Investigating SeaWinds Terrestrial Backscatter: Equatorial

02-111.qxd 10/8/03 12:22 PM Page 1243
Investigating SeaWinds Terrestrial Backscatter:
Equatorial Savannas of South America
Perry J. Hardin and Mark W. Jackson
Abstract
Because tropical grasslands play an important role in the
storage of global carbon, monitoring them is critical to
evaluating global climate change. The goal of this research
is to model seasonal SeaWinds Ku-band backscatter in five
savanna areas of Colombia, Venezuela, and Brazil as a function of biophysical changes in the savanna landscape. Multiple regression modeling demonstrates that savanna Ku-band
backscatter is a function of (1) savanna grass biomass/leaf
area, (2) soil moisture, and (3) other soil characteristics. Fit
for the regression models is excellent (R 0.87 and 0.81,
respectively, for the horizontal and vertical polarization case).
The horizontal—vertical polarization difference is also moderately related to precipitation (R 0.71). The results from
this modeling are consistent with theory predicated on previous C- and X-band research. The possibility of monitoring
savanna vegetation, soil moisture, and rainfall using Ku-band
radar and scatterometry is discussed.
Introduction
Grasslands
Midlatitude and tropical grasslands cover less than 10 percent
of the world’s land surface (Atjay et al., 1979) but play a critical role in global carbon storage. The burning of tropical savannas alone contributes approximately 1660 teragrams (Tg)
of carbon per year to the atmosphere in the form of CO2. In
contrast, tropical forest burning contributes 570 Tg. Of the
8700 Tg of global dry matter burned per year, tropical savannas account for 42 percent (Andrea, 1991; Levine et al., 1999).
An understanding of tropical grasslands is thus important to
comprehending the global CO2 balance.
The monitoring of large regional grasslands from space
has primarily been conducted using the Advanced Very High
Resolution Radiometer (AVHRR) and its derivative vegetation
index products. The large pixel size provided by the AVHRR
has been valuable for the monitoring of regional grassland extents while not producing huge data volumes typical of higher
resolution sensors such as Landsat, SPOT, and MODIS. While
the AVHRR has been a patently successful instrument, its optical data acquisition is hampered by the presence of darkness,
clouds, and smoke. These limits constrain its year-round use
for monitoring grasslands in critical grassland ecosystems
such as the boreal tundra and tropical savanna. In contrast,
active spaceborne instruments such as scatterometers are not
limited by these constraints.
Terrestrial Scatterometry
Spaceborne scatterometers are active microwave instruments
originally designed to measure ocean wind vectors. However,
scatterometers have been successfully used for regional
studies of land and sea ice as well. Encouraged by early successes with the Seasat-A scatterometer made by Kennett and
Department of Geography, Brigham Young University,
Provo, UT 84602 ([email protected]; [email protected]).
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Li (1989), studies of land cover predicated on scatterometry
have flourished. While much has been done over boreal regions (e.g., Wismann, 2000), terrestrial scatterometry has been
used to investigate tropical forests (Hardin and Long, 1994;
Woodhouse et al., 1999), as well as soil moisture (Magagi and
Kerr, 1997; Wagner et al., 1999a; Wagner and Scipal, 2000;
Woodhouse and Hoekman, 2000) and vegetation in arid and
semi-arid regions (Wagner et al., 1999b). A complete review
of terrestrial scatterometry is beyond the scope of this paper,
but is available from the Institute for Applied Remote Sensing,
2002.
Because their primary mission is to measure ocean winds,
scatterometers over land are limited by the low spatial resolution (e.g., greater than 25 km2) of the scatterometer instrument
itself. To alleviate this deficiency, research began in 1993 to
create enhanced-resolution regional radar backscatter (o)
images of the world’s vegetation using reconstructed scatterometer imagery from the Seasat-A scatterometer (Long
et al., 1993). The continuing goal of this research is to produce
land-cover scatterometer data at resolutions similar to AVHRR
global area coverage. This goal is achieved by compositing
low-resolution single-pass scatterometer imagery acquired
over a temporal span using a model to account for incidenceangle backscatter dependence in the process. The result of the
multi-pass reconstruction is a series of time-slice o images,
termed “A” images. In simple terms, A images are enhancedresolution weighted averages of local o that are normalized to
a single incidence angle (Long et al., 1993). Like the scatterometer data from whence they are derived, the A image pixels
measure backscatter on a decibel (dB) scale. Horizontal and
vertical A images are designated Ahh and Avv ,1 respectively.
Hypothesis and Objectives
Can SeaWinds backscatter measurements of five savanna2
study sites in Venezuela, Colombia, and northern Brazil be
statistically modeled as a function of landscape biophysical
parameters? This research will demonstrate that both
1
Hereafter, hh and vv are used to represent the two SeaWinds
polarization modes. The letters vv represent energy that is
vertically polarized both at transmission and upon reception.
The letters hh represent the horizontal equivalent.
2
To simplify the presentation, we will refer to the grassland
study sites in this research as “the savanna” with an implicit
recognition that other equatorial, tropical, and mid-latitude
grasslands with far different vegetative and climatic characteristics enjoy the designation too. We do not suggest that
these results apply universally to all these savannas.
Photogrammetric Engineering & Remote Sensing
Vol. 69, No. 11, November 2003, pp. 1243–1254.
0099-1112/03/6911–1243/$3.00/0
© 2003 American Society for Photogrammetry
and Remote Sensing
November 2003
1243
02-111.qxd 10/8/03 12:22 PM Page 1244
horizontal and vertical backscatter measurements contained
in reconstructed SeaWinds data can be successfully modeled
as a function of vegetation, soil, and rainfall parameters.
Within this hypothetical context, this research has
two primary objectives. The first objective is to quantify the
seasonal backscatter change of five savanna study sites of
northern South America using reconstructed SeaWinds A
imagery. The product of this quantification is a seasonal
backscatter signature for each savanna site. The second objective is to interpret seasonal change in both horizontal (hh) and
vertical (vv) backscatter as a function of biophysical change in
the landscape through the 12 calendar months in 2000 studied. This is done largely by correlation and multiple regression analysis using other collected data as independent variables. Convergence of evidence, based on (1) the strength of
statistical analysis and (2) backscatter theory, are offered as
support for the hypothesis.
This research builds on previous work and extends it in
significant ways.
•
•
•
It demonstrates that grassland backscatter theory developed
primarily from empirical C- and X-band research over small
midlatitude study plots can be successfully applied to explain
large-region Ku-band grassland backscatter of the tropical savanna. The addition of a regional grassland study at Ku-band
is unique.
Noting that grassland theory and empirical results have sometimes disagreed, this research provides insight into the role
played by biophysical parameters in explaining Ku-band
backscatter over large resolution cells. The role of moisture in
explaining polarization dependence in Ku-band backscatter is
a particularly important contribution.
This research validates the use of SeaWinds as an instrument
for savanna study and implies that a spaceborne Ku-band
instrument of higher resolution would be a valuable tool for
grassland monitoring. Given the importance of the savanna to
the global CO2 balance, the theoretical potential to monitor the
savannas at Ku-band is significant. In particular, the results
of this research indicate that rainfall, biomass, and burning
can be monitored in the savannas using Ku-band data as an
adjunct to optical and infrared sensors.
