Improving the Upscaling of Land Cover Maps
by Fusing Uncertainty and Spatial Structure
Information
Peijun Sun, Russell G. Congalton, and Yaozhong Pan
Abstract
Upscaling land cover maps is broadly employed to fill data
gaps or match the spatial-resolution of preexisting projects.
However, existing methods introduce systematic errors in the
area information and the landscape pattern. We developed
an upscaling method fusing the spatial structure informa-
tion (i.e., class Membership probability) and the uncertainty
information of the base map (e.g., Confidence level prob-
ability), called Fusing class Membership probability and
Confidence level probability (
FMC
). The results showed that
FMC
obtained higher upscaling efficiency, and mitigated
the negative influence of landscape heterogeneity and the
negative influence of unequal proportions of land cover
in the base maps, on the upscaling compared to Majority
Rule Based (
MRB
) method. Additionally,
FMC
can reduce the
uncertainty/error when these upscaled maps are used as
input to Earth observation model (e.g., land cover change).
Introduction
Land cover maps have been increasingly emphasized as an
essential variable for earth science modeling (Inglada
et al
.,
2017; Mack
et al
., 2017). Land cover distribution and area
provides fundamental data for various applications (Lu and
Weng, 2007; Fuchs
et al
., 2015; Ge
et al
., 2016), such as eco-
system services (Dong
et al
., 2015), environmental planning
(Grafius
et al
., 2016), climate change (Feddema
et al
., 2005),
hydrological processes and policy making (Chen
et al
., 2016).
Various research investigations require land cover data at
specific spatial resolutions (Stein
et al
., 2001; Ju
et al
., 2005;
Kitron
et al
., 2006; Bai
et al
., 2014).
To meet the requirements of these diverse models and ap-
plications, a variety of land cover maps at global or regional
scale have been produced (Congalton
et al
., 2014) including:
GLC 2000 (Bartholomé and Belward, 2005);
MODIS
Land Cover
(Friedl
et al
., 2010);
FROM-GLC
(Wang et al., 2015); and Glob-
Cover 2009 (Sophie and Pierre, 2010). These products extend
our understanding of the extent and distribution of land cover
types. However, the applications of these products can be
problematic depending on the availability of remote sensing
data from sensors with specific spatial resolutions (Yang and
Merchant, 1997; Raj
et al
., 2013; Wang
et al
., 2015).
To address this spatial resolution issue, rescaling of these
products becomes crucial for filling the data gaps (Stein
et al
.,
2001; Gardner
et al
., 2008). The literature on rescaling or pro-
ducing land cover maps over a variety of spatial resolutions
includes numerous ways of upscaling and downscaling (Tang
et al
., 2015). Upscaling decreases the spatial resolution while
the downscaling method increases the spatial resolution. The
focus of this research is in the improvement of upscaling land
cover maps given various different landscape scenarios.
In the past decades, many land cover mapping efforts have
been conducted to generate fine resolution land cover maps
using high spatial resolution sensors at regional to national
scales (Hansen and Loveland, 2012). These land cover maps
contribute by providing more detailed land information and
can also be upscaled to fill data gaps at coarser spatial resolu-
tions. For example, the National Agricultural Statistics Service
(
NASS
) Cropland Data Layer (
CDL
) is annually produced at
30 meter spatial resolution for the US. The Finer Resolution
Observation and Monitoring-Global Land Cover (
FROM-GLC
) is
produced at a global scale (Gong
et al
., 2013). These datasets
are derived with greater local expertise and can provide a more
accurate representation of the characterizations and details of
land cover (Strahler
et al
., 2006; Song
et al
., 2014; Gao
et al
.,
2017; Gonzales and Searcy, 2017), which enables their use as
base maps to obtain more reliable and accurate upscaled maps.
Producing upscaled land cover maps relies on two differ-
ent aggregation logic operations: numerical aggregation and
categorical aggregation (Raj
et al
., 2013; He
et al
., 2002). The
numerical aggregation determines the class type in the up-
scaled map based on a selected function between the coarse
pixel value (e.g., digital number (
DN
)) and its corresponding
fine pixel values (Raj
et al
., 2013). After acquiring the new
value for the coarse pixel, the class type is obtained by classi-
fication techniques (i.e., categorical aggregation). Two numeri-
cal algorithms, mean aggregation, and central pixel resam-
pling (
CPR
), are commonly used to calculate the coarse pixel’s
value. The mean aggregation algorithm estimates the mean of
pixel values (
DN
or reflectance value) from the fine-resolution
map within the area of the associated coarse pixel. This
algorithm is based on the principle that the
DN
or reflectance
value for any coarse pixel is the mean
DN
or reflectance value
over the associated area on the ground (Raj
et al
., 2013). The
CPR
employs a non-overlapping
n
×
n
window corresponding
to the coarse-resolution map and selects the fine resolution
central pixel’s value as the output course pixel’s value (Bian
and Butler, 1999).
Peijun Sun and Yaozhong Pan are with the State Key Laboratory
of Remote Sensing Science, Jointly Sponsored by Beijing
Normal University and Institute of Remote Sensing and Digital
Earth of Chinese Academy of Sciences, Beijing 100875, China;
and with the Institute of Remote Sensing and Engineering,
Faculty of Geographical Science, Beijing Normal University,
Beijing 100875, China (
).
Russell G. Congalton is with the Department of Natural
Resources & the Environment, University of New Hampshire,
Durham, New Hampshire 03824.
Photogrammetric Engineering & Remote Sensing
Vol. 84, No. 2, February 2018, pp. 87–100.
0099-1112/17/87–100
© 2018 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.84.2.87
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
February 2018
87
