.033 0.043 0.02 0.567 0.016 0.004
1991 14 14 14 28 0.034 0.033 0.049 0.02 0.586 0.018 0.005
Western conifers 1990 14 14 28 28 0.047 0.057 0.070 0.02 0.670 0.010 0.007
1991 14 14 14 14 0.057 0.042 0.066 0.03 0.681 0.011 0.007
.8
.7
.6
711
Reed, B.C. et al.
The NDVI values at which these events occur and the
median
time-integrated
NDVI, however, are consistent
between the years (Table 3). This shows that
while the
timing
of phenological
events
is highly
variable,
the
low
total production
is relatively
constant.
Forests
Three different
forest types were included:
(1) the
southern
deciduous forest
of the Ozark
and
Appalachian
mountains,
(2) the southern
coniferous forest, and (3)
the western coniferous forest of the Coast Range of
Oregon and California.
The southern
deciduous
forest,
made
up primarily
of Quercus
and
Carya,
has
consistent
phenological
metrics in the two years and a high time-
integrated
NDVI (Tables
3 and
4). These characteristics
are consistent
and distinct
in comparison
with other land
cover types, and
may provide
valuable
information
that
could potentially
be used when classifying land cover
types.
The coniferous forests that were analyzed
included
the loblolly, longleaf, and slash pines of the south,
extending
from
Mississippi
through
Florida and north
to
South Carolina,
and a western coniferous
forest in the
Coast Range of Oregon and California.
Both of these
regions
have very consistent
metrics,
but
because
of the
method used to compute time-integrated
NDVI (i.e.,
relative to the onset and the end of greenness), time-
integrated
NDVI is low in comparison
to other land
cover types (Table 3). However,
the high NDVI values
at which greenness begins and ends, coupled with the
low rates of greenup
and senescence can assist in the
.8
.7
.6
.5
z .4
.3
.2
.1
0
Fig.
6.
Multimodality
of southeastern coniferous
forest caused
by cloud contamination of AVHRR data.
classification
of coniferous
forest.
The interannual
vari-
ability of the coniferous forest, as expected, is low
(Table
4).
The southern
coniferous forest shows the highest
multimodality
of all the cover types (38.7 %
in 1990 and
20.7 %
in 1991) (Table
3). These results
are
unexpected
since there is not a pulsed
growth
pattern
for this type of
vegetation. Closer inspection of this area's seasonal
NDVI profiles shows that the reason for the multi-
modality
is cloud contamination.
There were extended
periods
of very low NDVI (related
to cloud contamina-
tion)
alternating
with
high
NDVI that the
median
smooth-
ing technique
did not completely correct
(Fig. 6). This
results
in a falsely high multimodal
metric.
Discussion
Results from the metrics show strong coincidence
with expected
phenological
characteristics
for the vari-
ous land cover types. In particular, they identified
interannual
variability
of spring
wheat
in North
Dakota
that was consistent
with climatic reports,
illustrated
the
phenological
consistency of corn and soybeans,
identi-
fied an area
of potential
crop
change
in an
irrigated
area,
characterized
the phenology
of four
types
of grasslands,
showed the phenological fluctuations of shrublands,
and established
the phenological
consistency
of decidu-
ous and
coniferous
forests.
While
the
results are
positive, there
are some caveats
in interpreting
these results. Since we are using the
moving average
method
to identify onset of greenness,
we are
actually
measuring
the beginning
of the greatest
increase
in NDVI during
the year.
This is not always
the
onset of greenness 'event' that
we are seeking to iden-
tify. For example, in the northern
Great Plains, the
greatest
increase
in NDVI actually
occurs
at the time of
snowmelt.
As a result,
the onset metric
employed
in this
study
sometimes
gives an earlier
greenup
than
expected
for this region. While the region may not be thought
of
as green at this point in the year, this rapid
change in
NDVI
may
be thought
of as a precursor
to spring
greenup.