Conceptual Framework
Backscatter Mechanisms
Early landmark research into radar backscatter of grasslands
was conducted in the tall-grass Konza Prairie in northeastern
Kansas. As reported by Zoughi et al. (1987), this X- and
C-band research was designed to determine the mechanisms
of grassland backscatter with emphasis on differentiating between burned and unburned sites. The authors concluded that
(1) the dominant high-frequency vv backscatter mechanism
for tall, dense prairie grass is the upper portion of the grass
blade itself; (2) the lower portion of the canopy, including the
thatch and soil, dominates hh backscatter; (3) the overall difference in o between hh and vv at a 50° incidence angle (i.a.)3
is about 1 dB (with vv greater than hh) regardless of whether
the site is burned; and (4) canopy attenuation of vv polarized
energy is nearly twice that of hh.
This early research was followed by efforts to model the
Konza Prairie empirical backscatter data as a function of the
grass and soil biophysical properties. These properties included grass permittivity, stalk orientation, stalk cross-section
radius, grass density, and canopy height (Bakhtiari and
Zoughi, 1991). The most significant mismatch between the
model results and empirical data was the model’s overestimate of the X-band difference between vv and hh backscatter
3
The SeaWinds data used in this research are standardized to
50° i.a. This explains the particular focus on 50° i.a. within
the conceptual framework.
1244
November 2003
amplitude. While the model predicted 8 dB more vv backscatter than hh backscatter at 50° i.a., the difference measured
in the field was only 1 dB (Bakhtiari and Zoughi, 1991).
Research also conducted on the Konza Prairie by Martin et al.
(1989) determined that C-band backscatter over grassland is
primarily a function of soil moisture, with only low correlation with biophysical properties of the grass itself (e.g., leaf
water potential).
In more recent research conducted by Hill et al. (1999),
multiple regression equations designed to relate dry matter
yield and biomass to C-, L- and P-band o over temperate
grasslands in New South Wales, Australia produced poor fits.
Higher correlations were obtained with linear equations modeling herbage height parameters (e.g., mean grass height, standard deviation of grass height) as a function of backscatter.
Recent theoretical work validated by empirical research supports the New South Wales study postulating a strong relationship between grass height and o. In studies conducted
using an Ann Arbor, Michigan wheat field as a grassland
surrogate, backscatter at L-, C-, and X-band were related to
(1) soil, leaf, and stalk moisture, (2) leaf height, (3) leaf count,
(4) stalk diameter average height, and (5) stalk standard deviation height. The resulting C-band models demonstrated that
scattering at smaller incidence angles is primarily a result of
leaf and stalk “‘ground-bounce’ while the direct scattering
component remains small. As the incidence angle increases,
the electric field of the vertically polarized incident waves
increasingly couples into the vertical structure of the wheat
plant.” At 50° i.a., ground bounce from stalk is attenuated
and becomes insignificant whereas backscatter increases
from the upper portion of the stalk and grain head (Stiles and
Sarabandi, 2000; Stiles and Ulaby, 2000).
In contrast to the C-band results, the X-band model in the
same study indicates that the dominant backscatter source for
both X-band vv and hh energy is direct bounce from the leaf
elements themselves; direct bounce from grass stalks only
becomes a significant source of vv backscatter above 40° i.a.
Furthermore, where ground bounce from both leaf and stalk is
not completely attenuated, its effect will be greater at hh than
vv polarization. Despite the polarization-dependent difference
in backscatter mechanisms, the model ultimately demonstrated that X-band backscatter at 50° i.a. would be nearly
equal at both hh and vv because the backscatter mechanisms
are mutually compensating (Stiles and Sarabandi, 2000; Stiles
and Ulaby, 2000). This equality between vv and hh backscatter at 50° i.a. diverges from another X-band model cited earlier indicating a vv backscatter greater than hh (Bakhtiari and
Zoughi, 1991) but agrees with field measurements (Zoughi
et al., 1987).
In summary, it is generally thought that grassland mechanisms for short-wavelength, high incidence angle vv and hh
backscatter are somewhat different, although the sum of the
mechanisms creates equivalent backscatter amplitude. At 50°
i.a., vertically polarized X-band radar energy responds primarily to the upper portion of the canopy and the vertical orientation of the grass stalks themselves, whereas horizontally polarized backscatter responds to the lower portion of the canopy
and the horizontally oriented leaves. However, the sum of
backscatter from different mechanisms creates comparable values of o at 50° i.a.
There is less agreement concerning the precise role of
necromass and soil in grassland scattering at high frequencies.
In one model, horizontally oriented moist necromass lying on
the ground (as thatch) is considered an attenuating layer
primarily for hh energy, acting as a cloud of independent
water droplets. As wetness increases, so does attenuation of
the energy striking the soil (Saatchi et al., 1994). In other
models, the effect of thatch is either ignored or incorporated
as a permittivity and soil roughness adjustment to a soil
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
02-111.qxd 10/8/03 12:22 PM Page 1245
scattering submodel or constraint (Bakhtiari and Zoughi,
1991, Stiles and Sarabandi, 2000; Stiles and Ulaby, 2000).
Seasonal Change in Backscatter
Both the seasonal change in the savanna landscape as well as
morphological change in the savanna grass itself suggests that
savanna backscatter would change throughout the season.
This change applies to both backscatter magnitude and causal
mechanisms.
As discussed by Olson and Lacey (1996), the most obvious morphological change which takes place in grass itself
during the growing season is increasing vertical plant height,
terminating with seedheads on plants with reproductive
stems. Less obvious is the growth of shoots from buds (tillering) which increases (1) leaf and stem densities within the
bunch and (2) leaf area over the landscape. As grass increases
in height, the vertical orientation of stems becomes the predominant morphological characteristic of the grass tussock
and primary reservoir of soil moisture. Given these growth
patterns, seasonal grass growth increases biomass, grass
height, landscape surficial permittivity, and vertical orientation of the dominant landscape elements (i.e. the grass
stalks).
An early agricultural study conducted at Ku-band by
Ulaby et al. (1975) implies that morphological changes in leaf
facet orientation caused by heavy rain also affect backscatter
magnitude from grass at high frequencies. Personal observation of grassland indicates that heavy rain encourages lodging,
moving grass stems toward a more horizontal orientation than
the prototypical vertical orientation associated with grass. In
assessing the relative importance of these changes to other
variable plant factors such as stalk moisture, researchers have
observed that, at high radar frequencies, this emphasis on
morphological change does not “imply that plant moisture is
not an important parameter in the backscattering process; it
simply indicates that other factors such as morphological
changes, which are influenced by plant moisture and growth
changes, are more critical” to backscatter. As frequency
increases between 8 GHz and 18 GHz, this dependence on
morphology increases proportionally (Ulaby et al., 1975;
Ulaby et al., 1978).
Where bare ground is extensive within sparser savanna
grassland, backscatter will modulate predictably as surficial
soil moisture increases and decreases in phase with precipitation seasonal patterns (Ulaby et al., 1978; Ulaby et al., 1979;
Dobson and Ulaby, 1981). However, at larger incidence angles,
the effect of soil moisture will decrease as vegetation increases to full canopy (Ulaby et al., 1975). As suggested by
studies cited previously, the effect of soil moisture on backscatter magnitude is dependent on polarization—the magnitude
of the copolarized ratio (hh/vv) at both X-band is directly proportional to soil moisture (Oh et al., 1992). However, see
Schmullius and Furrer (1992) for a counterexample. In contrast to surficial soil moisture, increases in seasonal open
water at X-band would probably decrease o over the larger
pixels typical of scatterometers. Backscatter from intermediate
conditions between moist soil and open water (e.g., ponded
water beneath a seasonally flooded grassland canopy) would
be dependent on grass height and density, but may create
high values of o (Henderson, 1995; Wismann and Boehnke,
1996).
