An object-oriented pattern recognition approach for urban classification - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 122

9.1 Introduction

The purpose of this chapter is to introduce readers to the prin-
ciples of OBIA and demonstrate how it can be applied to achieve
satisfactory accuracy in urban LULC mapping. We employed
two case studies with two example subset images extracted
from Quickbird multispectral satellite data and demonstrated
two object-based classification approaches, namely rule set (i.e.,
membership function) and nearest neighbor classifiers. These
classifiers are supported in the Definiens Developer 7.0 OBIA
software. Background is provided from the Definiens software
perspective. The software package uses a region-based, local
mutual segmentation routine, a type of region growing approach
to generate image objects or segments, prior toperforming object-
based classification (Baatz and Schape, 2000; Benz et al ., 2004;
Yu et al ., 2006).

Remotely sensed image data have been used extensively for many
years for the identification and mapping of land-use and land-
cover (LULC) classes for urban environments. Image analysis
techniques such as object-based image analysis (OBIA) have
been developed in the recent past to facilitate semiautomated
mapping of LULC classes. For instance, the Definiens commercial
OBIA software utilizes image segmentation and object-based
classification to delineate and classify objects within an imaged
scene. Such software automates processes and incorporates expert
knowledge to enable object-based classification. A further benefit
of partially automating the OBIA process is that once a processing
sequence has been created, a model incorporating this sequence
can be distributed to and used by others, and only site-specific
calibrations are normally necessary to achieve comparable and
consistent results.
The human mind is extraordinarily adept at pattern recogni-
tion, as when extracting relevant image-derived information such
as urban LULC classes fromremotely sensed data. Ahuman inter-
preter can recognize and identify a large number of urban objects
such as roads, buildings, or LULC polygons using tone/color,
texture, contextual, size, pattern, orientation, height, and shape
information containedwithin fine spatial resolution satellite data.
These are traditionally known as basic elements of image inter-
pretation (Lillesand, Kiefer and Chipman, 2008). While our eyes
and brains can distinguish the difference among different types
of LULC sharing similar spectral responses (e.g., cement roads,
sidewalks, and driveways), these tasks have been challenging for
traditional single or per-pixel classifiers that attempt to identify
LULC classes in a more automated fashion.
Attempts to perform more automated and effective image
analysis and information extraction from digital image data have
been pervasive for a few decades. Yet, fundamental advances in
automated or semi-automated digital image analysis remain a
challenge, especially when classifying detailed LULC classes from
fine spatial resolution data. Particularly challenging is the gener-
ation of computer algorithms that perform in the same manner
that the human brain functions to extract image information.
Object-based approaches to semi-automated LULC mapping
have been a major focus area of remote sensing and image
processing research in the past decade. Relative to single- or per-
pixel approaches, OBIA attempts to exploit spatial relationships
of groups of pixels in order to delineate and identify objects within
an imaged scene (Benz et al ., 2004). Object-based approaches are
most applicable to the high or H-resolution remote sensing scene
model (Strahler, Woodcock and Smith, 1986), where objects of
interest are larger than the ground resolution element associated
with a pixel. Such objects may be related to natural features of
urban landscapes (e.g., trees and lakes) or human-made features
(e.g., buildings or roads). Object-based image processing tech-
niques have been developed to support environmental remote
sensing for over 30 years (Ketting and Landgrebe, 1976). How-
ever, a greater research emphasis on such techniques has occurred
in the past several years due to: (1) the greater availability of dig-
ital remote sensing image data having fine spatial resolution
that are generally not amenable to achieving highly accurate
mapping and monitoring results when generated with per-pixel
image classification routines, and (2) the greater availability and
affordability of high performance computers and object-based
image processing software.

9.2 Object-oriented
classification

As explained earlier, the object-oriented classification employed
to map urban LULC classes in this chapter focuses specifically
on Definiens software known as eCognition and the various rou-
tines it contains to achieve OBIA. Here we provide background
on the components of an OBIA approach to LULC classifi-
cation, while describing some of the specific processing steps
and parameter choices associated with the Definiens software
environment. Even though this chapter is specifically centered
around the use of Definiens software package, an overview of
object-oriented classification approach, consideration of many
different features and parameters, image segmentation proce-
dure in relation to scale levels, overview and general procedure
to perform a rule set approach, nearest neighbor classifier using
training samples, and limitations and uncertainties associated
with object-based classification technique are useful and applica-
ble to any image processing software that contains object-based
image analysis functions.

9.2.1 Image segmentation

The first step in the process of an object-oriented analysis is to seg-
ment an image, representing a scene or the ground area covered
by the image extent, into image objects (Lee and Warner, 2006;
Stow et al . 2007; Im, Jensen, and Tullis 2008). An image object
is generally defined as a group of pixels sharing similar spectral
and/or textural properties (Navulur, 2007). When we interpret a
particular subset of the image, our brain tends to focus-in on small
areas or patches consisting of homogeneous tone/color or texture
that may have characteristic shapes or sizes. Hence, visually such
portions of the image are recognized as objects. For example, we
see and can delineate a green rectangular area and then identify it
first as landscaped grassland, and then in its urban context, as the
lawn of a park or an institutional complex. Similarly, a house, a
tree, a groupof trees, anopen cement parking lot, a street segment,
or a swimming pool can be considered as image objects. Thus,
an image object is a group of connected pixels having similar or
characteristic image properties. Each image object represents a
specific, small region in an image, and an urban image contains
many different objects. Image segmentation is a primary function