Integrating Landsat with MODIS Products for Vegetation Monitoring - Satellite-based Applications on Climate Change - page 246

The original STARFM code has been released for public use since 2006. The
STARFM and derived approaches have been tested and used for fusing Landsat and
MODIS reflectance (Gao et al. 2006 ; Hilker et al. 2009a ). Several ongoing
researches are expanding STARFM approach to different biophysical parameters
and use it for sensors other than Landsat and MODIS.

16.3.2 The Enhanced STARFM Approach

In order to better handle heterogeneous pixels even if no “pure” neighbor pixel
exists, an enhanced STARFM approach was recently developed based on a pixel
unmixing theory (Zhu et al. 2010 ). An additional assumption in the enhanced
STARFM is that the percentages of land types contained in the mixed coarse-
resolution pixel remain the same during the prediction period. Therefore, the
reflectance of a mixed MODIS pixel can be described as the linear mixture of
Landsat pixels for two input pairs with the same percentages of land cover types.
The ESTARFM approach introduces a conversion coefficient into the prediction.
The conversion coefficient indicates the ratio of the change of reflectance for the
end-member to the change of reflectance for a mixed coarse-resolution pixel from
input pairs. When the end-members are taken as fine-resolution pixels within a
mixed coarse-resolution pixel, the conversion coefficient can be computed by
linearly regressing the reflectance changes of fine-resolution pixels of the same
end-member and coarse-resolution pixel. Taking into consideration the spectral
similar pixels, the final prediction for the center pixel ( w/ 2, w/ 2) in the moving
window with end-member i can be revised to

X

w

X

w

X

n

(16.4)

Lðw= 2 ; w= 2 ; t 0 Þ¼ W ijk ðL i ðx; y; t k Þþv i ðx; yÞðMðx; y; t 0 ÞMðx; y; t k ÞÞÞ

x¼ 1 y¼ 1 k¼ 1

where v i ( x, y ) is the conversion coefficient for the ith end-member in the mixed
pixel ( x, y ). It can be computed based on two acquisition pairs ( t m and t n ):

L i ð x ; y ; t n Þ L i ð x ; y ; t m Þ
Mðx; y; t n ÞMðx; y; t m Þ

ν i ðx; yÞ¼

(16.5)

Results from simulated data and real satellite data show that the enhanced
STARFM can improve the accuracy of prediction, especially for complex heteroge-
neous landscapes, and preserve spatial details for small patches. In a homogeneous
area, the prediction of ESTARFM is slightly better than STARFM with the average
absolute difference of 0.0106 (vs. 0.0129) for the NIR band. In a complex mixed area,
the prediction accuracy of ESTARFM is improved significantly when compared to the
original STARFM (0.0135 vs. 0.0194) for NIR band (Zhu et al. 2010 ).