Synthesis
The reviewed literature above leads to a preliminary framework for interpreting high-frequency radar imagery of the
savannas of northern South America. Given the large pixel
size as well as the paucity of empirical data available for
Ku-band, the framework is necessarily qualitative, statistical,
and tentative until future physical modeling can be
completed.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
•
•
•
•
As dried necromass with minimal permittivity, senescent
grass vegetation will be transparent at Ku-band frequencies.
Continuing dry-season decreases in backscatter will primarily
be a function of progressive soil desiccation. Therefore,
Ku-band backscatter values in the savanna will reach their
nadir at the terminus of the dry season. Because polarization
differences are a result of soil moisture, polarization dependence at this time will be virtually nonexistent.
Early-season increases in soil moisture responding to the
onset of regional rainfall will cause an increase in backscatter
over the savannas. As new growth is initiated and grows
taller, the contribution of soil moisture to overall backscatter
will decrease. However, total backscatter will continue to increase in response to increased canopy height and density.
The relative magnitude between hh and vv backscatter will be
a function of soil moisture, with hh backscatter exceeding vv.
The relationship between soil moisture and backscatter will
be more pronounced as bare ground of the site increases.
Flooded grasslands will exhibit a backscatter dependent on
moisture content, standing water extent, and grass height.
Slightly ponded water will have an effect similar to high
soil moisture. Increasing inundation of grass will result in
decreased backscatter.
Data and Methods
The fundamental methodological approach of this research
was to (1) identify large homogeneous areas of grassland
savanna in northern South America, (2) determine the seasonal
backscatter signatures for the study regions (for year 2000)
using reconstructed SeaWinds A imagery, and (3) interpret the
backscatter changes in the context of existing backscatter
theory and by reference to other data sources. Linear models
explaining backscatter as a function of other biophysical variables were created as an intermediate step to interpretation.
Study Site Selection
A single definition of “savanna” is not universally accepted,
and is best defined geographically. For this study, the term
savanna includes the Guyanan grasslands of northern Brazil
and well as the grassy llanos of Colombia and Venezuela. This
savanna is classified as intact grasslands on the Map of the
Vegetation of South America Based on Satellite Imagery produced by Stone et al. (1994) and is distinguished from shrubland, scrubland savanna, or savanna with coppices of palms
or other tree species. These grasslands are also distinguished
from the subhumid savannas of Argentina and Uruguay
frequently termed Pampas.
The primary task in successful site selection was extracting the largest possible homogeneous savanna regions uncontaminated by other classes such as degraded grasslands and
agriculture. This task was accomplished by a combination of
region growing, filtering, and manual intervention using inhouse software to create homogeneous regions larger than
the SeaWinds pixel resolution. The map was the only data
source used for region delineation. Neither the NDVI data nor
SeaWinds data (discussed below) were referenced. The result
of this processing was the five study sites depicted in Figure 1.
As shown in Table 1, the sites range from 178,700 km2 to
533,900 km2.
The Pacaraima region is representative of the Guyana
savanna. Like the other four sites, the Pacaraima area is
dominantly a Trachypogon savanna with Paspalum and
Andropogon grasses codominant. The precise mix of these
three genera is largely a function of topography, edaphic
factors, and seasonal flooding. Among the five sites, the
Pacaraima site is unique due its high rainfall and dominance
of severely weathered soils (Baumgardner, 1986).4 The Apure,
4
All soil nomenclature follows the modern FAO/UNESCO
designation accompanying the FAO global soils database
(Baumgardner, 1986).
November 2003
1245
02-111.qxd 10/8/03 12:22 PM Page 1246
TABLE 1. SUMMARY CHARACTERISTICS OF THE FIVE SAVANNA STUDY SITES
Site
Area
(km2)
Center
Latitude/Longitude
Elevation (M.S.L.)
2000 Total
Precipitation
(mm)
Meta
262800
N 4.7° W 71.0°
140 m
2104
Apure
533900
1476
Guarico
178700
N 7.9° W 69.1°
80 m
N 8.3° W 66.3°
90 m
Viuda
217900
N 8.8° W 63.7°
130 m
1157
Pacaraima
205000
N 3.7° W 60.6°
230 m
2400
Figure 1. Location of the study sites within South America.
Guarico, and Viuda sites form the classic Venezuelan llanos of
the Orinoquia dominated by Trachypogon plumosus and
Trachypogon vestitus with Paspalum fasciculatum typical of
seasonally flooded regions. As shown in Table 1, these three
sites don’t differ substantially from each other except in regard to their soils. While all three of these Venezuelan llanos
sites have soils typical of extreme tropical weathering, nearly
70 percent of the Guarico site soils show evidence of poor
drainage and extended flooding during the year, whereas the
Viuda site, albeit weathered, possesses better drainage and
more intermittent flooding during the wet season.5 The Meta
site is part of the Llanos Orientales of Colombia and is generally well-drained but intermittently inundable during the wet
season. A small proportion of this study region contains areas
with lengthy periods of inundation.
SeaWinds A Imagery
A discussion of Scatterometer A image reconstruction is
beyond the scope of this paper, but can be found elsewhere
(Long et al., 1993; Early and Long, 2001). The A imagery used
for this study was NASA scatterometer pathfinder climate
5
The inferences regarding soils and inundation were visually
confirmed by viewing Landsat imagery acquired throughout
the 2000 season. This Landsat imagery was an important
adjunct to interpretation, although it was not a quantitative
data source for the research.
1246
November 2003
1619
Dominant Soils
69% Ferralsols
16% Arenosols
15% Gleysols
48% Gleysols
31% Luvisols
52% Gleysols
28% Acrisols
20% Vertisols
42% Arenosols
25% Acrisols
14% Ferralsols
49% Ferralsols
26% Plinthsols
25% Regosols
record data (NASA/BYU, 2002). This electronically available
SeaWinds A imagery is differentiated by (1) reconstruction
algorithm (egg versus slice), (2) polarization (hh versus vv),
and (3) orbit node (ascending versus descending). For this
research, imagery produced by the egg algorithm was selected
for analysis because of its lower image noise than corresponding slice algorithm images. The egg algorithm models the
SeaWinds measurement footprint as an ellipse whereas the
slice algorithm treats the footprint as a set of smaller cells.
Each approach has advantages and disadvantages. Descending
and ascending passes were combined, whereas hh and vv
imagery were analyzed separately.
The SeaWinds pathfinder images for 2000 were primarily
four-day composites. In-house software was used to combine
the four-day composites into monthly images using simple
pixel averaging over the time periods represented by the
pathfinder image source. The result of the monthly averaging
was 24 monthly A images; 12 vv images and 12 hh images
were produced. The mean backscatter values from these
vertical and horizontal A images are designated A
vv and Ahh ,
respectively.
