A MULTI-RESOLUTION ANALYSIS AND CLASSIFICATION FRAMEWORK FOR
IMPROVING LAND USE/COVER MAPPING FROM EARTH OBSERVATION DATA
DongMei Chen
Department of Geography, Queen’s University, Kingston, ON K7L 3N6, Canada — chendm@post.queensu.ca
WG VII/1 Fundamentals Physics and Modeling
KEYWORDS: Land use, Land cover, Classification, Analysis, Information, Multiresolution, Multisensor, Multispectral.
ABSTRACT:
The last few years have seen satellite platforms with a large number of sensors (e.g. Terra and ENVISAT) coming on-line and the
launching of a huge number of satellites with more than one sensor (e.g. IKONOS and QuickBird). Various satellite images with
spatial resolutions ranging from 0.5 to 25,000 m are available for different applications. This development offers new and significant
changes and challenges in the approach to analysis, integration, and the efficient spatial modelling of these observation data. This
paper presents a multi-resolution analysis and classification framework for selecting and integrating suitable information from
different spatial resolutions and analytical techniques into cl
examination of image structural using different spatial analytical
assification routines. The proposed framework focuses on the
techniques in order to select appropriate methods in different stages
of classification such as training strategy, feature extraction, scene models, and classification accuracy assessment. The multi-
resolution approaches are tested using simulated multi-resolution images from IKNOS data for a portion of western part of the
Kingston Metropolitan area. It was demonstrated that the multi-resolution classification approaches can significantly improve land
use/cover classification accuracy when compared with those from single-resolution approaches.
1. INTRODUCTION
Earth observation data at multiple resolutions have been widely
used in studies of environmental changes, natural resource man-
agement, and ecosystem and landscape analysis in general.
With the development of new remote sensing system, very-high
spatial resolution images provide a set of continuous samples of
the earth surface from local, to regional scales. The spatial reso-
lution of various satellite sensors ranges from 0.5 to 25,000 m
now. Furthermore, high resolution airborne data acquisition
technology has developed rapidly in recent years. As an in-
creasing number of high resolution data sets become available
such as Digital Globe (Quickbird), Space Imaging (IKONOS),
Orbimage, Indian Remote Sensing (IRS), Digital Orthophoto
Quarter Quads (DOQQ), etc., there is an increasing need for
more efficient approaches to store, process, and analyze these
data sets. The development of efficient analysis methods of us-
ing these multiscale data to improve land use/cover mapping
and linking thematic maps generated from high resolution to
coarse resolution has become a challenge (Foody, 2002).
The effects of spatial resolution on the accuracy of mapping
land use/cover types have received increasing attention as a
large number of multi-scale earth observation data become
available (Dungan, 2001). Scale variation and sensitivity have
played an increasingly important role in the employment of
earth observation data in different application areas (Marceau et
al, 1994; Atkinson and Curran, 1999; Chen et al., 2003). For
example, the resolution range to identify an individual tree is
much smaller than that to identify a large commercial building
block. Spatial autocorrelation existing in each class is an impor-
tant factor influencing classification results at each resolution
level (Chen and Stow, 2003). Although many methods of semi-
automated image classification of remotely sensed data have
been established for improving the accuracy of land use/cover
classification during the past forty years, most of them were
employed in single-resolution image classification. Due to the
more heterogeneous spectral-radiometric characteristics within
land use/cover units portrayed in high resolution images, many
applications of traditional single resolution classification ap-
proaches have not led to satisfactory results (Barnsley and Bar,
1996; Chen et al., 2003).
Several techniques have been employed to assess appropriate
(or optimal) spatial resolutions. Although a particular classifica-
tion can achieve the best result from a single resolution appro-
priate to the class, there is no single resolution which would
give the best results from all classes (Marcean et al., 1994).
Clearly landscapes are characterized by multiple scales of spa-
tial heterogeneity. Landscape objects (e.g. land cover/use poly-
gons) are not the same size and vary in different structures.
Some objects are better classified at finer resolutions while oth-
ers require coarser resolutions. Therefore, as suggested by
Woodcock and Strahler (1987), various objects require different
.analysis scales according to the image scene model. Scene
models may be either high (H) resolution with pixels smaller
than objects, or low (L) resolution with pixels larger than ob-
jects to be mapped. From a practical standpoint, building a
framework to represent, analyze and classify images repre-
sented by multiple resolutions is necessary in order to capture
unique information about mapped classes that vary as a func-
tion of scale.
Many previous studies show the importance of developing and
evaluating spatial analytic methods and models to support mul-
tiscale databases (Emerson et al. 1999, Li et al. 2000). The ap-
plication of multiscale or multi-level approaches to earth obser-
vation data research, however, is very recent and remains
limited and undeveloped. Several researchers (such as Solberg
et al., 1996; Li et al., 2000) have devoted considerable effort to
the development of methods to integrate and analyze multi-
sensor, multi-scale and multi-temporal satellite imagery. How-
ever, compared with rapidly expanding data sets, there is an
obvious lag in the development of spatial analytic methods and
models for handling the increased multi-resolution images
(Quattrochi and Goodchild, 1997; Tate and Atkinson, 2001 ):
1187
