The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
more information than coarse resolution imagery for detailed
observation on ground objects, single pixels no longer capture
the characteristics of classification targets, and the increase in
intra-class spectral variability causes a reduction of statistical
separability between classes with traditional pixel-based
classification approaches. Consequently, classification accuracy
is reduced, and the classification results show a salt-and-pepper
effect. Because of the characteristics of high resolution imagery,
lots of object-oriented image analysis was done and research
results prove the effectiveness of object-oriented image analysis
for information extraction from high spatial resolution imagery
(Baatz and Schape,1999; Blaschke et al., 2000; eCognition,
2002; Benz et al., 2004; Volker, 2004; Yan G et al., 2006).
An object-based approach to image analysis is composed of
four steps, which are multi-resolution segmentation to generate
objects resembling ground objects closely and to create object
hierarchy which allows a simultaneous representing of image
features of various scale and establishes a network allowing
relations between objects to be utilized, image object feature
extraction and parameter assessment which help to find useful
features and ways to separate classes, classification which uses
iterative steps to classify image objects, as well as accuracy
analysis and evaluation. In this research, object-oriented fuzzy
classification to optical imagery was performed in Definiens 5.0,
which is an object-based processing software from Definiens
Imaging GmbH.
2.2 Information Extraction from SAR Imagery Based on
Textural Features
Texture, a representation of the spatial relationship of grey-
levels in an image, is an important characteristic for the
automated or semi-automated interpretation of digital imagery.
SAR image is the representation of backward scattering
characteristics of ground objects to radar waves. If the ground
objects have same or similar backward scattering values, it is
difficult to get them separated, especially in the single band and
single polarization SAR imagery. In addition, because of the
influence of speckle noises, it is hard to distinguish ground
objects only by intensity values. Textural information should be
utilized (Guo et al., 2000).
Nowadays, different methods have been proposed for analysis
of image texture, and there is no general agreement on an
overall best analysis method, which outperforms all the others
on various tasks. In this research, we propose a multi-scale and
multi-texture feature fusion method based on SVM (Support
Vector Machine) for information extraction from single-band
and single-polarization SAR data by employing the multi-scale
textural analysis technique and the fractal analysis technique.
The method makes use of the ability of GLCM (Grey Level Co
occurrence Matrix) which uses spatial correlated characteristic
of grey values for texture description, takes the multi-scale
features of ground objects into account, and incorporates multi
fractal features which have the great capability in description of
complex spatial structure information and detailed texture
features; on the basis of SVM and by using feature-level image
fusion technique, it integrates seven textural features to realize
the land use/cover classification of SAR imagery. SVM is
introduced in this research because it is established on the basis
of statistical learning theory, and has favourable classification
performance in the feature space which is nonlinear, high
dimensional and has small samples. The seven-dimensional
textural features are:
Feature 1: Correlation feature with the processing window size
11;
Feature 2: Standard Deviation feature with the processing
window size 21 ;
Feature 3: Entropy feature with the processing window size 13;
Feature 4: Fractal dimension feature;
Feature 5-6: two multi-fractal features;
Feature 7: Second-order statistic lacunarity;
Where, the first three multi-scale GLCM features are obtained
according to multi-scale textural analysis, which can be referred
to Zeng et al.(2007a); the latter four features are based on
fractal theory and are obtained according to fractal analysis.
The calculation of fractal dimension feature can be referred to
Zeng et al.(2007b).
Improving the interpretation accuracy of SAR data is the basis
for change detection with optical imagery at the level of
information processing.
2.3 Soft-decision Change Detection Based on Rules
On the basis of above research, the change information can be
obtained by combining individually extracted information of
different temporal. Aiming at the change overestimation caused
by error propagation and the traditional hard-decision method
for change detection, this research proposes a soft-decision
method based on rules for change detection. In China, land
use/cover change usually occurs at the urban fringe areas, and
most land use/cover changes are the consequence of urban
growth. In this instance, we can think that change to built-up
area from other land use types is irreversible. With the rapid
economy development and population growth, the emphases of
land use change monitoring in China are urban expansion and
decrease of plantation areas. Taking the characteristics of land
use change at the urban fringe areas and the research focus into
account, this method employs a series of logic rules in turn to
evaluate the rationality of the detected changes. The logic rules
are made according to the status of pixels which are changed or
unchanged, land use change trajectories, as well as by taking
the spatial features of detected changes which are shape, size
and location into consideration.
Let Num denotes the number of detected changes. Land use
types are presented as Ci. Here, C! = “built-up area”, C 2 =
“water”, C 3 = “vegetation”, C 4 = “bare land”. The detected
change trajectory is denoted as j{C i ,,...)• Let N refers to “no
change”, W refers to “wrong classification and applying
masking”, and Y refers to “correctly detected changes”. These
logic rules are:
Rule 1 : if Num = 0 THEN N.
Rule 2: jp Num = \ AND r(C,.,C y ) (i * 1; / * j) THEN Y.
Rule 3: IF Num = \ AND T(C t ,Cj) 0 *1) THEN W.
Rule 4: River, lake, canal and its affiliated works are regarded
as N.
Rule 5: Land use types in between large area of cultivated lands
are regarded as N.
Rule 6: Isolated 3x3 detected change regions are regarded as W.
By employing these rules in turn to analyze the rationality of
the detected changes, the unchanged areas, falsely detected
changes, and possible changes can be separated. The change
overestimation can be decreased and the change detection
accuracy can be improved.
1078