Figure 2 shows the change in A
hh throughout 2000 for
northern South America. While the tropical forest remains
relatively invariant throughout the season, the savanna study
regions show substantial seasonal change. The decrease in
backscatter between the wet season (April, May, June) and
the progressive dry season (October, November, December)
is manifest visually in Figure 2 by an increase in dark
blotchiness.
The difference in incidence angle between SeaWinds vertical and horizontal data was a minor impediment to the data
analysis. SeaWinds on QuikScat is a dual-pencil-beam conically scanning scatterometer operating at 13.4 GHz over a
1600-km swath. The outer beam is vertically polarized whereas
the inner beam is horizontally polarized. Unlike SASS and
NSCAT, SeaWinds is primarily a single-incidence-angle instrument with nominal incidence angles of 46° and 54° for horizontal and vertical beams, respectively. If left uncorrected, the
incidence angle difference between the two beams would
complicate the interpretation of grassland targets showing significant incidence angle dependence in o at high frequencies.
The disparity would also complicate the calculation of typical
copolarization indexes such as ratios (A
hh Avv ). Given this
complication, the A
hh and Avv data were standardized to an
incidence angle of 50° by utilizing incidence angle dependence data produced in the reconstruction of NSCAT imagery
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
02-111.qxd 10/8/03 12:22 PM Page 1247
Figure 2. Monthly A images of northern South America. While the tropical forest remains relatively invariant through the year,
the savanna areas show substantial change.
collected in 1996 and 1997. The standardization process is
briefly summarized in Appendix A.
NDVI Data
In order to model the seasonal backscatter patterns apparent
in the SeaWinds A imagery, monthly pathfinder AVHRR land
data sets for the year 2000 were acquired (NOAA/NASA,
2002), resampled, and reprojected to overlay properly on the
A imagery. The NDVI pixels covering the five study regions
were extracted and statistically summarized by month. The
monthly mean NDVI for each region is represented symboli. Because NDVI has demonstrated its utility as a
cally by VI
surrogate for biomass, leaf area, and photosynthetically active
radiation for some vegetation classes, a correlation between
and A for the five savanna targets was expected. The
VI
relationship between NDVI and biophysical parameters is
discussed below.
Climate and Soils Data
Given that climatic factors were expected to play an important
role in the interpretation of seasonal SeaWinds backscatter
patterns, basic rainfall and temperature data were gathered for
the five study regions. The primary source for these data was
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
the National Climate Data Center Global Surface Summary of
Day (NCDC, 2002). This dataset incorporated daily weather
statistics from approximately 8000 stations worldwide participating in the World Meteorological Organization (WMO) World
Weather Watch Program. Of the 13 variables included for each
daily station weather record, only four were used for this research: (1) maximum daily temperature, (2) minimum daily
temperature, (3) mean daily temperature, and (4) total daily
precipitation (PT).
Of the stations listed for South America, one station was
selected as most representative for each of the study regions.
Several criteria had to be met for a meteorological station to
be accepted as representative. First, the elevation difference
between region center and meteorological station could not
exceed 100 m. Second, no topographic obstruction above
300 m could intervene between the station and the region.
Because of their unique climatology, coastal stations were also
excluded. Completeness of record (for the year 2000) and distance from the region centroid were also obvious criteria for
selection.
From the daily weather data, monthly temperature means
and total precipitation were calculated for each of the five
November 2003
1247
02-111.qxd 10/8/03 12:22 PM Page 1248
TABLE 2. SEASONAL EXTREMES AMONG THE MODEL VARIABLES
Site
Meta
Minimum
Maximum
Apure
Minimum
Maximum
Guarico
Minimum
Maximum
Viuda
Minimum
Maximum
Pacaraima
Minimum
Maximum
PT
Average
Temperature
17 mm/Apr
406 mm/Jul
23 C/Jun
28 C/Sep
0.29/Mar
0.56/Sep
12.9 dB/Feb
11.5 dB/Jun
12.8 dB/Feb
10.4 dB/Jun
0.1 dB/Feb
1.0 dB/Jun
7 mm/Jan
278 mm/Jun
25 C/Feb
28 C/Jan
0.29/Mar
0.74/Nov
11.6 dB/Mar
9.6 dB/Aug
11.1 dB/Mar
8.7 dB/Aug
0.3 dB/Jan
1.0 dB/Aug
1 mm/Jan
592 mm/Jul
21 C/Feb
25 C/May
0.26/Mar
0.56/Nov
13.2 dB/Feb
11.8 dB/Aug
12.7 dB/Apr
10.5 dB/Aug
0.5 dB/Mar
1.3 dB/Oct
1 mm/Mar
199 mm/Jun
25 C/Feb
29 C/Oct
0.34/Mar
0.62/Nov
12.6 dB/Apr
11.4 dB/Jul
12.1 dB/Apr
10.6 dB/Jul
0.6 dB/Apr
0.9 dB/Jul
51 mm/Dec
485 mm/May
27 C/May
29 C/Oct
0.31/Mar
0.52/Jun
13.5 dB/Mar
11.8 dB/Jun
12.9 dB/Mar
10.5 dB/Jun
0.6 dB/Mar
1.34 dB/Jul
VI
study regions. The availability of these summary statistics
allowed monthly backscatter values to be correlated with both
temperature and precipitation.
Most of the detailed vegetation information available for
the savanna study areas are found in written reports rather
than in digital or mapped form. Because of this lack of accessible mapping, soil data were used as a surrogate for broad
savanna vegetation community types. This did not pose a
problem because the linkage between savanna community
composition and soil association has long been delineated
(Blydenstein, 1967; Brown et al., 1987). The soils data originated with the 1:1,000,000-scale FAO Global and National
Soils and Terrain Digital Database (Baumgardner, 1986). After
some initial statistical analysis and comparison with the
Landsat imagery, two soil variables were selected for use in
the final models explaining SeaWinds backscatter: (1) percentage of study site covered by Gleysols (GP) and (2) percentage
of study site covered by Ferralsols (FP). These variables were
used as surrogates to represent the percentage of inundable/
wet savanna and humid/dry savanna, respectively. More
simply, they were also interpreted as poorly drained (Gleysols)
and well drained (Ferralsols) soils. As shown in Table 1,
these two soils were the most common among the five study
sites.
Statistical Analysis
The primary model building approach for this research was
multiple linear regression (Hair et al., 1998). The dependent
variables were A
hh , Avv , and D because these were the primary backscatter variables of interest. Independent variables
, PT, GP, FP, and the temperature variables. With
included VI
the exception of GP and FP, none of the independent variables
were significantly correlated with one another, so multicollinearity did not pose a problem in the regression analysis.
The goodness of the regression models was judged by the
simple correlation coefficient (R) and the standard error of
the estimate (S.E.E.) (Marascuilo and Levin, 1983; Hair et al.,
1998). The usual diagnostic tools associated with regression
were used when indicated.
Although multiple linear regression was used for model
building, partial correlation was employed to illuminate relationships among the variables. Partial correlation is a method
which measures the relationship between two variables on the
same scale used by R, while controlling for the covariance due
to a third exogenous variable (Marascuilo and Levin, 1983;
Hair et al., 1998). Partial correlation coefficients are designated as RPart in the results below.
