ons 
G) 
sity 
ing 
cult 
the 
fier 
the 
sity 
)6), 
of 
sity 
^M) 
lass 
to 
lass 
the 
(6) 
«ov 
(7) 
we 
(8) 
(9) 
    
where ó,- Dirac-delta function, if k-0, ój-1; kZ0, 0,0. 
c7 (c|c€ C, s€ c). 
Therefore, from (5) and (7) we know: 
Ws) » argmin Y (-In p(X | w(s))+ BY (1-6, (5,)] 
w(s) ses ceC 
(10) 
3. EXPERIMENTS 
3.1. Study Area and Database 
The measures reports in this study are conducted during the 
Watershed Airborne Telemetry Experiment. The study area 
locates in Grass Station of Lanzhou University in Zhang Ye 
district, Gansu province. Its geographical coordinates are 
39.25043?N, 100.005871?E, the altitude is 1385 meters. Land 
use mainly consists of country, bare salinization land and 
irrigative agricultural fields. The field experiment was 
conducted from June to July in 2008, at which time the crops 
were corn, clove, barley and other crops. 
Satellites over the study area provided TM and ASAR data 
on 7 July 2008 and 11 July 2008, respectively. ASAR 
(Advanced Synthetic Aperture Radar) is a synthetic aperture 
radar carried by the ENVISAT-1 satellite and operates in the C- 
band (central wavelength 5.63 cm), with multi-polarization, 
seven observation angles and five operating modes. In this 
study, we chose to use the ASAR data, and the operating mode 
was Alternation Polarization corresponding to two kinds of 
polarization (VV and VH) and high space resolution (12.5x12.5 
m per pixel). Figure 1(a) and 1(b) illustrate the false color 
composite image composed by TM3, 4, 5 and ASAR image in 
VV polarization. 
    
T ss fais E: i d 
(a) The false color composite image (b) ASAR image in VV 
polarization 
Figure 1. The images used in the paper 
When the initial class labels and conditional probability 
density function of the multi-source remote sensing data are 
determined by MLMM, formulation (10) is used to perform the 
local minimization at each pixel in a specified order and get the 
updated category. If changes occur then repeat estimating. The 
iteration continues until no more updates occur for all the pixels 
inside the lattice, then, the classification completes. 
Comparing with the conventional Markov model for iterative 
classification, our method needn't to assume the conditional 
probability density function in advance; With joining the spatial 
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 
correlation of the class labels, our method also can get a better 
classification accuracy than the ordinary maximum likelihood 
classification model. Moreover, the classification is achieved 
through the iterative process, which takes into account the 
characteristics of pixel attributes. 
3.2. Classification experiment 
VV, VH polarization of ASAR and all the TM bands are taken 
as the input of the classifier, and the multisource images are 
resampled to 30m*30m and geometrically corrected. The study 
area is separated to 12 classes, which are corn, other corps, 
garden, woodland, meadow, fallow land, sand, mountain, saline, 
desert, building, and water. The training samples and validation 
samples are shown in Table 1. When the training samples and 
validation samples are selected, we use the method in Section 2 
to classify the TM and ASAR images, the results are shown in 
Figure 2, In which, the basic form of oasis is similar with the 
Figure 1(a), which is consistent with the dual ecological 
environment of western semiarid regions, “oasis accompanies 
with water, and desert accompanies with no water”. 
Fhe Classification of Heihe Bivel By ASAR and TM 
dhesest 
taciding 
water 
Figure 2. The Classification map of study area 
i} ; 
LM tomes 
3.3. Validation 
To verify the necessity of coupling optical radar data for 
classification, the output of the classification of ASAR and TM 
in Section 3 were compared with the classification only by TM, 
all using Bayesian and MRF classifier. Table 2 presents the 
statistical errors among the three algorithms. 
In Table 2, the accuracy of each type of classification with 
single TM is lower, and reaches a total precision of 77.9%. 
When the ASAR dual polarization is jointed, the total precision 
increases to 89.496. The reason may be that the ASAR 
information can increase the surface characteristics and make 
them easy to distinguish, for example, corn, other corps, garden 
and woodland are similar in spectrum, and we can identify them 
by their various structural features revealed by their ASAR 
backscattering coefficients and finally obtain a better accuracy.