International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
The objective of this paper is to develop and test a multi-
resolution classification framework for selecting and integrating
information from different spatial resolutions to improve land
use/cover classification. The multi-resolution framework is
based on the development of analytical techniques and strate-
gies of selecting and integrating suitable information from dif-
ferent resolutions into classification routines. The following il-
lustrates the multi-resolution classification framework.
2. MULTI-RESOLUTION IMAGE ANALYSIS AND
CLASSIFICATION FRAMEWORK
A typical computer-assisted classification involves six major
steps: classification scheme design, feature transformation,
training, application of classification decision rules, post-
processing, and accuracy assessment. Spatial autocorrelation in-
fluences many aspects of image classification, depending on
sensor resolution and landscape fragmentation. The entire clas-
sification process, especially the selection of training data, ap-
propriate resolutions, classification algorithm, and sampling
data for accuracy assessment, may be affected by the autocorre-
lation of neighboring pixels. A proposed multi-resolution image
analysis and classification framework is based on the frame-
work developed in Chen and Stow. (2003) and presented in
Figure 1.
mevounsnsa sn ss en aGHAUUUNANAUOUNUNANAUAUOUAULS HA HO UAUUSUAURENHASAU AU SL ANCHSUAUO NAS CUUCU CUS SCA CU ALU CL CUES
.........>...>...——>N
Feature representative ( Establish
sites for each class I4—| Classificatio
^ n
Error modeling and Multi-scale
accuracy classification decision
Figure 1. A proposed multi-resolution analysis framework. Rn
stands for variable image spatial resolution.
The matic
map
The framework proposed here, focuses on the examination of
image pattern, structural, and autocorrelation using different
multi-scale spatial analytical techniques in order to select ap-
propriate methods in different stages of classification such as
training strategy, feature extraction, scene models, and classifi-
cation accuracy. Spatial analysis techniques for measuring the
pattern size and degree of autocorrelation are computed for
each training class in order to determine whether they can guide
selection of training data, high-resolution or low-resolution
classification models, the range of spatial resolutions used for
classification, and error patterns. The following sections de-
scribe the major parts of Figure 1.
2.1 Image resolution pyramids
Multi-resolution images can be created in two ways: (1) by in-
tegrating different resolution images acquired by different sen-
sors; and (2) aggregating fine resolution images into: different
coarse resolution levels (i.e., image pyramids). Obtaining im-
1188
ages of different resolutions from different sensors could have
advantage of including more spectral information that can be
used to identify different objects, but is expensive. The misreg-
istration between different images also would increase the
processing cost and reduce classification accuracy.
It is more efficient to extract spatial information over a range of
resolutions from a single high resolution image or degradation.
There are several types of methods used in the operation of ag-
gregation including simple aggregation methods (e.g. averag-
ing, central-pixel, median, sub-sampling, nearest neighbor, cu-
bic convolution, etc.), scale-space transform (Lindeberg 1994),
and Wavelet decomposition (Mallat 1989).
2.2 Multi-scale analysis
The purpose of multi-scale analysis is to establish a statistical
model describing relationships between variables measured at
different resolutions. The proper application of image classifi-
cation procedures requires knowledge of the variables of the
data to determine the appropriate classification methodology
and parameters to use. The concept of spatial autocorrelation
has been introduced as a basis for understanding the effect of
scale. Spatial autocorrelation is an important factor in selecting
1) appropriate training methods for different homoge-
nous/heterogeneous classes (Chen and Stow, 2002), 2) appro-
priate spatial resolutions and image scene models (L- or H-
resolution), 3) suitable classification methods and parameters;
and 4) understanding classification errors at the different spatial
resolutions (Foody 2002).
Prior to performing image classification, exploratory spectral-
radiometric data analysis and visual assessment of the spatial
characteristics of each image band is recommended. Explora-
tory spatial data analysis should be employed to determine re-
quired information on image spatial characteristics and to en-
sure that appropriate scene methods and classification
parameters are used.
Many spatial statistical measures have been used to establish
the spatial characteristics and scene model characteristics of
images and to assess the scale of spatial variation in remotely
sensed data (such as local variance (Woodcock and Strahler,
1987), multi-scale spatial variability (Myers, 1997), and Fractal
analysis (Emerson et al., 1999). Several studies have explored
spatial autocorrelation measures to examine the autocorrelation
of pixel DNs and to determine the optimal spatial resolution of
a remotely sensed application (Atkinson 1997).
2.3 Multi-scale classification decision rules and algorithms
Three strategies were developed for maximum likelihood
classifier by Chen and Stow (2003) to exploit information
obtained from different resolutions and thus, to improve the
classification results.
The first strategy is a simple means of using information from
multiple resolutions by incorporating them simultaneously in a
classification routine. In this way feature measures obtained at
various resolutions are merged. This approach is simple and no
other algorithms are needed to organize the data. The major
drawback is that computation cost may be high
The second strategy is to compare the posteriori probabilities
obtained from different resolutions. For this approach, the clas-
sifier is applied at each resolution to obtain the probability for
