3. EXPERIMENTS AND ANALYSIS 
The land cover data used in this paper is acquired from the 
National Land Cover Dataset 2001 (abbreviated to NLCD2001) 
for conterminous United States. Sixty-five mapping zones and 
sixteen land cover types are involved in NLCD2001. All 
NLCD2001 products were generated from a standardized set of 
data layers mosaiced by mapping zone. Typical zonal layers 
included multi-season Landsat-5 TM and Landsat-7 Enhanced 
Thematic Mapper (ETM+) imagery centred on a nominal 
collection year of 2001 (Homer, 2007). All of the images are 
geo-registrated to the Albers equal area projection grid, and 
resampled to 30m grid cells. 
    
  
Open Water 
Forest 
Grassland/Shrub 
Barren/Sand 
Cropland 
Wetland 
red (c) 
Figure 2. (a): Land cover types in NLCD2001 database; (b) and 
(c): an ETM- image and six land covers of the test area 
Because all the images utilized in NLCD2001 are provided by 
U.S. Geological Survey (USGS) Centre for Earth Resources 
Observation and Science (EROS), a corresponding Landsat 7 
ETM- image was downloaded from EROS. It was acquired on 
July 13, 1999. After the registration with the land cover data, a 
study area ranging from the latitude of 47 4l' N to 47 54' N, 
and from the longitude 109'01' W to 10921' W was clipped 
from the ETM+ image. The clipped image is made up of 500 by 
500 pixels (Figure 2(a)). In our experiment, six classes 
including open water, forest, grassland/shrub, barren/sand, 
cropland and wetland are chosen from the NLCD2001 (Figure 
2(b)). The tasseled cap transformation is applied to the ETM+ 
image after which the two components of soil brightness and 
greenness are selected. 
Thanks to NLCD2011, the ground truth of each pixel in the 
image is known. Hence, 1500 training samples are random 
selected with others allocated as testing samples. The ratio of 
training samples to testing samples is rather small. The support 
vector machine (abbreviated to SVM) is utilized as the classifier 
in this paper. And the package libsvm interfacing with the 
statistical software R is adopted to implement K-class ( K » 2) 
land cover classification (Meyer, 2009). 
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 
   
For this Data set, the estimated Arif index is 0.411 which 
manifests a moderate separability and corresponding to the 
potential highest accuracy of 79.47%. Compared to the overall 
accuracy of 73.79% obtained by the SVM classifier, 79.47% 
not only reflects that part of the information is consumed by the 
classifier, but indicates an improvable accuracy of about 5%. Tt 
is easy to compute indicator transforms for training samples of 
known class labels. And, after prediction of the posterior 
probabilities pertaining to six classes by the SVM classifier, the 
residuals can be calculated as differences between binary 
indicators and predicted class probabilities. Table 1 lists the 
variogram models of simple kriging with local mean which 
reflect the spatial distribution of residuals of the training 
samples. And it also shows the vairogram models of the 
primary variable and the secondary variable, and the 
covariogram of the cokriging method. It exhibits the spatial 
variation of the target variable at the locations of training 
samples and testing samples (i.e., all pixels except for training 
samples in this experiment), respectively. The trend of the 
cokriging method in this paper is obtained by applying spatial 
smoothing to the posterior probabilities. 
In general, as is shown in Table 2, the cokriging method obtains 
a considerable improvement in overall accuracy and kappa 
coefficient, and simple kriging with local mean is no exception 
and even more effective. The former achieves 2 percents 
improvement in overall accuracy and an increase in kappa 
coefficient from 0.58 to 0.65, while the latter witnesses a 5% 
accuracy increase and an improved kappa coefficient of 0.68. 
The reason the SK method gains higher accuracies than 
cokriging may be that the trends of the primary and secondary 
variables of cokriging are obtained through smoothing in spatial 
domain, while the trend of the SK method is localized to each 
pixel. Therefore, the residuals are more accurate. Moreover, the 
latter demands less variogram models and thus costs less time 
for modelling. Therefore, the method of simple kriging with 
local mean is more worthy of recommendation for the fusion of 
input information and spatial information. 
Furthermore, the SK method is made as an example to account 
for the effects of kriging paradigm. Four groups of testing 
samples, each of which contains fifteen samples with ground 
truth as farmland but classified as other land cover types are 
randomly chosen and exhibited, shown in Figure 3. The 
posterior probabilities predicted by the SVM classifier and 
those revised by the SK method are compared in this figure. 
Generally, the posterior probabilities obtained by the classifier 
would first be corrected by residuals and then be normalized. 
However, in order to more clearly reveal the probability 
fluctuations before and after residual corrections, the 
normalization procedure was skipped over. In Figure 3, the 
abscissa represents the number of testing samples, while the 
ordinate denotes the probabilities. For the selected testing 
samples, the red circle € denotes probabilities pertaining to the 
land cover type of farmland after SVM prediction, while the 
black triangle A represents the highest posterior probabilities 
pertaining to the prevailing class type other than farmland after 
the prediction of SVM. Figures 3(a)-(b) illustrate that the 
classifier failed to make accurate predictions. Correspondingly, 
the reversed purple triangle ¥ denotes the probabilities 
pertaining to farmland after residual corrections by the SK 
method; while the blue squares æ represent the revised ones 
corresponding to the original black triangles A.