Categorical aggregation assigns a class label to the upscaled
maps based on the class labels in the corresponding fine-res-
olution map. Three categorical aggregation methods: majority
rule based (
MRB
) (e.g., Saura, 2004), random rule-based (RRB)
(e.g., He
et al
., 2002), and point-centered distance-weighted
moving window (PDW) (Gardner
et al
., 2008), have been de-
veloped to implement this categorical aggregation logic.
MRB
assigns the class label for the coarse-resolution map by choos-
ing the most frequently occurring class of the fine-resolution
map (He
et al
., 2002; Saura, 2004). If more than one major
class is present in the base map, the class label is randomly
determined from these major classes (Raj
et al
., 2013). RRB is
based on the random selection of a class from the specified
pixels of the fine-resolution map. The corresponding ag-
gregated pixel is then assigned to that class (He
et al
., 2002).
PDW uses a weighted sampling grid of variable resolution to
sample the existing map, and then randomly selects from the
frequency of cover types derived from the samples to assign
the cover type for the corresponding location in the rescaled
map (Gardner
et al
., 2008).
Previous studies have focused on the performance analy-
sis of these different methods. For example, Raj
et al
. (2013)
conducted an experiment to analyze the advantages and dis-
advantages of numerical and categorical aggregation methods.
The results showed that the dominant class becomes more
clumped while the sparse classes become less clumped when
using
MRB
. The results agree with what was also found by
Yang and Merchant (1997). RRB performs better for maintain-
ing the landscape characteristics (structure) than
MRB
(He
et al
., 2002). PDW is more suitable for ecological resource
management.
CPR
preserves the dominant class type, while
the mean aggregation method increases the proportion of
the non-dominant class types as reported by Bian and Butler
(1999). Additionally,
MRB
is recommended for obtaining
upscaled maps for agriculture planning, which is especially
relevant since in our work here has emphasized agricultural
maps. Overall, these previous studies focused on evaluating
the performance of different aggregation methods and analyz-
ing their discrepancies.
However, these previous studies did not consider the
uncertainty of the land cover map or the spatial structure in-
formation derived from neighboring pixels, both of which are
important components in the rescaling including upscaling
(e.g., Bian and Butler, 1999; Ju
et al
., 2005) and downscaling
(e.g., Boucher, 2006; Ling
et al
., 2014). Curran and Atkinson
(1998) pointed out that the objects on the Earth are more alike
when they are closer to each other than when they are farther
away. In remote sensing, the spectral response of one pixel
will be more similar to the spectral response of the closer
pixels (Hay
et al
., 2001). This concept is further illustrated in
land cover or thematic maps (i.e., the class types of neighbor-
ing pixels have a higher probability to be same as the class
type of the center pixel). The application of this concept
becomes the basis of using class membership probability
(
CMP
) in remote sensing techniques. For example, existing
downscaling approaches using spatial structure information
successfully constructed the land cover distribution in the
fine-resolution map. Ling
et al
. (2014) generated downscaled
maps fusing class membership probability (
CMP
) at two
resolutions: one for fine-resolution maps and the other one at
coarse-resolution. The results demonstrated that integrating
multi-sources of spatial structure information could improve
the accuracy of prediction for downscaling maps.
Although the importance of spatial structure information
has been emphasized in remote sensing and was explored
in downscaling techniques, upscaling research has rarely at-
tempted to employ it to generate coarser-resolution maps. Be-
sides using this spatial structure information, the uncertainty
information of the land cover maps or thematic maps (e.g.,
confidence level of the land cover maps, posterior probability
of the classification maps) has only infrequently been applied
to test if this information can improve the accuracy of either
downscaling or upscaling. Moreover, fusing multi-sources of
information from the land cover maps has been conducted as
an effective way to assist in the downscaling (e.g., Boucher
et
al
., 2008; Ling
et al
., 2014), while rarely investigated in reduc-
ing the error in the upscaling. Therefore, this paper aims at
exploring if fusing multi-sources of information (i.e., spatial
structure information or uncertainty information) can be an
efficient way to upscale maps.
The objective of this study was to propose a new ap-
proach (referred to as Fusing class Membership probability
and Confidence level probability,
FMC
) to generate upscaled
agricultural maps and their associated confidence level maps.
The class membership probability (
CMP
) was determined by
the majority rule-based upscaling method (
MRB
) to predict the
potential spatial distribution for the coarse-resolution (target)
maps. The confidence level probability was used to represent
the uncertainty information of the base map. Additionally, as
previous studies have illustrated, landscape pattern impacts
the upscaling performance. Therefore, we selected six study
sites with different landscape patterns to perform our experi-
ments. Finally, the upscaling efficiency of
FMC
was assessed
by comparing the results to
MRB
using: overall consistency
(
OC
), proportional error (
PE
) (Yang and Merchant, 1997), and
changes in landscape pattern.
Methodology
This section describes the land cover maps, the upscaling
methods, the assessment of the upscaled land cover maps,
and the uncertainty analysis of upscaling. The methods used
in this study were conducted in three steps. In the first step,
the initial land cover maps were downloaded (from
https://
nassgeodata.gmu.edu/CropScape/
) to produce the agricul-
tural maps to be used as the base maps at 30 meter resolution
for the upscaling analysis. Then, the upscaling methods (i.e.,
MRB
and
FMC
) were conducted to generate upscaled maps for
the six study areas. Finally, the third stage was the assessment
and the comparison of the two upscaling methods.
Study Areas and Data Description
The impact of upscaling as evaluated by
OC
and
PE
(Yang and
Merchant, 1997) varies based on the landscape patterns of
each map. Hence, six Agriculture Statistic Districts (ASDs)
with different landscape patterns were selected for analysis
(Figure 1). The landscape patterns/structure were measured
using two landscape metrics: Patch-per-Unit (
PPU
) as a mea-
sure of heterogeneity (Frohn, 1997) and the Aggregation index
(
AI
) as a measure of aggregation degree (Frohn, 1997). These
districts can be divided into two categories: heterogeneous
districts and homogeneous districts. The heterogeneous dis-
tricts (
PPU
>10.0,
AI
<92.0 (Table 1)), ASD1810, ASD1830, and
ASD1870, are located in Indiana in the US. The homogeneous
districts (
PPU
<10.0,
AI
>92.0 (Table 1)), ASD4530, ASD4550,
and ASD4580, are located in South Carolina in US. The agri-
culture fields in heterogeneous districts are more regular in
shape and their size is larger than the homogeneous districts.
In addition, the agricultural fields comprise more than 45
percent of each heterogeneous district, while in the homoge-
neous districts, the agricultural fields comprise lower than 20
percent of the area. The patches in the heterogeneous districts
are distributed more evenly than the other districts due to its
lower dominance and percentage area of cropland (Table 1).
The National Agricultural Statistics Service (NASS)
Cropland Data Layer (
CDL
) have been extensively used in
various research projects (Wright and Wimberly, 2013) due to
88
February 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
