values for obtaining the finer-resolution upscaled maps does
not contribute much to the performance of
FMC
.
Furthermore, the determination of
FMC
on the optimal ex-
ponent values (Table 3) illustrates that the
CMP
did not make
a big contribution to improve the
OC
of the upscaled maps in
the most homogeneous areas. Considering the results shown
in Table 3 and Figure 9, for very homogeneous areas, we
recommend to use
τ
CLP
=1 and
τ
CMP
=0 as the optimal exponent
values for
FMC
to produce the upscaled maps. Nonetheless,
caution should be taken that the results of
FMC
using
τ
CLP
=1
and
τ
CMP
=0 may still not be the optimal results as further
experiments need to be explored to quantitatively determine
the definition of “very homogenous”. More work considering
the percentage of the dominant land cover using metrics such
as
PPU
,
SqP
, etc., is needed to judge if a study area should be
treated as very homogeneous when implementing the
FMC
.
Uncertainty Analysis of Upscaling
The effect of upscaling on the land cover change detection is
illustrated using
RE
(Figure 7), which shows that upscaling
leads to increasing in the underestimation and overestimation
of area for land cover types. For example, for the upscaled
maps at 960 meter resolution, the changes in the non-crop
map class in
ASD
4550 was underestimated about 14 times
more than the changes in non-crop estimated by the thematic
map at the 60 meter resolution when using
MRB
(Figure 7).
The reason for this big underestimation is due to the increase
in the proportion of non-crop (about 7.5 percent or 87709 ha).
The results of
RE
highlight the importance of reporting an un-
certainty analysis for application of upscaled maps for use in
various Earth’s observation models. Moreover, Earth science
researchers should be aware of the uncertainty/error derived
from upscaling.
Additionally, comparison of
RE
between
FMC
and
MRB
dem-
onstrates that
FMC
can reduce the negative effect of upscaling
on the
LCC
detection in both heterogeneous and homogeneous
areas (Figure 7). For example, when the thematic map was
upscaled to 60 meters, the
RE
of the corn map class change
was reduced from 15.85 percent to 7.87 percent by using
FMC
.
These results further strengthen our confidence to recom-
mend the
FMC
to be applied for generating upscaled thematic
maps for diverse Earth science models.
Limitations and Future Work
Our focus has been on fusing uncertainty information ob-
tained from land cover distribution and uncertainty informa-
tion from the base map to predict upscaled, coarse resolu-
tion agricultural maps. We have explored and compared the
performances of a new approach,
FMC
, with an established
approach,
MRB
. The results extend our understanding of using
uncertainty information to upscale thematic maps.
However, our research has several limitations. The first
major limitation is that
FMC
uses the
MRB
results to predict the
CMP
, which results in
FMC
being impacted by
MRB
, to some
degree. To solve this issue, two approaches could be explored
in future work: (a) simulate randomly coarse-resolution maps
to optimize these maps by integrating various measures of
uncertainty, and (b) employ different upscaling methods to
predict the initial upscaled maps for calculating the
CMP
.
The second major limitation is that
FMC
does not reduce
the changes in landscape pattern (it performs similarly to
MRB
) that are an area of concern in several research fields (e.g.,
landscape ecology modeling). Geospatial methods should
be adopted to improve the
FMC
for acquiring a better perfor-
mance both in reducing
PE
and lowering changes in landscape
pattern.
Another limitation is that we explored 101 combinations
of the exponent values to predict the
P
k
IMC
(
V
m
). In future
work, the optimization algorithm (e.g., simulated annealing
algorithm) should be applied to solve the issues for selecting
the optimal exponent values, which may provide different
results than we obtained in this paper. Additionally, to eas-
ily implement the
FMC
, the definition of homogeneous areas
should be quantitatively explored and determined in future
work. The exponents,
τ
CLP
=1 and
τ
CMP
=0, can be then directly
applied to these very homogeneous areas. More attempts need
to be performed to obtain this standard based on the various
landscape metrics.
Finally, to investigate the uncertainty of upscaling and as-
sess the
FMC
’s performance, a simple land cover change model
was selected. There is no doubt that Earth science researchers
use far more sophisticated models (e.g., carbon flux model)
than the simple land cover change model used here. There-
fore, future work should focus on the uncertainty analysis
of the applications of upscaled maps and further explore the
practical upscaling methods per the requirements of differ-
ent models using various land cover maps at wide range of
resolutions.
Conclusions
This study proposed a new upscaling method for thematic
maps (called
FMC
) based on the Tau model.
FMC
fused two
sources of the uncertainty information: land cover distribu-
tion (i.e., class membership probability,
CMP
) and uncertainty
information of the base map (i.e., confidence level probabil-
ity,
CLP
). The
CDL
data for 2016 at 30 meter resolution were
used to extract the agricultural maps as the base maps for
conducting an experiment on six study areas with different
landscape patterns (degrees of heterogeneity). The results
showed that: (1) The
FMC
approach obtained higher overall
consistency (
OC
) and lower proportional error (
PE
) compared
to
MRB
. (2) Although the results were not very different,
FMC
can reduce the negative effects of class proportion in the
base maps on upscaling. (3) The analysis of the relation-
ship between upscaling (i.e., spatial resolution for coarse
maps), heterogeneity, and the performance of
FMC
showed
that increasing either the upscale resolution or the hetero-
geneity of landscape has less influence on
FMC
compared to
MRB
. Additionally, the influence of the exponent values on
the performance of
FMC
showed that
CLP
contributes more to
improve the accuracy of upscaling. Further, we recommend
using
τ
CLP
=1 and
τ
CMP
=0 as the optimal exponent values for
FMC
to generate upscaled maps in very homogeneous areas.
Note that caution must be taken that
FMC
can probably not ob-
tain the optimal result when the exponent values,
τ
CLP
=1 and
τ
CMP
=0, are applied. Moreover, a simple and straightforward
land cover change (
LCC
) model was employed to investigate
the effect of upscaling on the
LCC
detection.
CDL
data at two
dates (2008 and 2016) were adopted to conduct the
LCC
. This
analysis highlighted that considerable attention must be paid
when employing upscaled maps to perform any Earth science
modeling. Also, the comparison of
FMC
and
MRB
in the
LCC
detection further confirmed
FMC
should be considered as an
effective way to reduce the uncertainty/error using upscaled
maps. Overall, the implications of these findings suggest that
FMC
achieved a more accurate result, which can be applied to
acquire upscaled thematic maps.
Future research should be directed towards the following
aspects: (1) Exploring algorithms to predict the
CMP
that is not
based on existing upscaling method, because that the existing
methods have influence on the
FMC
. (2) Multi-sources of infor-
mation, such as geospatial distribution, uncertainty informa-
tion, spectral information from imagery, etc., should be fused
to predict the coarse maps. (3) Optimization methods should
be explored to select the best exponent values for outputting
the coarse-resolution maps. (4) The effect of upscaling on the
98
February 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
