ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS”, Bangkok, May 23-25, 2001
280
analysts have to manually process many critical tasks related
to image processing, feature extraction and feature delineation
and so on. There are many problems associated with these
manual or semiautomatic process such as time-consuming,
inconsistency, and being difficulty to apply them to large-scale
and global information systems. Therefore, a reliable and
automated change detection system is required.
3.A framework for automatic change detection
This research is designed for the development of a
framework for automatic change detection. Based on this an
automated spatial change information extraction system from
various geo-spatial data sets will be created. This study has
four interrelated subsystems as shown in Fig 1. In the first
subsystem, we proposed to develop an automated image
registration subsystem that registers images (or image to map)
accurately and quickly with little or no operator supervision.
After automatically registered images, a automated change
finding subsystem, which takes multi-sensor and/or multi
temporal, space-borne and/or air-borne, remotely sensed
imagery or map and other data sources as its inputs, and
finding changed area would be developed. In this subsystem,
the output change information is divided into two groups:
certain information (the changes with minimal uncertainties
that can be easily extracted based on spectral and perhaps
some other image characteristics) and uncertain information(
the changes involving high level of uncertainties).The
uncertain information is then presented to the second
subsystem: automatic change feature extraction and
identification subsystem in which changed image
features(edge and area)are extracted. The changed
information entered into the fourth subsystem: intelligent
change recognition subsystem that is built upon knowledge-
based techniques utilizing knowledge stored in the database.
In this subsystem, the changed magnitude, changed type will
be recognized based on some recognition algorithms. The
procedures for reducing uncertainties including verification and
accuracy assessment of both certain and uncertain changes
are explicitly built into the proposed methodology. This is the
fifth subsystem. Finally, the identified and recognized changed
features are transmitted to information update system for data
update and visualization for decision and policy-making
purpose
Fig 1 .A Framework for automated change detection system
3.1 Automated image registration subsystem
In automated change finding subsystem, the impact of
misregistration on remotely-sensed change detection is
quantitatively investigated and potential techniques to remove
or reduce this impact is developed. Because of anticipated
large data volume and high data rates of these current and
future high-resolution sensors, the traditional approach of
visual identification of tie-points to register multi-temporal and
multi-sensor data is not an acceptable solution. The elements
of this procedure include image segmentation, control point
selection and correspondence, transformation parameter
estimation and so on. An automated feature-based image
registration procedure is one of the choices. In fact it includes
two parts: one is a image-image registration procedure,
another is a image-map registration. Knowledge about
matching control points and recognition can be obtained from
database.
3.2 Automated change identification subsystem
After accurate registration we use the image difference
technique (one kind of widespread used change detection
techniques) to get the difference image. The selection of
threshold is automatic and the algorithm is suggested by
Bruzzone (2000). Although because of the difficulty and
uncertainties of threshold selection we can’t obtain the very
precise changed area, it is possible to identify those areas with
larger probability (this is certain change information) And those
area with smaller probability would be divided into uncertain
change information naturally. Obviously, this procedure is
based on pixel level.
3.3 Automated change feature extraction subsystem
For certain and uncertain changed information especially
uncertain information, the best way to discriminate them is to
analyze further by additional information. The main features in
images are edge and area that are the most important factors
for object recognition. Although radiance values in images
change at two different dates because of the differences in
atmosphere conditions, illumination condition, relief condition,
noise and so on, the basic shape and outline of features in
images always keep invariant. So some image processing
algorithms such as edge enhancement, edge detection, area
segmentation and edge extraction are applied into the system.
The goal of extracting a lot of edge and area image features
lies in three aspects: one is to compare them between the two
images for identify changed area, another is to provide image
features to be recognized and the third is to update the
database using changed image features. For those uncertain