1248
November 2003
A
vv
A
hh
D
Results
Seasonal Patterns
A strong seasonal pattern of precipitation is archetypical of
the five study regions. Deep precipitation minima generally
occur from January to March, followed rapidly by precipitation maxima in early summer. These extremes form an initial
context for observing SeaWinds backscatter through the savanna season. As shown in Table 2, the lowest study region
always occurred in March, and was remarkably uniform in
VI
value among the five sites (range 0.26 to 0.34). There was
during the year (range
much more difference in the highest VI
0.52 to 0.74) among the sites. For three of the regions, this
occurred in November. In the regions of
maximum VI
Pacaraima and Meta this maximum took place in June and
September, respectively, reflecting the southward to northward movement of the Intertropical Convergence Zone
through the wet season.
In every case, the minimum in A
hh , Avv , and D occurred
during the same month. This, in turn, took place within a
minimum. In contrast, with the exception of
month of the VI
Pacaraima, there was no obvious relationship between the
and month of maximum A
month of the maximum VI
hh and
A
,
except
to
say
that
the
highest
backscatter
occurred
some
vv
maximum. This observation is immonths previous to the VI
and SeaWinds
portant, since it is the first indicator that VI
backscatter responded differently to savanna landscape
changes throughout the season.
Figures 3a through 3e show the relationship between PT,
, A
VI
hh , and Avv for the five study regions. The graphs indicate an obvious seasonal relationship between backscatter and
the other two variables, but it is difficult to detect whether the
mechanism responsible for the seasonal change in the backscatter was rainfall or changes in the landscape vegetation, espe and PT varied seasonally too. The graphs also
cially when VI
indicate that backscatter reached its lowest point at the terminus of the dry season and reached its maximum earlier than
. For Meta and Pacaraima, VI
and backscatter were closely
VI
synchronized throughout the season. Viuda, Guarico, and
Apure show a different pattern. In these three regions, although
continued to increase throughout the wet season to reach a
VI
late season high, backscatter tended to decrease after an early
wet-season maximum.
In all the sites (Figures 3a through 3e) both vv and hh
curves showed similar seasonal patterns of change. Furthermore, some polarization dependence was apparent during the
wet season, where A
hh exceeded Avv more than other months.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
02-111.qxd 10/8/03 12:22 PM Page 1249
Figure 3. (a) Seasonal signatures for the Meta study site.
(b) Seasonal signatures for the Apure study site. (c) Seasonal
signatures for the Guarico study site. (d) Seasonal signatures
for the Viuda study site. (e) Seasonal signatures for the
Pacaraima study site.
As the dry season progressed, polarization dependence decreased and reached a minimum before the onset of the next
wet season.
Overall Models of Backscatter
Results from the regression model building illuminate the
relationships depicted graphically in Figures 3a through 3e
while including the soil factors GP and FP. The overall model
explaining backscatter as a function of the independents is
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
show in Table 3. Horizontal backscatter was best explained by
, PT, and GP. Neither FP nor the tema model incorporating VI
perature variables provided significant discriminatory power
, PT, and GP were accounted for. The
after the influence of VI
fit of this general model was surprisingly strong (R 0.87,
p 0.05) for such a large area.
Based on the observations regarding the relationship be made above, it is not surprising that
tween backscatter and VI
comparison of the standardized coefficients in Table 4 indicate
November 2003
1249
02-111.qxd 10/8/03 12:22 PM Page 1250
TABLE 3. OVERALL MODEL EXPLAINING hh AND vv BACKSCATTER AS A FUNCTION
OF NDVI, PRECIPITATION, AND SOIL VARIABLES. IN BOTH THE hh AND VV CASE,
THE SAMPLE SIZE WAS 60
A
hh
A
vv
Raw
Standardized
Raw
Standardized
Coefficient
Coefficient
Coefficient
Coefficient
VI
PT
GP
Constant (dB)
R
S.E.E. (dB)
7.28
2.04E03
1.22E02
15.15
0.87
0.54
0.71
0.25
0.26
7.08
†
1.12E02
15.54
0.81
0.59
0.75
†
0.26
† Variable not significant contributor in regression ( 0.05).
AND BACKSCATTER FOR
TABLE 4. PARTIAL CORRELATIONS (RPart) BETWEEN VI
THE FIVE STUDY AREAS. THE EFFECT OF MONTHLY RAINFALL HAS BEEN
REMOVED FROM THE CORRELATION COEFFICIENTS. ALL OF THE COEFFICIENTS
ARE SIGNIFICANT AT 0.05 (n 12)
Region
VI
A
hh
RPart
A
vv
RPart
Apure
Viuda
Meta
Guarico
Pacaraima
0.74
0.62
0.56
0.56
0.52
0.84
0.65
0.81
0.80
0.62
0.88
0.70
0.83
0.70
0.62
that A
hh was more highly related to VI than to either PT or GP
created a
per se. Whereas a standard deviation change in VI
0.71 standard deviation change in A
,
a
similar
effect on
hh
backscatter required a three standard deviation change in
either rainfall or the percentage of gley soils within the study
region. In the general A
hh model, VI was much more significant to explaining backscatter than either rainfall or soil. In
contrast to the model for horizontal backscatter, rainfall had
no significant effect in explaining A
vv once the effects of VI
and soils were accounted for.
A simple t-test designed to test equality of regression
coefficients between the hh and vv models (Marascuilo and
Levin, 1983) indicated that there was no statistically signifi coefficients, GP coefficients,
cant difference between the VI
and
A
and constants for the A
hh
vv models (p 0.50). The two
models represented the same relationship between backscat. Statistically speaking, any observed difference
ter, GP, and VI
between A
hh and Avv (A HH A VV D ) could be attributed to
(1) rainfall, (2) other unmeasured variables, and (3) measurement errors or noise in the data.
Polarization dependence during the wet season was noted
above in the seasonal backscatter curves (Figures 3a through
and rainfall is further illus3e). The relationship between D
trated in Figure 4. While a variety of curves fit the regional
data reasonably well, the quadratic equation of the form
0.51 0.0027 * PT 3.0E-06 * P T2 provided the best fit
D
with a correlation coefficient of R 0.71 (p 0.05). Simple
to precipitation created
region-by-region models relating D
much better fits (data not shown), indicating that any future
would be most successful when
attempt to predict PT from D
done for each site separately, or perhaps month-by-month. In
, linear relationships were
all these site-specific models of D
adequate to explain its relationship with PT.
Discussion
Perhaps the most obvious statement is that both hh and vv
backscatter over the region were most strongly related to bio. The
physical landscape characteristics also measured by VI
1250
November 2003
Figure 4. Relationship between rainfall and polarization
difference.
was most apparent
relationship between backscatter and VI
during the wet season, and weakened with the progression
of the dry season. Similar relationships between NDVI and
rainfall were found by Hess et al. (1996).
and
To further illustrate the relationship between VI
backscatter, partial correlation coefficients were calculated to
show the relationship between the two variables with the
effects of rainfall and soil removed for each of the five study
and
areas. As shown in Table 4, the relationship between VI
backscatter remained moderate to very strong. The relation was the strongest in the
ship between backscatter and VI
Apure site and weakest in the Pacaraima site—the two
,
sites with the highest and lowest maximum seasonal VI
respectively.
was correlated with backscatter, the precise
Although VI
cause of the correlation can only be inferred by convergence
of evidence. As demonstrated by Mougin et al. (1995) and Lo
Seen et al. (1995) among many others, grassland NDVI is primarily related to leaf area, photosynthetically active radiation,
live biomass, percentage of vegetation covered ground, and
net primary productivity. In a marked seasonal grassland like
tropical savannas, all of these factors are themselves highly
interrelated, making the inference regarding causality difficult.
