across the study area (Figure 5a). The ecosystem-based end-
member time series exhibits the spatial variation of the two
endmembers. Although the woody or herbaceous vegetation
have some variation across ecosystems, the two endmem-
bers are still distinct from each other. In general, the woody
endmember has steeper increase, starts earlier (in September),
remains high longer (until May), and declines later than the
herbaceous endmember (Figure 5b).
Validation of the Derived Products
Verifying the method using ground observations was not prac-
tical since measurements of woody and herbaceous canopy
cover, leaf area index, or
NDVI
were not available at a suitable
scale for comparison. However, we were able to validate the
method using three approaches:
a. Simulation of woody and herbaceous
NDVI
by adding varia-
tion to endmember time series;
b. Comparisons of decomposed woody and herbaceous com-
ponents with cover fractions derived from high resolution
satellite imagery (Helman
et al.
2015)
c. Transect-based visual sampling of high resolution Google
Earth images to create
MODIS
pixel scale estimates of woody
and herbaceous cover and retrieval of supplementary point
data derived from published studies for comparison.
Simulated Data
Data for validation were simulated by adding random val-
ues (range from -0.1 to 0.1) to each observation of the two
endmember time series to simulate the two components of
a mixed time series. The mixed time series is calculated as
sum of fractions of the two simulated components. We then
decomposed the mixed time series based on the two endmem-
bers using our method. The results from the decomposition
were then compared to the simulated components.
Cover Fractions from High Resolution Imagery
High-resolution images including
IKONOS
and Google Earth
data were used to generate the fractional cover dataset. Two
IKONOS
images acquired during the dry season (Figure 1) on
11 May and 11 June 2010 were available. The multispectral
bands (4 m resolution) were fused with the panchromatic
band (1 m resolution) to create pan-sharpened multispectral
images. Each pan-sharpened image was previously clustered
into 100 groups and further classified into photosynthetic
vegetation (
PV
), nonphotosynthetic vegetation, bare soil, and
other land cover classes (Hill
et al.
2016). In this study, the
cluster groups within the
PV
class were further separated into
green woody and green herbaceous classes (Figure 6). Classifi-
cation results were evaluated using manually selected woody
and herbaceous patches from high-resolution images. The
overall classification accuracy was 86.2% for the 11 May im-
age and 92.6% for the 11 June image at 1 m resolution (Table
2). The classification images were converted to the fractional
cover of green woody and herbaceous vegetation within
the grids of the
MODIS
images. We then developed a regres-
sion tree model that correlates the decomposed
NDVI
with a
subset of the fractional cover dataset. Regression tree models
Figure 6. Illustration of generating
MODIS
-resolution fractional cover validation data from high-resolution images. The top box
shows the procedure of converting an
IKONOS
image to 500 m fractional cover image. The bottom box shows the fractional
cover calculation from Google Earth images at selected sites.
Table 2. IKONOS image classification accuracy.
Image Date Class
Product Accuracy (%)
User Accuracy (%)
Overall Accuracy (%)
Kappa Coefficient
May 11
green woody vegetation
77.03
96.57
86.20
0.73
Green herbaceous vegetation
96.83
78.43
June 11
Green woody vegetation
87.82
97.68
92.62
0.85
Green grass vegetation
97.76
88.22
514
July 2019
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
