9
MULTISCALE SEGMENTATION AND
CLASSIFICATION OF REMOTE
SENSING IMAGERY WITH ADVANCED
EDGE AND SCALE-SPACE FEATURES
ANGELOS TZOTSOS, KONSTANTINOS KARANTZALOS, AND
DEMETRE ARGIALAS
9.1 INTRODUCTION
The current need for automated image analysis and computer vision technological
tools requires a processing scheme able to encapsulate effectively the content of
remote sensing data. However, Earth’s landscape structure is complex, the context
varies, and so does the appearance of the images, being a combination of many
different intensities, representing natural features such as vegetation, geomorphological and hydrological features, man-made objects (e.g., buildings, roads),
and artifacts caused by variation in illumination of the terrain (e.g., shadows).
Furthermore, roads, infrastructure, vegetation, landforms, and other land features
appear in different sizes and geographical scales in images (e.g., country road vs.
interstate, tree stands vs. forest, maisonette vs. polygon building, rill vs. river). Only in
a few special circumstances do the objects of interest belong to a certain scale while
the remaining ones, to be discarded, belong to another. In most cases such a global
scale threshold is not possible since the desired information is present at several scales
(Witkin, 1983; Lindeberg, 1994; Weickert, 1998; Meyer and Maragos, 2000).
Scale-space representations and multiscale image analysis provide the framework
to explore the entire image content by detecting the scale(s) at which objects or
patterns appear and are most distinctly identified. Toward this end and parallel to the
170
Scale Issues in Remote Sensing, First Edition. Edited by Qihao Weng.
Ó 2014 John Wiley & Sons, Inc. Published 2014 by John Wiley & Sons, Inc.
MULTISCALE SEGMENTATION AND
CLASSIFICATION OF REMOTE
SENSING IMAGERY WITH ADVANCED
EDGE AND SCALE-SPACE FEATURES
ANGELOS TZOTSOS, KONSTANTINOS KARANTZALOS, AND
DEMETRE ARGIALAS
9.1 INTRODUCTION
The current need for automated image analysis and computer vision technological
tools requires a processing scheme able to encapsulate effectively the content of
remote sensing data. However, Earth’s landscape structure is complex, the context
varies, and so does the appearance of the images, being a combination of many
different intensities, representing natural features such as vegetation, geomorphological and hydrological features, man-made objects (e.g., buildings, roads),
and artifacts caused by variation in illumination of the terrain (e.g., shadows).
Furthermore, roads, infrastructure, vegetation, landforms, and other land features
appear in different sizes and geographical scales in images (e.g., country road vs.
interstate, tree stands vs. forest, maisonette vs. polygon building, rill vs. river). Only in
a few special circumstances do the objects of interest belong to a certain scale while
the remaining ones, to be discarded, belong to another. In most cases such a global
scale threshold is not possible since the desired information is present at several scales
(Witkin, 1983; Lindeberg, 1994; Weickert, 1998; Meyer and Maragos, 2000).
Scale-space representations and multiscale image analysis provide the framework
to explore the entire image content by detecting the scale(s) at which objects or
patterns appear and are most distinctly identified. Toward this end and parallel to the
170
Scale Issues in Remote Sensing, First Edition. Edited by Qihao Weng.
Ó 2014 John Wiley & Sons, Inc. Published 2014 by John Wiley & Sons, Inc.