PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
January 2015
59
Sub-pixel-scale Land Cover Map Updating
by Integrating Change Detection and
Sub-Pixel Mapping
Xiaodong Li, Yun Du, and Feng Ling
Abstract
Coarse-resolution remotely sensed images are high in tem-
poral repetition rates, but their low spatial resolution limits
their application in updating land cover maps. Our proposed
land cover updating method involves the use of coarse-reso-
lution images to update fine-resolution land cover maps. The
method comprises change detection and sub-pixel mapping
methods. The current coarse-resolution image is unmixed,
and the previous fine-resolution map is spatially degraded
to produce current and previous class fraction images. A
change detection method is applied to these fraction images
to create a fine-resolution binary change/non-change map.
Finally, a sub-pixel mapping method is applied to update the
fine-resolution pixel labels that are changed in the change/
non-change map. The proposed method is compared with a
pixel-based classification method and two sub-pixel mapping
methods. The proposed method maintains most of the spatial
patterns of land cover classes that are unchanged in the
previous and current images, whereas other methods cannot.
Introduction
Remotely sensed images can provide reliable land cover infor-
mation at different scales and are the primary data utilized in
the production and updating of land cover maps. At the glob-
al scale, coarse-resolution images, such as those obtained with
a moderate-resolution imaging spectroradiometer (
MODIS
),
have been applied to build land cover products, such as the
MODIS
land cover product (Friedl
et al
., 2002). Coarse-resolu-
tion images are high in temporal repetition rates, which allow
the timely updating of land cover maps and the creation of
long-term land cover products. However, the spatial reso-
lution of coarse-resolution images is low. Coarse-resolution
land cover products fail to satisfy regional-scale land cover re-
source and landscape analyses. At the regional scale, fine-res-
olution remotely sensed images are the primary data utilized
to generate land cover maps. For instance, Landsat images at a
spatial resolution of 30 m are utilized to produce and update
the National Land Cover Database (
NLCD
) of the United States
(Homer
et al
., 2007). However, owing to the tradeoff between
spatial and temporal resolution, fine-resolution images have
their limitations because they are often acquired at a relative-
ly low temporal resolution. The land cover products from
fine-resolution images are derived only from remotely sensed
data acquired during one or several years, and these products
represent the land cover characteristics of a specific period.
Therefore, they lack not only long-term but also timely land
cover change information.
Using a current coarse-resolution image and a previous
fine-resolution map to timely update fine-resolution land cov-
er products at the regional scale is meaningful and challeng-
ing. This task necessitates the use of multi-resolution images,
which provide mutually supplementary land cover informa-
tion at different scales. A popular approach that combines
fine-resolution and coarse-resolution images is the use of
coarse-resolution images that cover the entire area as the pri-
mary data source, as well as fine-resolution images that cover
a part of the area as training samples. Braswell
et al
. (2003)
combined coarse-resolution and fine-resolution images to
extract land cover fraction images at the sub-pixel scale using
soft classification, which predicts land cover class fractional
information within each coarse-resolution pixel. The fine-res-
olution images were utilized to train endmember signatures,
and the coarse-resolution images were utilized for spectral
unmixing. Lu
et al
. (2011) integrated
MODIS
and Landsat im-
ages to map a fractional forest cover in the Brazilian Amazon.
MODIS
images were unmixed to forest fraction images, whereas
Landsat images were utilized to calibrate the forest fraction
images. However, the aforementioned methods can only detect
land cover fraction within each coarse-resolution pixel and
cannot produce fine-resolution land cover maps.
Sub-pixel mapping (
SPM
) or super-resolution mapping is
a technique that transforms a coarse-resolution image or a
spectral unmixing result into a fine-scale hard classification
map by dividing pixels into sub-pixels and assigning different
classes to these sub-pixels (Foody, 2006; Atkinson, 2009).
SPM
provides more information than spectral unmixing during
the downscaling of coarse-resolution images because
SPM
can
specify the location of each class within the coarse pixels.
Generally,
SPM
adopts mono-temporal coarse-resolution
remotely sensed images as input. In fact,
SPM
is an ill-posed
inverse problem of transforming a coarse-resolution fraction
image to a fine-resolution land cover map, and
SPM
accuracy
is influenced by the uncertainty in determining fine-reso-
lution pixel labels (Nguyen
et al
., 2006; Ling
et al
., 2010).
The combination of a current coarse-resolution image and a
previous fine-resolution land cover map is useful in reduc-
ing
SPM
uncertainty. Ling
et al
. (2011) developed a sub-pixel
scale land cover change mapping method by using a current
coarse-resolution remotely sensed image and a previous
fine-resolution land cover map. This method was directly
used on land cover fraction images obtained by spectral
unmixing applied to remotely sensed images; fraction image
errors reduced the accuracy of the result.
The integration of a previous fine-resolution land cover
map into land cover classification and map updating accura-
Institute of Geodesy and Geophysics, Chinese Academy of sci-
ences, 340 XuDong Rd. Wuhan 430077, Hubei, China (lingf@
whigg.ac.cn).
Photogrammetric Engineering & Remote Sensing
Vol. 81, No. 1, January 2015, pp. 59–67.
0099-1112/15/811–59
© 2014 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.81.1.59
