×

You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

Title
Fusion of sensor data, knowledge sources and algorithms for extraction and classification of topographic objects
Author
Baltsavias, Emmanuel P.

International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
202
enhanced with multispectral image values. Not only the mean
value of the multispectral channels can be used; other measures
derived from the original bands, like NDVI, LAI or sigma
values, can be displayed. Object analysis opens up a new
chapter in textural analysis of the panchromatic data. It offers
the possibility to analyse the (relative) amount of panchromatic
sub-objects per multispectral pixel-object (since a multispectral
pixel represents an area of more than 400 m2, it is meaningful
to treat each pixel as a separate object). For this purpose, the
multispectral pixel layer is segmented in objects, each pixel
receiving a unique code. Then, the panchromatic dataset is
segmented, taking into account the borders of the multispectral
pixels. In a typical example of alpine forest (Fig. 1), this would
mean that a SPOT4 pixel of 23 x 23 m 2 contains 529
panchromatic pixels of 1 x 1 m 2 . After a two level segmentation,
the single SPOT pixel covering a dense coniferous forest area
contains an amount of ± 35 sub-objects, representing the
amount of crowns and shadow areas inside that multispectral
pixel-object. Using the same tolerance parameters for the
segmentation, the forest/road pixel contains only 25 objects.
The amount of panchromatic objects depends on the user-
defined tolerance parameters. Their relative amount is therefore
more important than the absolute one. Texture analysis, which
depends upon window based algorithms, such as the first and
second moment filter techniques, can then be expanded with the
relative number of sub-objects. This allows a mixed pixel
analysis, where the operator directly sees the spectral behaviour
of the multispectral channels and the spatial properties of the
high resolution panchromatic data. This is very helpful for
advanced mixed pixel analysis.
3. EVALUATION
The visual evaluation is still a reliable praxis in high resolution
image classification. Confusion matrixes (Richards, 1992),
which are used in pixel based analysis are not useful for object
evaluation.
In principle, an object classification is false or correct. When a
particular object receives a false classification, the statistics of
the area represented by that object-class can be seriously
affected. The fuzzy logic functions can determine the chance per
object of falling into a specific class. For example, two objects
belong to class A. Object 1 has a chance of 96 % and object 2 a
chance of 75 %. Furthermore, the chance of object 1 to belong
to class B is 92 % and for object 2, 15 %. The difference
between the first and second option is therefore also a powerful
indication of reliability. The evaluation procedure shows how
important it is for the operator to be an expert on the objects of
interest. A forester needs to know about the ecology of his
forest area and the parameters that influence spatial and spectral
behaviour of the objects of interest. This is similar for the
hydrologist and the geologist or any other field expert. Object
classification using fuzzy logic expects a high input of the field
expert to define the decision functions. This evaluation part of
the image classification is an area where a lot of research still
lies ahead.
4. DISCUSSION
Although the literature mentions the possibilities of object-
based image analysis since a few decades (Kettig and
Landgrebe, 1976), only latest-technology hardware, intelligent
software and high resolution images can advance this concept.
In practice, a Pentium II-like processor, >256 MB RAM and >6
GB hard disks are needed to have acceptable speed levels for
tests with large amounts of image objects. The limitations of
hardware development have not been reached yet. Therefore,
this field of study is practically driven by ‘getting the most out
of the modem machine’. Other interesting ways of image
classification should not be overlooked in the meantime. The
interaction with object-based remote sensing data and GIS
objects has a high potential. This concept needs a lot of
attention and might offer possibilities for map updating.
The object-based analysis can only be successful, if a proper
automatic evaluation is accepted among RS specialists. The user
is interested in high accuracy with a certain cost factor for a
specific set of objects of interest. Lower overall accuracy is
often acceptable for lower cost, but objects of interest might
need high accuracy. That makes evaluation standards for image
classification quite complicated, when object-oriented
classification is used. Then, evaluation might also become user
dependent. The increasing demand for the field expert
knowledge, however, is a good development.
5. CONCLUSIONS
Rapid improvements in hardware capacity, clever software and
modem sensors make it possible to realise ideas, which have
been developing for decades among the remote sensing
specialists. In practice, during the realisation of these ideas, new
challenges appear. This leads to increasing demands for
knowledge of ground truth. Ground truth involves not only the
current state of the landscape but also an insight in the
processes that formed and are forming the present scene of
interest. Not the whole scene needs to be interpreted; areas of
interest have to be analysed using the huge amount of
information available in high resolution data. Deriving
information from imaging data is often limited due to the
associated high costs. Only automatic procedures can reduce
these costs. Object-based analysis offers an interesting option to
improve automatic information derivation from an ‘exploding’
amount of data from different sensors.
The representation of multispectral data is based on three
domains: image space, spectral space and feature space
(Landgrebe, 1999). The terrain expert is very familiar with
image space. Object-based analysis allows a direct link between
the image space display and the database information on the
spectral and feature spaces. This makes object-based software
quite intuitive and user-friendly.
The robust query possibilities, well known in GIS, should be
used in the interpretation of imaging data. The automatic