Likelihood-based image segmentation and classification: a framework for the integration of expert knowledge in image classification procedures

This paper describes a likelihood-based segmentation and classification method for remotely sensed images. It is based on optimization of a utility function that can be described as a cost-weighted likelihood for a collection of 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 user's point of view the result was either a maximum probability for truth (likelihood) or maximum utility (benefit). The method 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. As likelihoods are defined in the probability for spectral class domain, the method avoids problems with the segmentation of multi-spectral data encountered in methods based on edge detection. Due to these capabilities, the method represents a significant 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 was designed and used for experimentation. The result showed that 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 was used.