246
2. FEATURE EXTRACTION
Feature extraction and image interpretation are the most
time consuming tasks in photogrammetric mapping and
are regarded as the typical job of a human operator in a
traditional photogrammetric environment. A digital
approach offers the prospect for automation of these
tasks. There have been research activities in
photogrammetry, remote sensing and computer vision
areas in automated feature extraction (Albertz and Konig,
1991). Knowledge required for this include physics of the
imaging process, geometry and photometry of specific
objects and the spatial relationships and constraints
between objects (Quam and Strat, 1991). To date, full
automatic feature extraction is not operational, some
semi-automatic procedures or techniques have been
developed over the years. Forstner (1993) has classified
features into three categories: low level features which are
attributes of the pixel arrays of the images such as
spectral features used in multispectral classification, mid-
level features which are either geometric primitives such
as points, edges or regions or they are aggregates of these
primitives including relations, high level features which
are already interpreted, with meanings or labels attached.
In other words, the feature extraction can be characterized
as thematic information extraction or spectral
classification, geometric feature extraction and the
combination of both. In the field of remote sensing,
feature extraction algorithms were developed based on
spectral properties of the objects and their relationships.
In photogrammetry and computer vision fields, most of
the activities were focused in extracting features by
exploiting geometric knowledge of the objects under
investigation. In the following subsections, we review the
feature extraction methods used in these fields
respectively.
2.1 Feature Extraction in the Photogrammetry and
Computer Vision Context
Feature extraction from digital images involves two steps,
firstly to identify the objects by interpreting,
understanding and classifying the image, then to track the
objects by measuring the coordinates of the object
outlines. Usually, a geometric primitive (point, line or
region) is defined by an algebraic equation with a number
of parameters. The extraction algorithm looks for subsets
of points in the data set that lie on a geometric primitive
or close to it, in other words, the algorithm is essentially
looking for subsets with low fitting cost. Depending on
the context, additional constraints may be imposed on the
subset (Veelaert, 1997).
Edges carry most of the information in an image and are
relatively robust to changes in image contrast and
radiometry (McKeown, 1990). Therefore, edge detection
has been an important process in image processing,
pattern recognition and computer vision and it can be
achieved by detecting the maxima of the gradient or zero-
crossings of the second derivatives including the
Laplacian (Bennamoun, et al., 1997, Shen, 1996). There
are many detectors, such as Canny detector (Canny,
1986), developed over the past by researchers in the field
of photogrammetry and computer vision. Research in
both fields also showed that there exists no universal edge
detector which can be applied to a digital image function
to both identify and track edges with sufficient success
(Agouris and Stefanidis, 1996).
Attention has been given by the researchers in the
photogrammetry and computer vision fields to semi-
automatic or automatic linear feature extraction,
especially detection and delineation of roads. In the most
recent development in semi-automatic extraction of roads
from satellite and aerial images, a generic road model was
represented by using photometric and geometric
properties with several constraints and merit functions
(Gruen and Li, 1996, Li, 1997, and Trinder and Li, 1995).
Li (1997) has given an overview on some existing feature
techniques, and shows that most application presented in
the literature used black and white small scale aerial
images or single band satellite images, and no spectral
properties were taking into consideration. Trinder and
Wang (1997) have proposed a knowledge-based system
for automatic road extraction. Knowledge about the
objects was limited mainly to geometric properties. The
radiometric property applied was the average intensity or
gray value of the roads.
Extracting buildings has been another focused area for
photogrammetrists and computer scientists in the past. It
requires knowledge about the structure of built objects,
existing techniques of edge-line analysis, shadow analysis
and stereo imagery analysis to produce building
hypotheses, but no single technique can perfectly
delineate the structures in every scene. McKeown (1991)
identified the problems associated with building
extraction and pointed out that there is a need for
information fusion techniques and for incorporating
information of spectral properties into the extraction
process. However, no literature to date has presented such
applications . Multispectral images offer a richer dataset
image processing and interpretation, but they place
greater demands on technology and algorithm
development. Currently the application of multispectral
images has been limited to classification techniques in
remote sensing. Hence, most image understanding
procedures have not incorporated multi-spectral data
(Trinder and Sawmya, 1997).
2.2 Levels of Information Extraction in Remote
Sensing
Remote sensing technology has provided the capability of
extracting information about objects or surfaces on the
Earth’s surface and in the atmosphere. The level of
information extraction can range from manual image
interpretation using raw images, image interpretation
using ortho-rectified images together with existing base
maps, using enhancement, segmentation and
transformation techniques to improve the interpretability,
and digital classification using supervised or
unsupervised methods. More detailed information on
image enhancement, transformation and classification can
be found in Richards (1993) and McCloy (1995). Reed
and Du Buf (1993) gave a good review of segmentation
and feature extraction techniques. Segmentation
applications can be found in Ryherd and Woodcock
(1996), Dong et al (1997) and Dong and Forster (1998).
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
m 5" py em,
t" U t U» nm to
m^ o^ f^»