However, for spring wheat, the onset of greenness is
coincident with snowmelt and the metric serves its
purpose
well. Similarly,
the end-of-greenness
metric is
related
to the most rapid
decline in NDVI and in some
cases may be related
to abiotic
factors,
such as extended
periods of cloudiness, which do not have direct
phenological
significance.
The smoothing algorithm,
applied to minimize the
effects of cloud contamination,
is an area that may
potentially
be improved.
The running
median method
employed in this study is effective at minimizing the
effects of extremely
low NDVI values that
are related
to
)NDJ FMAI
TIME
712
- Measuring phenological variability from satellite imagery -
cloud contamination,
but has the undesired effect of
lowering
high NDVI values, which are
presumed
to be
valid. While the consequences of the smoothing are
lessened by returning
to the raw data
to compute
many
of the metrics, other measures
such as time-integrated
NDVI are still adversely affected (although
not as se-
verely as if raw data were used). Future
investigations
should address the use of a different smoothing algo-
rithm, such as the best index slope extraction
(BISE),
which retains
high
values and eliminates
only extremely
low values (Vivoy et al. 1992).
The definition of time-integrated
NDVI also has
shortcomings
in regions
that have
year-round photosyn-
thetic activity. This metric is measured from the base-
line of the onset and end of greenness
NDVI values. In
evergreen communities, this baseline value is higher
than
in any of the other land cover types because these
communities
are
actively photosynthesizing throughout
the year when temperatures
are above 0 ?C. A method
for incorporating
the vegetative activity
taking
place at
NDVI levels less than this baseline should be devel-
oped. This example illustrates that some metrics have
more applicability
for selected cover types. The time-
integrated
value, for example,
is not as useful for conif-
erous
forests as for agricultural crops,
but other
metrics,
such as the low rates
of greenup
and senescence, effec-
tively characterize
the coniferous forests.
Conclusions
The
results obtained
in this study,
while preliminary,
illustrate
techniques that can be used to evaluate the
variability
or stability of the phenology of land cover
types. Derived measures of phenological variability,
such
as the median absolute
deviation
about
the median,
have potential
value for ecosystem monitoring.
An in-
crease
in within-class
phenological
variability
may sig-
nal a change of land cover in at least a portion
of that
class. In a regional
or global
monitoring system, such an
increase in variability
may serve as a signal to obtain
higher
resolution
imagery
(such
as Landsat
TM
or
SPOT)
over this region and perform
more detailed
land cover
analysis.
The results also have implications for land cover
mapping,
suggesting that
remotely sensed data are ap-
propriate as input to vegetation mapping, but more
stable information contained in ancillary data should
also be incorporated, particularly
in areas with high
phenological variability
such as grasslands
and shrub-
lands. The variability
shown
in this study
argues
against
the idea of signature
libraries,
whether
they be spectral,
temporal,
or a combination of spectral
and temporal.
This study represents
a first generation
of pheno-
logical feature
extraction from NDVI temporal profiles.
Further studies may incorporate
climate data, experi-
ment with different smoothing algorithms, and use a
longer
time series of satellite data as they become avail-
able. While the seasonal measures
derived
from
satellite
data
appear
to be valid and have several
potential
appli-
cations, detailed ground-based
documentation of the
measures needs to be undertaken.
Acknowledgements.* Work
conducted
by Hughes STX staff
was performed
under U.S. Geological Survey contract 1434-
92-C-40004. Partial
support
for Ms. Brown was provided
by
the University of Nebraska-Lincoln and funded under
U.S. Environmental
Protection
Agency grant
(X0075626-01).
Partial
support
for Mr.
VanderZee was provided by the United
Nations Environment
Programme's
Global Resources Infor-
mation
Data
base (UNEP-GRID)
project through
a Memoran-
dum of Understanding
between the U.S. Geological Survey,
the National Aeronautics and
Space
Administration,
and
UNEP.