Nevertheless, field data collected at the Estación Biológica de
Los Llanos (EBLL) in 1986 provide a useful point-of-evidence
to interpreting the backscatter curve of the Apure and Guarico
sites (San Jose and Montes, 2002).6 The field data were collected between the boundaries of the Guarico and Apure study
sites, and included a variety of information on biological variables for the trachypogon savanna such as above ground total
biomass and leaf area index.
The correlation between Guarico/Apure backscatter and
the EBLL field data was displayed visually (Figure 5) and
examined statistically to determine if a relationship between
the two could be inferred. As shown in Table 5, the fit between backscatter, biomass, and leaf area index was excellent
and somewhat better than the fit between the same biophysi.
cal parameters and VI
Despite the difference in years, the similarity between
the backscatter data and biophysical variables leads to the
6
The applicability of the 1986/1987 Estación Biológica data to
the year 2000 is discussed in Appendix B.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
02-111.qxd 10/8/03 12:22 PM Page 1251
Figure 5. Horizontal backscatter of Guarico and Pacaraima
sites compared to EBLL field data.
AND
,
TABLE 5. SIMPLE CORRELATION BETWEEN EBLL FIELD DATA, VI
BACKSCATTER. ALL OF THE COEFFICIENTS ARE SIGNIFICANT AT 0.01
(n 12)
Apure
Guarico
Variable
A
hh
A
vv
A
hh
A
vv
1986 EBLL Biomass
1986 EBLL Leaf area index
2000 VI
1986 VI
0.94
0.92
0.86
0.87
0.95
0.94
0.89
0.90
0.93
0.91
0.81
0.88
0.91
0.90
0.75
0.84
was used in the regional model
conclusion that, while VI
(Table 3) to predict A
hh and Avv , the second-order causal link
between backscatter and the vegetation itself was biomass and
leaf area. In turn, biomass and leaf area were probably surrogates for grass height, density, canopy permittivity, or plant
served as a surromoisture. In Guarico and Apure at least, VI
gate for these several variables. This conclusion is in harmony
with the high-frequency literature discussed above (Schmullius
and Furrer, 1992; Mougin et al., 1995; Lo Seen et al., 1995;
Stiles and Sarabandi, 2000; Stiles and Ulaby, 2000; Herold
et al., 2000) as well as scatterometer studies of the Pantanal
wetlands of South America conducted at C-band (Wismann
and Boehnke, 1996). There was no evidence to link the late
wet-season decrease apparent in many backscatter curves
with increased attenuation from increasing grass height.
As discussed above, the regional regression coefficients
relating NDVI and backscatter (Table 3) were statistically
identical for both the A
hh and Avv models. This finding supports both empirical and theoretical research indicating that
Ku-band horizontal and vertical backscatter of grass should be
nearly identical under dry ground conditions at 50° i.a. (Stiles
and Sarabandi, 2000; Stiles and Ulaby, 2000). At the regional
scale and i.a. used in this research, there was no evidence that
one polarization was superior to the other at measuring the
vegetative properties of the savanna grasslands.
The effect of surficial soil moisture on Ku-band backscatter over the region was represented in the regression by the
inclusion of PT in the A
hh model. Its failure to account for
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
significant variance in the regional A
vv model agrees with
theoretical work that X-Band vv energy interacts primarily
with the top of the grassy vegetation itself whereas hh energy
includes a stronger ground component to the scattering (Stiles
and Sarabandi, 2000; Stiles and Ulaby, 2000). This phenomenon is manifest in all the study areas.
Both regionally and in all five of the study sites, there
and PT, ranging from
was a significant relationship between D
R 0.85 for Apure to R 0.64 for the Meta site (p 0.05).
This polarization dependency with soil moisture is in
harmony with X-band literature already reviewed above
(Oh et al., 1992).
The last explanatory factor in the region models (Table 3)
is the variable percentage of gleysols within the individual
study sites (G%). Gleysols are typical of many regions in the
world, and represent soils with a seasonally high water table
sufficient for anaerobic processes to dominate the soil biology
and chemistry. These soils frequently have a grey or blue-gray
cast to them, and can be periodically flooded. They are not
necessarily infertile, and can support both grass and tree
vegetation.
and PT , the interUnlike the interpretation of factors VI
pretation of G% remains unclear. The relationship between
backscatter and gleysol area was direct rather than inverse;
both A
hh and Avv increased with increasing gleysol. Examination of the Landsat imagery indicated that vegetation in many
of the poorly drained gleysols remained green for an extended
period throughout the dry season, while other vegetated areas
became brown and desiccated in a fashion similar to the
better-drained ferralsol areas. The Landsat imagery also indicated that vegetation was much denser at the height of the wet
season in the gley areas, and evidence of some ponding and
flooding was noticed. Thus, whether bare and moist or heavily vegetated, the gleysol areas represented areas of high
backscatter within the 16-km2 reconstructed SeaWinds pixels
utilized in this research. This was true regardless of season.
or PT model
Why this effect was not captured in either the VI
components remains speculative. Because the gley soils
reflected soil moisture and ponding as well as a vegetative
factor, G% may represent the interaction between soil moisture
or PT variables
and vegetation not represented by the VI
alone.
Conclusion
Summary
Given the relationships cited above, the following generalizations can be drawn, albeit with different levels of certitude:
•
•
•
•
•
Both A
hh and Avv of tropical savannas are largely a function of
the savanna vegetation itself, most likely plant biomass and
is a convenient surrogate. The
regional leaf area for which VI
relationship between backscatter and NDVI is linear and positive. The relationship between biomass and backscatter is
identical for both A
hh and Avv .
A
is
more
sensitive
to
soil
moisture changes than is A
hh
vv.
cited above, increases in soil
Like the relationship with VI
moisture are reflected in an increase of backscatter, and can
also be modeled as a straight line. Whereas rainfall is impor
tant to explaining A
hh backscatter, it is not statistically signifi
cant in the regional A
vv model.
), changes in VI
produce a threeOf the two factors (PT vs. VI
times greater increase in A
hh backscatter than does a corresponding change in rainfall. By logical extension, biomass and
leaf area are much more important to savanna Ku-band backscatter than is soil moisture (at 50° i.a.).
is a function of
Polarization dependence as measured by D
rainfall/soil moisture. The relationship is positive, and linear
or curvilinear.
The percentage of gley soil within the five study sites is
also an explanatory factor in explaining both A
hh and Avv
November 2003
1251
02-111.qxd 10/8/03 12:22 PM Page 1252
backscatter. Uncertainty remains about the causal link, but
it may represent an interaction term between vegetation and
or PT alone.
soil moisture not captured by either VI
These conclusions should be carefully interpreted in the
context of the research limitations. First, this pilot study was
conducted over a period of a single calendar year. Repetition
of the experiment for other years where SeaWinds data are
available is warranted to increase the size of the data set and
clarify the relationships discussed above. Second, it is important to remember that the data used in this project were a
monthly composite of daily values. It is common knowledge
that the process of monthly averaging removes significant
variance that would manifest itself in a daily or weekly study.
