The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
movement of traffic, shipping containers, and other transient
features. The effectiveness of the technique also depends, to a
greater or lesser extent, on the nature of the data that is to be
extracted. For example, a process that can identify changes to
road networks in a rural area may not work as well in an urban
setting, and may be entirely unsuitable for the detection of
changes to buildings. The nature of the input data also plays a
major role in the process. The resolution (ground sample
distance) of the images will often dictate which techniques are
most appropriate to use, while the presence of an infra-red
component in an image can be of great importance when
separating vegetation from the built environment.
One task of the data collection area of a mapping agency is to
detect changes between the features held in a topographic
database, and the features present in an image or set of images.
In an ideal situation, an automatic technique is required which
compares the existing topographic data with a single up-to-date
image and identifies the differences between them. It would be
very useful if the process could also filter out any changes
which are unlikely to be of interest (e.g. traffic) and present to
the photogrammetrist only those changes which are required by
the data capture specification. This paper presents various
methods of automatic change detection investigated by
Ordnance Survey during the last year. These methods are
largely based on image classification, followed by feature
comparison techniques. Per-pixel and per-object classification
methods were tested, using both off-the-shelf systems and
techniques developed in-house.
2. CHANGE DETECTION METHODOLOGY
2.1 Which changes are important?
Of the many different features present in the national
topographic database, changes to the built environment proved
to be the most critical. The construction and demolition of
buildings are both important to many users of spatial data and,
as uncovered by the manual change detection processes, are
often overlooked by third-party change intelligence sources. It
was therefore decided that our research should concentrate on
the detection of new buildings and demolitions. Once this
decision had been made, the exact method of change detection
had to be determined. Rather than taking one single approach,
several different methods were tried and compared, to find the
one showing the most promise for future implementation into
the production system.
2.2 Source data
Since the target of this research is the implementation of a
change detection process within the photogrammetric data
collection production system, it was important to use as inputs
only those types of data which would be readily available to
that system. For several years, Ordnance Survey has used an
Intergraph Z/I Imaging Digital Mapping Camera (DMC) as its
primary image data source. The inputs to this research were
therefore constrained to the data which can be extracted from
this DMC imagery. Both the panchromatic and 4-band 12-bit
multispectral images from the DMC were used at some point in
the research. Rather than relying solely on the spectral aspects
of the data, the projects also used a digital surface model,
created fully automatically from the overlapping panchromatic
imagery. This allowed us to discriminate more easily between
man-made and natural objects in the scene.
Two test sites were chosen, both of which have undergone
many changes recently: one near the Heathrow Airport
Terminal 5 junction of the M25 London orbital motorway, the
other in the urban centre of Bournemouth, a city on the south
coast of England. The images were collected during the 2005
and 2006 flying seasons. The two sites gave us the opportunity
to test the algorithms in different physical environments, to
indicate whether the process is likely to be transferable for use
in different parts of the country.
2.3 Change detection via image classification
Each of the change detection methods investigated involved an
image classification process. Both per-pixel and per-object
classifications were undertaken, using various methodologies
and several different software packages. Initially, the
classification techniques were applied to the images to identify
buildings, roads, trees, other vegetation, water bodies and roads.
These were then filtered to identify the changes to buildings,
which were the main focus of the research.
2.4 Per pixel classification
One of the main challenges in the detection of urban change is
the spectral heterogeneity of the urban land cover (Small 2001).
The many different surface types and objects present within an
urban scene can often generate different spectral responses for
essentially a single land cover. In order to deal with this
problem, two relatively new per-pixel image classification
techniques, which had the potential to discriminate between
objects in an urban scene, were applied to the images. The first
of these was the Support Vector Data Description (SVDD), a
one-class classifier developed by Tax and Duin (1999). The
SVDD is based upon the principles of Support Vector Machines.
Being a one class classifier, it works on the basis that only
target class data are used in the training stage. The target class
refers to the class of interest and it is assumed that it is sampled
well and that enough training data is available. However, in the
testing and validating stage the classifier will encounter outlier
data that was not present in the training stage. The classifier
must therefore have the capacity to distinguish if the data in a
testing set belongs to the target class or it is unknown and as
such belongs to the outlier class (Tax and Duin, 1999). In order
to give the maximum information about each class and to allow
comparison between different images taken at different times,
the variables used included band ratios and texture for each of
the bands, NDVI (Normalised Difference Vegetation Index) and
a DSM (digital surface model) for the area. The training set was
composed of a mixture of different pixels from different roof
top materials. The testing set was composed of pixels, 50%
belonging to the target class and 50% belonging to all the other
classes present in the image. All the pixels were chosen
randomly from across the image. The pixels that formed the
testing set were totally independent from those used in the
training sets to avoid any biases in the confidence level of
accuracy.
The second per-pixel classifier was a decision tree method,
specifically the CART decision tree software developed by
Salford Systems. Decision trees have long been popular in
machine learning, statistics and other disciplines for solving
classification problems. Decision trees are very flexible and
can handle non-linear relationships between features and classes
(Friedl and Botley, 1997. CART uses several different
approaches to the splitting of the decision tree, including the
Gini Index, entropy, and class probability. CART uses an over-