2250 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 51, NO. 4, APRIL 2013
Fusion of MODIS Images Using
Kriging With External Drift
Marcio H. Ribeiro Sales, Carlos M. Souza, Jr., and Phaedon C. Kyriakidis
Abstract—The Moderate Resolution Imaging Spectroradiome-
ter (MODIS) has been used in several remote sensing studies,
including land, ocean, and atmospheric applications. The ad-
vantages of this sensor are its high spectral resolution, with 36
spectral bands; its high revisiting frequency; and its public domain
availability. The first seven bands of MODIS are in the visible,
near-infrared, and mid-infrared spectral regions of the electro-
magnetic spectrum which are sensitive to spectral changes due
to deforestation, burned areas, and vegetation regrowth, among
other land-use changes, making near-real-time forest monitoring
a suitable application. However, the different spatial resolution
of the spectral bands placed in these spectral regions imposes
challenges to combine them in forest monitoring applications.
In this paper, we present an algorithm based on geostatistics to
downscale five 500-m MODIS pixel bands to match two 250-m
pixel bands. We also discuss the advantages and limitations of
this method in relation to existing downscaling algorithms. Our
proposed method merges the data to the best spatial resolution and
better retains the spectral information of the original data.
Index Terms—Forest monitoring, image fusion, kriging with an
external drift, Moderate Resolution Imaging Spectroradiometer
(MODIS), spatial resolution.
I. I NTRODUCTION
T
HE Moderate Resolution Imaging Spectroradiometer
(MODIS) measures reflected and emitted energy at 36
bands placed at a wavelength range of 0.405–14.385 μm, with
spatial resolutions of 250–1000 m [1], acquired at a near daily
basis. The spectral, spatial, and temporal resolutions of MODIS
render it a powerful tool for global and large-scale applications
of remote sensing. The first seven bands of MODIS cover the
wavelength range most related to features of land, cloud, and
aerosol properties [2], for example, chlorophyll absorption at
Manuscript received December 21, 2011; revised May 18, 2012; accepted
May 26, 2012. Date of publication September 6, 2012; date of current version
March 21, 2013. This work was supported in part by The David and Lucile
Packard Foundation and in part by the Climate Land Use Alliance. The work
of M. H. Ribeiro Sales and P. C. Kyriakidis was supported in part by the U.S.
National Geospatial Intelligence Agency.
M. H. Ribeiro Sales is with the Instituto do Homem e Meio Ambiente
da Amazônia—Imazon, Belém 66613-397, Brazil, and also with the Pro-
duction Ecology and Resource Conservation Graduate School, Wageningen
University, 6709 PB Wageningen, The Netherlands (e-mail: marcio@imazon.
org.br).
C. M. Souza, Jr., is with the Instituto do Homem e Meio Ambiente
da Amazônia—Imazon, Belém 66613-397, Brazil (e-mail: souzajr@imazon.
org.br).
P. C. Kyriakidis is with the Department of Geography, University of Califor-
nia at Santa Barbara, Santa Barbara, CA 93106-4060 USA, and also with the
Department of Geography, University of the Aegean, 81100 Mytilene, Greece
(e-mail: pckyriakidis@gmail.com).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TGRS.2012.2208467
red band (i.e., band 1, 0.620–0.670 μm) and canopy backscat-
tering at near infrared (band 2, 0.841–0.876 μm). These bands
can also be combined to retrieve biophysical properties of
forests using products such as vegetation indices available at the
Land Process Distributed Active Archive Center (LP DAAC)
Web site [3]. These features make the first seven MODIS bands
the most important ones for land-use and land-cover change
(LULCC) applications [1]. In Brazil, MODIS data have been
used to detect deforestation at near real time for forest law
enforcement and environmental policy verification [4], [5].
However, a limitation is imposed by the difference in the
spatial resolutions of these bands: Bands 1 and 2 are the
only ones available at the 250-m spatial resolution [6] while
bands 3–7 are in the 500-m spatial resolution. This difference
in spatial resolutions limits the generation of products based
on MODIS data at the 250-m resolution using the full range
of spectral information important for forest monitoring. The
250-m resolution bands 1 and 2 can, however, be spatially
degraded to 500 m to create a product at this resolution with
seven spectral bands, but the 250-m resolution is preferred
because many changes in the land cover occur at this spatial
scale [7]. Splitting a 500-m pixel into four 250-m pixels would
result in a product with seven spectral bands with no necessarily
direct spectral correspondence from the two spatial resolutions.
Spectral mixing, for example, is a scale-dependent [8] phe-
nomenon that could drastically affect this type of data-fused
product. Alternatively, bands 1 and 2 alone have been used
for land-cover mapping applications [9], [10]. However, the
addition of information on the infrared bands can significantly
improve the detection and analysis of LULCC [9].
Combining data from different spatial resolutions such
as 250- and 500-m MODIS bands is a typical image fusion
problem. The general objective of image fusion is to combine
two (or more) images in order to obtain a better image, retaining
the desired characteristic of each of the original images [11].
Image downscaling is a special case of image fusion, in which
fine spatial coarser spectral resolution images are combined
with coarser spatial higher spectral resolution images to yield a
spectrally enhanced product at the finer spatial resolution [12].
This type of approach is more appropriate than pixel splitting
the 500-m MODIS images because it retains the spatial infor-
mation at the finer resolution and the spectral correspondence
of the two spatial-resolution bands is guaranteed. Image
downscaling has the potential to greatly improve the utility of
MODIS data for several remote sensing applications. For ex-
ample, improvements in snow cover mapping have been found
after the application of a downscaling algorithm to MODIS
images [13].
0196-2892/$31.00 © 2012 IEEE