This is particularly true of the soil moisture component,
which changes rapidly with the daily weather (Scipal et al.,
2002a; Scipal et al., 2002b). Although the use of monthly
image composites makes the process of statistical modeling
easier, it hides complexity that must eventually be revealed
and examined at shorter temporal scales if the physical basis
of Ku-backscatter of tropical grasslands is to be properly understood. The unavailability of extensive above ground biomass data for validation purposes is a third limitation of the
study.
Implications
This study hints at several important possibilities. Foremost
is the clear ability to monitor vegetation of savanna areas
using high-frequency radar. By using vertically polarized
data, the effect of soil moisture is minimized, and vegetation
state can be evaluated. More fieldwork to clearly establish the
relationship between backscatter and plant biophysical properties is warranted to clarify the precise meaning of “vegetation state,” although biomass and leaf area remain strong
candidates.
Alternatively, if soil moisture rather than vegetation state
was the biophysical parameter desired, the polarization
dependence existing between vv and hh backscatter might be
used to infer it. The ability to directly infer rainfall amounts
from polarization dependence is also an exciting possibility
suggested by this research. However, if operationalization was
desired, this would probably require site-specific models relating soil moisture to rainfall for each soil unit in the region.
To maximize accuracy, the site-specific models would likely
change seasonally or monthly. This effort would also require
substantial field data collection.
The successful use of ERS-1, and ERS-2 C-band scatterometry for grassland and soil moisture studies has been briefly reviewed in the introduction. Given the success of this Ku-band
research, the savannas of South America may provide another
fertile laboratory for such similar research for reconstructed
ERS AMI scatterometry. The potential to link C-band imagery,
Ku-band imagery, and AVHRR/MODIS data for modeling change
in the savanna should not be ignored. As scientists seek to
better understand the savanna contributions to global CO2,
high-frequency radar data at both C and Ku-bands may be an
important contributor to this effort.
Appendix A
As mentioned earlier, it was necessary to normalize the
SeaWinds vertical and horizontal data to a common incidence
angle using NSCAT data. This appendix describes the normalization procedure.
Unlike the SeaWinds algorithm, the NSCAT scatterometer
image reconstruction process produces two images. The NSCAT
A image is a weighted average of the original low-resolution
measurements, standardized to 40° incidence angle. The second image (B image) is an image of slope; it is a measure of
the incidence angle dependence with units of db/degree of
1252
November 2003
incidence angle change. Like the A images, the B images can
also be “averaged” over a time period to determine the average incident angle dependence of particular targets during the
temporal span. For this research, the averaging process was
performed using NSCAT data individually for the five savanna
study areas, producing simple constants representing Ku-band
backscatter incidence angle dependence for the sites. These
constants were then used to adjust both A
hh and Avv for the
sites to 50° incidence angle month by month.
For any land cover, the accuracy of this correction is
dependent on (1) the linearity of the incidence angle dependence in target’s backscatter over the range of the original
NSCAT measurements (16° to 63°), and (2) any change in the
land cover itself between the NSCAT overflight period of 1996/
1997 and the SeaWinds 2000 timespan that might alter the
incidence angle dependence. Given that the adjustment re
quired to standardize the SeaWinds A
hh and Avv data to 50°
requires only a 4° adjustment to both polarizations; any failure to completely satisfy these two assumptions is of minor
consequence.
To validate the approach, an alternative method of correc
tion was also tested. Using the 12 monthly SeaWinds A
hh and
A
vv images, a simple difference operation was performed
(A
hh Avv ) over Brazilian rainforest for each month of the
year 2000. These values were then averaged by month to produce a single constant describing the “apparent” polarization
dependence of the forested target for the given month. Because Ku-band backscatter from tropical forest vegetation
should be polarization independent, any difference between
A
hh and Avv could be attributed to the difference in incidence
angle alone. The monthly constants produced in this alternative correction approach agree well with the approach
using the NSCAT-based method, and the differences do not
impact the conclusions drawn in the research.
Appendix B
As discussed in the paper, the relationships between Estación
Biológica de Los Llanos (EBLL) biomass/leaf area index, and
Apure/Guarico backscatter is statistically significant. However, the 14 years difference between the field data and SeaWinds data cannot be ignored. The probability of a serendipitously high correlation between the 1986 EBLL field data and
2000 SeaWinds backscatter data is low, but the issue is central
enough to inferences made in the paper that the possibility
must be eliminated.
The argument supporting the logical comparison requires
evidence that (1) seasonal weather patterns of rainfall and
(2) vegetation patterns of green-up/senescence are sufficiently
similar between 1986 and 2000 to justify the comparison.
Unfortunately, no argument for comparison can be made using
climatic data. Climate data for the EBLL site during 1986 and
1987 are unavailable. Weather data for nearby stations are incomplete and weather station data for more distant coastal
stations are not representative of the llanos in either geographic
site or situation.
Although climatic data were not available, 1986 NDVI data
1986) for the Apure and Guarico sites were obtained and com(VI
2000) to determine whether the
pared to the 2000 NDVI data (VI
seasonal change in the savanna surrounding the EBLL field site
was similar between the two dates. The correlation between
2000 and VI
1986 for Apure was extremely strong (R 0.94,
VI
n 12, p 0.01). Furthermore, on a month-by-month basis,
2000 and VI
1986 differed by only
the corresponding values of VI
0.04 (average), with a maximum NDVI difference of 0.14 recorded
2000 and
in November. An identical comparison of Guarico VI
1986 provided similarly high agreement. The correlation beVI
2000 and VI
1986 for Guarico was R 0.92 (n 12, p tween VI
0.01), with an average monthly deviation of 0.08.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
02-111.qxd 10/8/03 12:22 PM Page 1253
The last two rows of Table 5 show correlation coefficients
and backscatter for both 1986 and 2000. By combetween VI
paring the values for the two years, the relationship between
is approximately the same. This is particubackscatter and VI
larly apparent for the Apure site. We note the apparent anachro1986 and backscatter is
nism whereby the correlation between VI
2000. However, the test for
greater than its correlation with VI
equality of two correlation coefficients indicates that these differences are not statistically significant (p 0.5).
As measured by NDVI at least, there is little difference between the seasonal patterns in the Apure and Guarico sites between 1986 and 2000, and by extension, to the EBLL field site
between them. As long as the difference in dates is not forgotten, relating the 1986 EBLL field data to the 2000 SeaWinds
data presents no serious logical problems.
References
Andrea, M.O., 1991. Biomass burning: Its history, use, and distribution and its impact on environmental quality and global climate,
Global Biomass Burning (J.S. Levine, editor), MIT Press,
Cambridge, Massachusetts, pp. 3–21.
Atjay, G.L., P. Ketner, and P. Duringneaud, 1979. Terrestrial primary
production and phytomass, SCOPE 13: The Global Carbon Cycle
(B. Bolin, E. Degens, S. Kempe, and P. Ketner, editors), Wiley,
Chichester, United Kingdom, pp. 129–181.
Bakhtiari, S., and R. Zoughi, 1991. A model for backscattering
characteristics of tall grass prairie grass canopies at microwave
frequencies, Remote Sensing of Environment, 36:137–147.
