IEEE Geoscience and Remote Sensing Magazine - September 2015 - 57

Fig. 8 demonstrates an
application of ET data fuCumulative ET (mm)
300
sion over a managed loblolly
1000
pine plantation in the coastal plains of North Carolina.
800
Studies in this plantation
600
have demonstrated a stand400
age dependence of water
use and drought response to
200
drought - presumably relat0
0
ed to tree rooting depth [37],
[38]. Strong heterogeneity in
a map of seasonal water use Figure 8. Cumulative ET from day of year 100 to 300 over a managed pine plantation in the North
is observed over this man- Carolina coastal plains (right). Yellow crosses in true-color Landsat image (left) indicate locations of flux
aged forest system, accumu- towers in mature and clear-cut stands.
lated from daily fused ET
approach (backup algorithm for the MODIS BRDF/Aldata at 30m resolution over the 2013 growing season, even
bedo products [39]) is used to correct MODIS daily surusing only a few clear-sky Landsat scenes. The fused results
face reflectance to daily NBAR data. BRDF parameters for
reasonable reproduce daily measurements of ET made in
each pixel are extracted from the MODIS BRDF parama 20-year old stand and a newly clearcut stand within the
eters (MCD43A1) product. Since MODIS BRDF is a 16-day
modeling scene. Multi-year high spatiotemporal resoluproduct with 8-day overlap, we select the 16-day period
tion ET datacubes, developed using STARFM data fusion,
that best covers the Landsat acquisition date.
can be effectively mined to investigate water use response
Landsat and MODIS images have different geolocation
to climate and management forcings.
accuracy. Although the geolocation accuracy for MODIS is
very good in term of MODIS pixel resolution, it may still in
D. OperatiOnal StarFM FraMewOrk
a few Landsat pixels. Since data fusion is performed at LandData fusion approaches assume remote sensing data
sat pixel resolution, a better co-registration between Landfrom different sensors are consistent and comparable.
sat and MODIS is required. In the operational data fusion
Even when Landsat and MODIS data have been careframework, we implemented the automatic image co-regisfully geo-registered, radiometrically calibrated and attration process by computing and searching the maximum
mospherically corrected, they may be inconsistent for
correlation between Landsat and MODIS images. If the locamany reasons. In order to reduce these inconsistencies
tion of the maximum correlation is different from current
and make the data fusion system more robust, an oplocation (directly reprojected using the MODIS reprojection
erational STARFM data fusion framework has been built
tool), the MODIS image will be shifted to the Landsat loca[24]. Compared to earlier implementations of STARFM,
tion which produces the maximum correlation.
several improvements have been incorporated in this
The operational framework allows selection of input
framework. These include viewing angle correction to
pair images either automatically or manually. The manual
the MODIS daily bidirectional reflectance, automated
option allows definition of pair images for any prediction
co-registration of MODIS and Landsat pair images, and
date. The automatic option was implemented in two ways.
automatic selection of Landsat and MODIS image pairs
One is to choose the pair image closest to the prediction
used as input to STARFM.
date. The other option is to select the MODIS-Landsat pair
Data fusion tools can use MODIS daily surface reflecimage that has a higher correlation coefficient with MODIS
tance or MODIS nadir-bidirectional reflectance distriimages between two bracket pair dates.
bution function (BRDF) adjusted reflectance (NBAR) to
The STARFM approach uses an overlapped searching
fuse with Landsat. The MODIS NBAR product is a 16window for each prediction. It can be very time consuming
day product and may not agree with Landsat data on a
in searching for spectrally similar pixels. Recently, we imspecific day especially during period of rapid vegetation
plemented an improved algorithm that allows combining
growth. The MODIS daily surface reflectance is a direcof multiple predictions together if the pair images are the
tional reflectance product which is affected by BRDF efsame. When the same pair images are used, the searching
fects; i.e., the surface reflectance varies with viewing and
process for spectrally similar pixels is the same and thereillumination angles, whereas the Landsat TM, ETM+ and
fore we only need search once and then run predictions for
OLI instruments are normally regarded as nadir-viewing
many prediction dates. This implementation greatly iminstruments. To minimize MODIS-Landsat differences
proves computing efficiency when running data fusion for
due to viewing angle, the operational STARFM framemany dates. In our tests, it is 15 times faster for 30 predicwork corrects the MODIS daily directional reflectance to
tions (same pair) using the improved algorithm comparing
nadir-viewing reflectance. A magnitude BRDF inversion
september 2015

ieee Geoscience and remote sensing magazine

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