References
Anon. 1990. Global Vegetation Index Users' Guide. U.S.
Department
of Commerce,
National Oceanic and
Atmos-
pheric Administration,
Satellite Data Services Division,
Washington,
DC.
Allen, T.R., Bara,
T.J. & Walsh,
S.J. In
press.
Observed Biases
in the Conterminous U.S. AVHRR Satellite Data Set: A
North Carolina
Case Study. GeoCarto Int.
Badhwar, G.D., Cames J.G. & Austin, W.W. 1982. Use of
Landsat derived temporal profiles for corn-soybean
fea-
ture extraction and classification. Remote Sens. Environ.
12: 57-79.
Barnes, P.W., Tieszen, L.L. & Ode, D.J. 1983. Distribution,
production,
and
diversity
of C3- and C4-dominated
com-
munities
in a mixed prairie.
Can. J. Bot. 61: 741-751.
Becker,
R.A., Chambers,
J.M.
& Wilks, A.R. 1988. The New S
Language.
Wadsworth and Brooks, Pacific Grove, CA.
Brown, J.F., Loveland, T.R., Merchant,
J.W., Reed, B.C. &
Ohlen, D.O. 1993. Using Multisource Data in Global
Land-Cover Characterization:
Concepts, Requirements,
and Methods. Photogramm.
Eng. Remote Sens. 59: 977-
987.
Changnon,
S.A. & Kunkel,
K.E. 1992. Assessing Impacts
of a
Climatologically Unique Year (1990) in the Midwest.
Phys. Geogr. 13: 180-190.
Curran,
P.J. 1981. Multispectral
remote
sensing
for estimating
biomass and productivity.
In: Smith, H. (ed.) Plants and
the Daylight Spectrum,
pp. 65-99. Academic Press, Lon-
don.
Eidenshink,
J. 1992. The 1990 Conterminous U.S. AVHRR
Data Set. Photogramm. Eng. Remote Sens. 58: 809-813.
* Any use of trade,
product,
or firm names in this publication
is for descriptive purposes only and does not imply endorse-
ment by the U.S. Government.
713
Reed, B.C. et al.
Flint,
H.L. 1974. Phenology
and
genecology of woody plants.
In: Lieth, H. (ed.) Phenology and Seasonality Modeling,
pp. 83-97. Springer-Verlag,
New York,
NY.
Goward,
S.N. 1989. Satellite
bioclimatology.
J. Clim.
2: 710-
720.
Goward,
S.N., Dye, D., Kerber,
A. & Kalb,
V. 1987.
Compari-
son of North and South American
biomes from AVHRR
observations. GeoCarto
Int. 1: 27-39.
Goward, S.N., Tucker,
C.J. & Dye, D.G. 1985. North Ameri-
can vegetation patterns
observed with the NOAA-7 Ad-
vanced Very High Resolution
Radiometer.
Vegetatio
64:
3-14.
Granger,
C.W.J. 1989. Forecasting in Business and Econom-
ics, 2nd ed. Academic Press, Inc., Boston, MA.
Hoff, J.C. 1983. A Practical Guide to Box-Jenkins
Forecast-
ing. Lifetime Learning
Publications,
Belmont, CA.
Huber,
P.J. 1981. Robust
Statistics.Wiley,
New York, NY.
Justice, C.O., Townshend, J.R.G., Holben, B.N. & Tucker,
C.J. 1985. Analysis of the phenology
of global vegetation
using meteorological
satellite data.
Int. J. Remote Sens. 6:
1271-1318.
Kidwell, K.B. 1991. NOAA
Polar Orbiter
Data Users' Guide.
National Oceanic
and
Atmospheric
Administration,
Satel-
lite Data Services Division, Washington,
DC.
Lewis, J.K. 1971. The grassland
biome: A synthesis
of struc-
ture and function, 1970. In: French,
N.R. (ed.) Prelimi-
nary Analysis of Structure
and Function in Grasslands.
