LYSIS
IMAGES
THM
PLE CUE
Ali Akbar Abkar
LIKELIHOOD-BASED IMAGE SEGMENTATION AND CLASSIFICATION:
CONCEPTS AND APPLICATIONS
A. A. ABKAR', M. A. SHARIFI"
Soil Conservation and Watershed Management Research Center (SCWMRC)
P.O. Box 13445-1136, Tehran, Iran
Abkar@itc.nl
“International Institute for Aerospace Survey and Earth Sciences (ITC)
P.O. Box 6, 7500 AA Enschede, The Netherlands
Alisharifi @itc.nl
Working Group IC-7
KEY WORDS: Classification, Knowledge Engineering, Remote Sensing, Segmentation.
ABSTRACT
This paper describes a likelihood-based segmentation and classification method for remotely sensed images. It is based
on optimization of a utility function, described as a function of likelihood of all objects and their parameters. As the
likelihood or posterior probabilities are calculated per object rather than per pixel, the variance in (spectral) likelihoods
will be greatly reduced. From a users point of view the result has either a maximum probability for truth (likelihood) or
maximum utility (benefit). It includes a new approach for segmentation, which is based on criteria derived from local
average likelihoods instead of local means or variances, making the segmentation method much less sensitive to
radiometric outliers. Due to these capabilities the method represents a step towards operationalization of remote
sensing. This approach can also be seen as a framework for integration of external knowledge with image classification
procedures. To evaluate the concept, a software tool has been designed and used for experimentation. The result showed
that the per-object maximum likelihood performs much better than the per-pixel method. It led to higher classification
accuracy and explicit utilization of the geometrical and topological information about land use objects and land use
processes. For generation and testing of the geometric models, the problem of deforestation in Thailand is used.
1 INTRODUCTION
Image segmentation and classification is usually applied at the pixel level (using solely the radiometric properties of
pixels on an individual basis) by processing the data in order to arrive at a meaningful partitioning of the image. Pixel-
based segmentation and classification procedures start off with “raw-data” (a set of samples of reflectance values) and
end up with a segmented and labeled image. In this process the preliminary features or segments are located in the
Remote Sensing (RS) data and then examined for consistency (to find correspondence with real world objects). The
major problem with these pixel-based approaches is that there is insufficient information to completely isolate the
required objects because of the complexity of the objects and of the object radiation interaction in RS images.
Considerable research has been dedicated to improve the pixel-based segmentation and classification results. For
example, "classification accuracy improvement” (Hutchinson, 1982), “knowledge-based image analysis” (Civco, 1989;
Mulder and Schutte, 1992; Abkar and Sharifi, 1995), “object-based classification” (Janssen ef al., 1990), “integration of
algorithms” (Schoenmakers, 1995; Gorte, 1998; Janssen, 1994), “model-based image analysis” (Mulder, 1993), "image
understanding” (Kohl and Mundy, 1994), "image fusion" (Fórstner and Lócherbach, 1992; Pohl and van Genderen, 1998).
It is therefore clear that the spectral data alone is not adequate for rigorous information extraction from RS images.
Combining RS spectral data with other multisource (ancillary) data allows the use of more knowledge, which can
improve information extraction. Notwithstanding much effort spent already, an operational and (integrated) analysis
method is as yet required to combine (fuse) all this Earth observation data with geo-information systems (GISs) and
models to make it useful for a user community that is concerned with mapping, monitoring, managing and modeling the
Earth surface (Ehlers, 1991; Wang, 1991; Hahn and Baltsavias, 1998; Baatz, 1999).
In this paper we present a method that allows integration of available data and/or knowledge contained in GIS about objects
and processes into the analysis of remotely sensed data through hypothesis driven image analysis and effective
representation of knowledge about scene objects, sensor, and sensing processes. The analysis, as such, is guided by
expectations from the domain and the analysis starts with a model (at the object level) to partition the image. This model-
based classification of objects focuses on the integration of RS and GIS; this is done by shape hypothesis generation using a
GIS and evaluation of shape hypotheses and shape parameters in the evidences provided by RS images.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 9