Baumgardner, M.F. (editor), 1986. Project Proposal World Soils and
Terrain Digital Database at a Scale 1:1M (Soter), International
Society of Soil Science, Wageningen, The Netherlands, 23 p.
Blydenstein, J., 1967. Tropical savanna vegetation of the Llanos of
Colombia, Ecology, 48(1):1–15.
Brown, K.S., Jr., and G.T. Prance, 1987. Soils and vegetation,
Quaternary History in Tropical America (T.C. Whitmore and
G.T. Prance, editors), Oxford Monographs on Biogeography,
No. 3, Oxford, Clarendon Press, Oxford, United Kingdom,
214 p.
Dobson, M.C., and F. Ulaby, 1981. Microwave backscatter dependence
on surface roughness, soil moisture, and soil texture: Part III—
Soil tension, IEEE Transactions on Geoscience and Remote
Sensing, GE-19(1):51–61.
Early, D.S., and D.G. Long, 2001. Image reconstruction and enhanced
resolution imaging from irregular samples, IEEE Transactions on
Geoscience and Remote Sensing, 39(2):291–302.
Hair, J.F., Jr., R.E. Anderson, R.L. Tatham, and W.C. Black, 1998.
Multivariate Data Analysis, Fifth Edition, Prentice Hall, Upper
Saddle River, New Jersey, 730 p.
Hardin, P.J., and D.G. Long, 1994. Discriminating between subtropical vegetation formations using reconstructed high-resolution
Seasat-A scatterometer data, Photogrammetric Engineering &
Remote Sensing, 60(12):1453–1462.
Henderson, F.M., 1995. Environmental factors and the detection of
open water areas with X-band radar imagery, International
Journal of Remote Sensing, 16(13):2423–2437.
Herold, M., C.C. Schmullius, and I. Hajnsek, 2000. Multifrequency and
polarimetric radar remote sensing of grassland—Geobiophysical
and landcover parameter retrieval with E-SAR data, Proceedings
of the 20th EARSEL Symposium Remote Sensing in the 21st Century: A Decade of Trans-European Remote Sensing Cooperation,
14–16 June, Dresden, Germany, pp. 95–102.
Hess, T., W. Stephens, and G. Thomas, 1996. Modelling NDVI from
decadal rainfall data in the north east arid zone of Nigeria,
Journal of Environmental Management, 48:249–261.
Hill, M.J., G.E. Donald, and P.J. Vickery, 1999. Relating radar backscatter to biophysical properties of temperate perennial grassland,
Remote Sensing of Environment, 67:15–31.
IFARS, 2002. Land Applications of Scatterometer Data, Institute for
Applied Remote Sensing, URL: http://www.ifars.de/eolas/
scatapps.htm, last accessed 22 April 2003.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Kennett, R.G., and F.K. Li, 1989. Seasat over land scatterometer data,
part I: Global overview of the Ku-band backscatter coefficients,
IEEE Transactions on Geoscience and Remote Sensing,
27(5):592–605.
Levine, J.S., T. Bobbe, N. Ray, R. Witt, and A. Singh, 1999. Wildland
Fires and the Environment: A Global Synthesis, Technical Report
99–1, Division of Environmental Information/EarthWatch, United
Nations Environment Programme, Nairobi, Kenya, 46 p.
Lo Seen, D., E. Mougin, S. Rambal, A. Gaston, and P. Hiernaux, 1995.
A regional sahelian grassland model to be coupled with satellite
multispectral data. II. Toward the control of its simulations by
remotely sensed indices, Remote Sensing of Environment,
52:194–206.
Long, D.G., P.J. Hardin, and P. Whiting, 1993. Resolution enhancement of spaceborne scatterometer data, IEEE Transactions on
Geoscience and Remote Sensing, 31(3):700–715.
Magagi, R.D., and Y.H. Kerr, 1997. Retrieval of soil moisture and vegetation characteristics by use of the ERS-1 wind scatterometer over
arid and semi-arid areas, Journal of Hydrology, 188/189: 361–384.
Marascuilo, L.A., and J.R. Levin, 1983. Multivariate Statistics in the
Social Sciences: A Researcher’s Guide, Brooks/Cole, Belmont,
California, 530 p.
Martin, Jr., R.D., G. Asrar, and E.T. Kanemasu, 1989. C-band scatterometer measurements of a tallgrass prairie, Remote Sensing of
Environment, 29:281–292.
Mougin, E., D. Lo Seen, S. Rambal, A. Gaston, and P. Hiernaux, 1995.
A regional sahelian grassland model to be coupled with satellite
multispectral data. I. Model description and validation, Remote
Sensing of Environment, 52:217–238.
NASA/BYU, 2002. Scatterometer Climate Record Pathfinder, Brigham
Young University, Provo, Utah, URL: http://www.scp.byu.edu,
last accessed 22 April 2003.
NCDC, 2002. Global Surface Summary of Day, National Climate Data
Center, URL: http://www.ncdc.noaa.gov, last accessed 22 April
2003.
NOAA/NASA, 2002. Pathfinder AVHRR Land Data Sets, NASA
Goddard Space Flight Center, URL: http://daac.gsfc.nasa.gov, last
accessed 22 April 2003.
Oh, Y., K. Sarabandi, and F.T. Ulaby, 1992. An empirical model and
an inversion technique for radar scattering from bare soil surfaces, IEEE Transactions on Geoscience and Remote Sensing,
30(2):370–381.
Olson, B.E., and J.R. Lacey, 1996. Basic Principles of Grass Growth
and Management, EB 35, Montana State University Extension
Service, Bozeman, Montana, 13 p.
Saatchi, S.S., D.M. Le Vine, and R.H. Lang, 1994. Microwave
backscattering and emission model for grass canopies, IEEE
Transactions on Geoscience and Remote Sensing, 32(1):177–186.
San Jose, J., and R.A. Montes, 2002. NPP Grassland: Calabozo,
Venezuela, 1969–1987, from Oak Ridge National Laboratory
Distributed Active Archive Center, Oak Ridge, Tennessee, URL:
http://www.daac.ornl.gov, last accessed 22 April 2003.
Schmullius, C., and R. Furrer, 1992. Frequency dependence of radar
backscattering under different moisture conditions of vegetationcovered soil, International Journal of Remote Sensing,
13(12):233–245.
Scipal, K., W. Wagner and R. Kidd, 2002a. Comparison of Ku- and
C-band backscatter time series over land, Geoscience and Remote
Sensing Symposium, 2002. IGARSS ’02, 24–28 June, Toronto,
Canada (IEEE International), 3:1343–1345.
Scipal, K., W. Wagner, M. Trommler, and K. Naumann, 2002b. The
global soil moisture archive 1992–2000 from ERS scatterometer
data: First results, Geoscience and Remote Sensing Symposium,
2002. IGARSS ’02, 24–28 June, Toronto, Canada (IEEE International), 3:1399–1401.
Stiles, J.M., and K. Sarabandi, 2000. Electromagnetic scattering from
grassland—Part I: A fully phase-coherent scattering model,
IEEE Transactions on Geoscience and Remote Sensing, 38(1):
339–348.
Stiles, J.M., and F.T. Ulaby, 2000. Electromagnetic scattering from
grassland—Part II: Measurement and modeling results, IEEE
Transactions on Geoscience and Remote Sensing, 38(1):349–356.
November 2003
1253