The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B4. Beijing 2008
grow/prune-back strategy in which the decision tree is allowed
to grow to a high level of complexity, and is then pruned back
to a more manageable level. This approach helps prevent the
classifier getting stuck at a local optimum point too early in the
classification process.
2.5 Object-based classification
An object-based classification was adopted by the third method,
using Definiens software. This object-based method offers
some advantages over the traditional pixel-based approach, in
that, in addition to the spectral data, it can incorporate shape,
texture and local context into the classification process (Benz et
al, 2004). Also, by segmenting the image into homogeneous
objects prior to classification, this approach helps to reduce the
spectral variability within each class.
The first stage of the classification process was the
segmentation of the image into homogeneous objects.
Definiens employs a region-merging segmentation method,
starting with individual pixels and merging adjacent regions
until a user-defined threshold of heterogeneity is reached (Benz
et al, 2004). After the initial segmentation, the objects were
classified using a hierarchy that divides the objects into
increasingly refined categories, using user-defined membership
functions. The membership functions are determined by
observation, i.e. by manual examination of the characteristics of
the objects of interest in the image. Although this is a slow
process, it should only be required once, in the expectation that
the objects of interest in other images will have similar
characteristics. Earlier splits were simpler and more reliable
(e.g. vegetated objects were split from non-vegetated objects
using only the Normalized Difference Vegetation Index, NDVI,
with very high reliability). Subsequent classes became
progressively more difficult to separate and the reliability of the
classification decreased. Spectral properties, shape and texture
features were all used in the identification of buildings.
Shadow was also used to constrain the buildings class to those
objects within a defined distance from a shadow object. This
methodology is discussed further in Sanchez Hernandez et al
(2007).
3. CLASSIFICATION RESULTS
3.1 Classification accuracy
The first test of the classification methods used a 300 m by
270 m subset of the Heathrow dataset. The aim of the test was
to determine the accuracy of each classification, in order to
assess its usefulness in a subsequent change detection step. A
training set of 600 pixels was selected, with 50% of the pixels
in the buildings class and 50% in the non-building class. A
further set of 200 randomly selected pixels formed the test set.
The overall accuracies of the test pixels are shown in Table 1.
The first set shows the accuracy achieved using only the
spectral components of the image, while the second set shows
the accuracy achieved when the digital surface model was also
included in the source data. As can be seen, the DSM data
increased the accuracy of the classification in all cases. This
was as expected, since the buildings can be spectrally very
similar to the surrounding man-made roads and other surfaces;
while the difference in height distinguishes them immediately
from the ground surface.
Classification
SVDD
Decision Tree
Object-Based
Spectral only
70.0%
73.0%
76.0%
Spectral + DSM
85.0%
90.5%
91.0%
Table 1: Overall accuracy of the classified Heathrow subset
For the object-based classification, a slightly different method
was applied when the DSM data were included. First, the
image was classified by a user-defined sequence of processes
within Definiens. The main features used in this classification
were shape, slope, and height in context to neighbouring
features. The spectral information used in the process was of a
lesser importance and was limited to the normalised difference
vegetation index (NDVI) and the brightness. These were
chosen since other spectral features are likely to differ
significantly between images captured in different areas, at
different dates and in different lighting conditions. Localised
differences in height between building features and the
surrounding land are less likely to differ between images and
these should prove more reliable. It is thought that this will
allow the process to be generalised and will therefore increase
the potential for transferring the rule-set to different images.
The results on a larger test area were similar to those shown in
Table 1. As with any classification process a certain amount of
misclassification is inevitable. In the per-pixel classifications,
the main reasons for misclassification were the presence of
objects on the surface which could not be distinguished from
the surrounding buildings. These objects included large
vehicles and shipping containers, which are similar in height
and area to small buildings. A further cause of mis
classification was the use of the DSM data as absolute height
values (rather than heights relative to the surrounding pixels).
Using the absolute heights works well in a flat area, but in areas
of undulating terrain there will be problems between rooftops
and man-made surfaces at the top of slopes.
In the object-based classification, vehicles and containers were
also often misclassified as buildings. This classification,
however, could distinguish successfully between rooftops and
manmade surfaces, by using the local slopes around each of the
objects. Some misclassifications still remained, due to very
low-rise buildings such as sheds and garages, which were not
recognised as buildings. This occurred because these objects
failed to meet the height and area thresholds within the rule-set.
This was expected to occur because the rule-set was devised to
meet the Ordnance Survey specification for significant changes
(“Category A” changes). Category A includes newly built
residential buildings and demolitions, but small, non-residential
buildings are not within this specification. A Definiens rule-set
could be constructed to detect these small buildings, but this
would inevitably lead to misclassifications of objects of a
similar size and height, such as vehicles. In the end, a trade-off
must be made between the proportion of building features
correctly identified as buildings (true positives) and the
proportion of non-building features that are misclassified (false
positives).