IEEE Geoscience and Remote Sensing Magazine - September 2015 - 53

all Landsat pixels within a MODIS pixel that belong to
same end-member class have same conversion coefficient, conversion coefficient of these same-class pixels
can be estimated from the regression slope computed
between their Landsat and MODIS reflectance values
in two observed pairs. Similar to STARFM, ESTARFM
also uses spectrally similar pixels within the moving
window. The predicted reflectance for the central pixel
at ^w/2, w/2 h in the moving window based on two observed pair images can be computed as:
2

w w
w w
L ( 2 , 2 t p) = / Tk # (L ( 2 , 2 , t k)
k =1

large land cover patches. The ESTARFM result shows
clearer land cover boundaries than STARFM. Another
comparison was conducted over a more heterogeneous
area in central Virginia with smaller land cover patches.
Fig. 5 shows the STARFM and ESTARFM fused images
compared to the actual Landsat ETM+ image acquired on
that same day (Feb. 26, 2002) over a more heterogeneous
area. The ESTARFM result captured spatial details better than the original STARFM approach when two pairs
of Landsat and MODIS image were used. However, the
STARFM approach is more flexible and can use one input
pair for prediction if it is properly selected. In Fig. 5, the

+ / W ik # v (x i, y i) # (M (x i, y i, t p)
i =1

- M (x i, y i, t k)))

(5)

where Tk determines the contribution from each pair of
observed images. It is computed based on the change magnitude of MODIS images from prediction date to each of
the two input pair dates at coarse resolution. The pair with
smaller change magnitude contributes more to the final
prediction in Eq. 5 assuming it gives a more accurate and
reliable estimate.
D. Comparison of the starfm, estarfm anD
staarCh fusion approaChes
These three data fusion approaches - STARFM, ESTARFM
and STAARCH - have been tested and applied in different
studies published in the literature. They are all based on a
common assumption - that temporal changes observed in
near-daily MODIS imagery are indicative of changes that
should be expected at Landsat scale, and therefore can create
dense time-series of Landsat-like imagery. Biophysical variables such as surface reflectance are used to build a model
that translates these changes between different sensors. All
three approaches use spatial information from neighboring
pixels to predict value for the central pixel. This is a fundamental difference from traditional data fusion approaches.
It is a reasonable assumption that pixels having similar spectral properties or land cover type within a local area (moving window) will generally have a similar temporal change
pattern. Despite these central core commonalities, the three
data fusion approaches described here differ in terms of
their specific application and limitations.
The ESTARFM approach considers the possibility of
different change patterns for different land cover types
within a heterogeneous MODIS pixel by introducing a
landcover-related conversion coefficient. Comparisons
between STARFM and ESTARFM in different sites have
demonstrated the strengths of ESTARFM for producing
Landsat-like images in more heterogonous landscapes
[9], [14]. Fig. 4 shows the observed Landsat ETM+ image
(sub-area in Fig. 2) and the STARFM and ESTARFM fusion results using two pairs of Landsat and MODIS images. Both approaches capture temporal variations and
september 2015

ieee Geoscience and remote sensing magazine

(a) Landsat7 ETM+

(b) ESTARFM

(c) STARFM
Figure 4. The observed Landsat 7 ETM+ (sub-area in Fig. 2) over
the BOREAS Southern Study Area, Canada (54ºN, 104ºW) on July
11, 2011 (a) and data fusion results from ESTARFM (b) and STARFM
(c) using image pairs from May 24 and August 12, 2011 (modified
from Zhu et al., 2010).

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2015