ial
ng
ng
he
ial
ns
pa
ith
int
ng
nd
he
nd
re.
ler
ity
he
he
he
ng
he
he
ter
he
ly,
les
K
CS
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
SK Cokriging
Class Variogram Variogr am of Marlogram of Covariogram
primary variable secondary variable
Open water 0.0007 Sph(30) 0.001 Sph(30) 0.0003 Sph(30) 0.0003 Sph(30)
Forest 0.046 Exp(21) 0.044 Exp(58) 0.011 Exp(60) 0.005 Gau(250)
Grassland/Shrub 0.012 Gau(37) 0.122 Exp(44) 0.018 Sph(50) 0.008 Exp(100)
Barren/Sand 0.130 Gau(53) 0.130 Exp(80) 0.011 Gau(65) 0.005 Gau(180)
Cropland 0.056 Gau(90) 0.056 Exp(430) 0.0009 Exp(75) 0.0002 Exp(100)
Wetland 0.019 Exp(92) 0.019 Gau(100) 0.001 Gau(200) 0.0007 Gau(100)
Table 1. Variograms and covariograms of SK and cokriging
Method SVM Classifie Cokriging Simple Kriging with Local Mean
Class Producer’s User’s Producer’s User’s Producer’s User's
Accuray Accuray Accuracy Accuracy Accuracy Accuracy
Open water 37.29 53.14 24.41 48.65 31.86 50.81
Forest 78.95 84.17 76.52 81.59 71.95 83.12
Grassland/Shrub 57.36 66.33 58.86 64.91 61.29 64.47
Barren/Sand 91.96 74.51 86.25 82.76 87.53 81.82
Cropland 15.92 59.75 73.28 76.11 72.80 73.93
Wetland 10.32 41.80 35.98 20.51 27.16 54.83
Overall Accuracy 73.79 75.77 77.02
Kappa Coefficient 0.58 0.63 0.65
Table 2. Accuracy Assessment Indexes of SVM Classification and Two Kriging Methods (Accuracy Unit: %)
a 1
a i
iv v o a
wv 2 x d i - i
A A
LA aay 4 ix A a A 4 & ^
* m A À a A |
A A * i
2 * y A "i = à A f
e LA = u 9 = = 2 a $ |
® = v ou t
* z * + 77 Y |
» P s A i x + + av :
+ Begg’ Eg d v AY a b
S $ " #- Ha + e: Wee e ron +»
a E
alg ® ® |
= =
T T T : T 1 7 T T f ? EU 7
2 4 6 8 0 12 A 2 4 6 8 10. (2. 4
(a) Classified as non-farmland; revised as farmland (b) Classified as non-farmland; revised as farmland
v
v E
vov $ sn
qa i j
S. o. 9 act P5. " Ea
7 Y \ i ®
> Fon i { = * ÿ
2 PO ? & €. i } % rate
wm / % x Ly = à i mi x: d f
X / : j X v og X = B. - d 3 wo" v / v : 8
\ í bf 7 » +. s * f
. ; tit o ; ;
a % / 4 ! » + ¥ + Y *v v
% i m mu ¥ w -
* f à a
x > $
2 T T T T T T t ; n ç T ¥ :
2 4 8 8 We 42. M 2 4 6 8 1024. 14
(c) Classified as farmland; revised as farmland
Figure 3. The Probabilities of the SVM Classifier and the SK Method (e.g., Farmland)
On the whole, Figures 3(a)-(c) reflect the effect of the SK
method, i.e., it utilizes the information of spatial distribution
provided by training samples to improve the posterior
probabilities pertaining to the target land cover type and
accordingly reduce those pertaining to the confusing types.
Take Figure 3(a) for example. Given the testing samples with
(d) Classified as farmland; revised as non-farmland
ground truth as farmland, the predicted posterior probabilities of
farmland are less than those of other confusing types, which
results in omission. However, after the residual corrections, the
probabilities of farmland are improved with a relative
probability decrease of other types, which may also be reflected
by the producer’s accuracy in Table 2. In addition, as is shown