The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
2
inverting the Li-Strahler model than using image-based
endmembers. This method provides an avenue to up-scale the
information from local to regional scale. We subsequently call
this method the regional scaling-based endmember extraction.
Additionally, in very densely forested regions, crowns touch
each other and almost no sunlit background component can be
derived, thus no estimation from the Li-Strahler model
inversion is obtained. With regard to increasing the
applicability of the Li-Strahler model inversion for mapping CC
on a per-pixel basis, finding ways, such as spatial interpolation,
to solve missing estimations in densely forested regions is
extremely important.
In summary, the major goal of this study is to use an inversion
of the Li-Strahler geometric-optical model combined with a
scaling-based endmember extraction method and a spatial
interpolation technique to derive forest crown closure over
larger areas and to map changes in crown closure within a time
span of 2 years.
2. STUDY SITE AND DATA
The Three Gorges region of China is chosen as study site in this
work. This region refers to a special area associated with the
Three Gorges Dam and Reservoir project along the Yangtze
River and it is also called the Three Gorges Reservoir region.
The total acreage of this region is about 58,000 km 2 (28°32'-
31°44'N, 105°44'-111°39'E), which includes 20 counties in
Hubei province (ranging from Yichang in the east to Badong in
the west) and Chongqing (ranging from Wushan in the east to
Jiangjin in the west). This study area belongs to the temperate
climate zone (Koeppen: Cwa-Subtropical monsoon (McKnight
and Hess, 2000)). The average annual precipitation is about
1000-1300 mm and the rainy season is between spring and
summer (April-October). Based on a land cover investigation of
2002, the Three Gorges region is occupied by about 43%
cropland, 30% forest, 20% shrub and 3% grassland (Huang et
al., 2006). The forested areas are mainly dominated by
coniferous, deciduous broadleaved and subtropical evergreen
broadleaved species.
The field data were collected in September 2006. With help of
1:50,000 topographic maps and a land cover map of 2002, a
total of 25 sample sites within the forest area of the Three
Gorges region were selected. For each sample site, we first
determined a homogeneous forest area with acreage above 500
m x 500 m. According to information from the yearly local
forest investigation, the sample sites were chosen to be in areas
without severe logging and replanting activities between 2002
and 2006. The central location of each sample site was recorded
by a GPS (±15 m spatial accuracy). At every sample site, at
least 2 sample plots (100 m x 100 m) were selected and forest
structural properties were measured. These include forest crown
closure (CC), crown diameter (CD), stem diameter at breast
height (DBH), tree height (H), trunk height (TH) and stem
density (SD).
Two Landsat TM images (Path 125/Row 39), acquired on
September 1, 2002 and October 8, 2004 respectively, are used
in this study as high spatial resolution data (30 m). The images
have been geometrically corrected and converted from digital
numbers (DNs) to top-of atmosphere (TOA) reflectance. Both
TM images cover the eastern part of the Three Gorges region.
The moderate spatial resolution data covering the whole Three
Gorges region are based on MODIS images, which were
collected at the same dates as the Landsat TM images. We use
the daily ‘surface reflectance’ product of MODIS (MODIS-09,
collection-4) with 7 spectral bands and 500 m spatial resolution
in this study. Due to the importance of the image viewing angle
in the model inversion and scaling approach used in this study,
we use the Aqua-MODIS, which has an observation direction
(nadir), located closer to the Landsat TM images. In addition, a
land cover map of the Three Gorges region from 2002 is used
for identifying the forest region. This map is derived from field
investigations combined with a remote sensing classification
(Zhang et al., 2007). The forest change detection between 2002
and 2004 in this study is only focused on the forest area present
in the 2002 land cover map. Other ancillary data needed for
modelling are the digital elevation model (DEM) with 25 m
spatial resolution over the Three Gorges region and the MODIS
Global Geolocation Angle product, which contains information
on solar illumination and instrument viewing geometry.
3. METHOD
3.1 Inverted Geometric-Optical Model
The Li-Strahler geometric-optical model (Li and Strahler, 1985;
1992) is based on the assumption that the Bidirectional
Reflectance Distribution Function (BRDF) is a purely
geometric phenomenon resulting from a scene of discrete 3-
dimensional objects being illuminated and viewed from
different positions in the hemisphere. For modelling a forest
scene, three components have to be estimated: sunlit canopy-C,
sunlit background-G and shadow-T (Li and Wang, 1995;
Peddle et al., 1999; Peddle et al., 2003). This model also
assumes that the resolution of the remote sensing image is
larger than the size of individual crowns but smaller than the
size of forest stands, and that the individual trees are ‘Poisson’
distributed within the pixel (Woodcock et al., 1994).
To derive CC by inverting Li-Strahler model, the fraction of
sunlit background (K g ) is required as input (Strahler and Jupp,
1990; Li and Strahler, 1992; Woodcock et al., 1997; Zeng et al.,
2007; 2008). Based on the field measurements in the forest area
of the Three Gorges region, the crown shape can be modelled as
an ellipsoid. Therefore, the measured values of tree height from
ground to mid-crown, crown radius in vertical direction and
crown radius in horizontal direction are the necessary inputs of
model inversion. Moreover, the slope and aspect images re
sampled to 500 m spatial resolution and the solar and viewing
angles are also required for model inversion (Schaaf et al., 1994;
Zeng et al., 2007). Since K g is the most critical input, accurate
extraction of the G, C and T endmembers from the MODIS
images of the Three Gorges region for both years is very
important.
3.2 Endmember Extraction
Traditionally, the linear unmixing model has been widely used
to calculate the percentages of several individual surface
components contained in each pixel of a remote sensing image
(Peddle et al., 1999; Goodwin et al., 2005). The model assumes
that the reflectance (S) of each pixel is a linear combination of
endmembers (R), which are the pure reflectance spectra for
each component. The general equations are: