International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
Table 1. Accuracy evaluation of mixed-pixel spectral decomposition of the CNP MODIS imagery
Sensor Name Spatial resolution Mean Variance absolute error Relative Error
(m) (%) (%)) (%) (%)
EOS-MODIS 250 24.1295 59.1619
CBERS02-CCD 19.5 23.9752 29.8883 A tac
gm me ] and significant degree of correlation, etc.) of a study region,
0. 0050 IE Rs based on the Moran's / and Geary's C and semivariogram
0040 A analysis. In this study, the corresponding absolute change rates
0030 — nme — — |a GearysC, | of the Moran's /, Geary's C and semivariogram representing the
£7 0.0020 EG CNP's spatial structure characteristics have the decreasing trend
0. 0010 Re T along a horizontal axis direction (See Figure 3), whereas there
00007 > AT exists a greatly distinct turning point nearby at a lag of 7.5km
0 0.5 1 1.5 2 2.5 3 . . * ; . .
Lag (km) and then their changes in size is very small, which indicates that
Figure 2. First order differences of means of local spatial
autocorrelation (Getis ord G;, Moran's I;, Geary's C;) under
difference spatial-sampling grains in the CNP, based on the
fraction of winter-wheat planting area in each pixel (RIP/RIV)
that is an important reference factor for sample point size
design of spatial sampling. In addition, it was thinking of the
spatial structuring and resolution of available MODIS images of
the CNP and real feasibility of spatial sampling that the scale
750mx750 was determined as an optimal (optimum) sample-
point size in spatial sampling of winter wheat area estimation of
this study. This is consistent with the sample-point size of
500mx500m that is adopted currently in the actual cropland
area remote sensing operation monitoring of winter wheat in
North China plain. Given the research results, we may
appropriately improve the existing sampling designs and
implementing solutions so as to reduce the cost of spatial
sampling for crop remote sensing monitoring and improve the
spatial sampling efficiencies.
4.2 Determining Sample Point Distances
Distances between sample points (namely, sample-point distances)
are important elements of spatial sampling design and can set up a
set of sample points in all sampling space to a certain extent to be
characterized and determine the implementation characteristics of
corresponding spatial sampling. For example, distances of sample
points, if too small, are likely to present the strong spatial
correlations and then lower the adequately random characteristics
of samples which are necessary of probability spatial sampling;
being under certain sample size conditions, if too large, there are
some difficulties to lay sample points in a sampling space, or,
deficiencies that aren't able to effectively use a priori knowledge or
information (e.g., its spatial structuring) or fully represent the entire
population features with samples (even though only using simple
probability sampling methods). Hence, there are inevitably larges
deviations between the sampling results and real populations and
they further lower the efficiencies of spatial sampling. As a result,
we should not only settle rational sample-point sizes but also
arrange them being divided with reasonable intervals (i.e., sample-
point distances), especially based on the minimum optimum
intervals that are derived from spatial structuring in a sampling
population space and the most foundational control requirement of
implementing spatial sampling.
Therefore,
global spatial autocorrelation analysis of the
RIP(s)/RIV(s) of objects of sampling space is an useful method
to explore the total spatial structure characteristics (including
average correlation of and spatial distribution pattern of objects,
sensitivities of this change trend are falling down along with
increasing lags (i.e., separating distances among spatial objects).
Besides, another distinct turning point appears at a lag of
22.5km, whereafter this trends to a more random stationary
status. Consequently, we could settle the separating distances
(from about 7.5 to 22.5km) as the sample-point distances
because they were able to reach the pre-requisite minimum
distances among sample points to reduce their spatial
correlations to enough small degree and satisfy probability-
sample random of and meanwhile, simultaneity certain
accuracy requirements of spatial sampling. The results are
consistent with the set sample-point distances (about 20~30km)
of the current large-area operating spatial sampling survey of
crop remote sensing monitoring in China North.
0. 06 Lower optimum minimum upper optimum minimum
sample-point distance ample point distance
and
t i 1
Aa, | Gear's C increment i
We tt roe, gone
Y rade gente grep ptt rea dt etit
0.02 9 10 15 20 1 25 30
0. 04 Moran's increment
-0. 06 *
(Distance Unit km)
(a)
Semivariance(^1) - Semivarianceti)
Lower optimum minimum
samplc-point distancc
|
upper optimum minimum
sample-point distance
0 5 10 15 20 25 30
(Distance Unit: km)
(b)
Figure 3. First order differences of (a) global spatial
autocorrelation (Moran's I, Geary's C) and (b) semivariance
with sample-point separating distances, respectively
4.3 Spatial Sampling by Stratifying Spatial Autocor-
relation Statistics
Given that we determined 750mx750m as an optimum pre-
sample-point scale in the CNP, now the RIP/RIV (ie,
percentage of winter wheat planting area in each pixel)
distributing maps in spatial resolution of 750mx750m served as
the basic operated data by aggregating the baseline MODIS
images (with spatial resolution of 250mx250m). Through local
spatial autocorrelation analysis with the Moran’s /,, Geary's C;
and Getis ord G; (or 6;, the results were obtained and
appropriately stratified, respectively. We could thus determine
the corresponding average minimum sample-point distances of