Range Science Series 10, pp. 317-387. Colorado State
University,
Ft. Collins, CO.
Lieth, H. 1974. Purposes
of a Phenology Book. In: Lieth, H.
(ed.) Phenology and Seasonality Modeling, pp. 3-19.
Springer-Verlag,
New York, NY.
Lillesand, T.M. & Kiefer, R.W. 1987. Remote Sensing and
Image Interpretation,
2nd ed. Wiley, New York, NY.
Lloyd, D. 1990. A phenological classification of terrestrial
vegetation cover using shortwave vegetation index im-
agery.
Int. J. Remote
Sens. 11:
2269-2279.
Lo, T.H.C., Scarpace, F.L. & Lillesand, T.M. 1986. Use of
Multitemporal Spectral
Profiles
in Agricultural
Land-Cover
Classification. Photogramm. Eng. Remote Sens. 52:
535-544.
Loveland,
T.R., Merchant, J.W., Ohlen, D.O. & Brown, J.F.
1991. Development
of a Land-Cover Characteristics
Data
Base for the Conterminous U.S. Photogramm.
Eng. Re-
mote Sens. 57: 1453-1463.
Loveland, T.R., Merchant,
J.W., Ohlen, D.O., Brown, J.F.,
Reed, B.C. & Steyaert, L.T. 1993. Conterminous
U.S.
Land Cover Characteristics
Data Set: 1990 Prototype.
EROS Data Center, U.S. Geological Survey CD-ROM
[includes
Readme
file.], Sioux Falls, SD.
Malingreau,
J.P. 1986. Global vegetation
in dynamics,
satel-
lite observations over Asia. Int. J. Remote Sens. 7: 1121-
1146.
Malingreau,
J.P. 1989. The vegetation
index and
the study
of
vegetation dynamics. In: Toselli, F. (ed.) Application
of
Remote
Sensing
to
Agrometeorology,
pp. 285-303. ECSC,
Brussels and
Luxembourg.
Maxwell, J.R., Sr. 1976. Commodity
Futures Trading with
Moving
Averages. Speer
Books, Red Bluff, CA.
Morain,
S.A. 1974. Phenology
and Remote
Sensing.
In:
Lieth,
H. (ed.) Phenology and Seasonality Modeling,
pp. 55-75.
Springer-Verlag,
New York,
NY.
Odenweller,
J.B. & Johnson, K.I. 1984. Crop Identification
Using Landsat
Temporal-Spectral
Profiles. Remote Sens.
Environ. 14: 39-54.
Rauzi,
F. & Dobrenz,
A.K. 1970. Seasonal variation
of chloro-
phyll in western wheatgrass
and blue grama. J. Range
Manage. 23: 372-373.
Rouse, J.W., Haas, R.H., Schnell,
J.A. & Deering,
D.W. 1973.
Monitoring
the vernal advancement
and retrogradation
(green wave effect) of natural vegetation. NASA/GSFC
Type II Progress Report.
Greenbelt,
MD.
Townshend,
J.R.G., Justice, C.O. & Kalb, V. 1987. Charac-
terization and classification
of South
American land cover
types using satellite data.
Int.
J. Remote
Sens. 8: 127-150.
Tucker, C.J., Townshend,
J.R.G.
& Goff, T.E. 1985. African
land-cover
classification
using satellite
data. Science 227:
369-257.
Tukey,
J.W. 1977.
Exploratory
Data
Analysis. Addison-Wesley
Publishing
Co., Reading,
MA.
Vivoy, N., Arino, 0. & Belward,
A.S. 1992. The Best Index
Slope Extraction:
A method for reducing
noise in NDVI
time series. Int. J. Remote
Sens. 13: 1585-1590.
Received 20 November 1993;
Revision received 21 April 1994;
Accepted 22 August 1994.
714
