Probabilistic segmentation of remotely sensed images

B. Gorte

Research output: Thesisexternal PhD, WU


<p>For information extraction from image data to create or update geographic information systems, objects are identified and labeled using an integration of segmentation and classification. This yields geometric and thematic information, respectively.</p><p>Bayesian image classifiers calculate class posterior probabilities on the basis of estimated class probability densities and prior probabilities. This thesis presents refined probability estimates, which are local, i.e pertain to image regions, rather than to the entire image. Local class probability densities are estimated in a non-parametric way with an extended k-Nearest Neighbor method. Iterative estimation of class mixing proportions in arbitrary image regions yields local prior probabilities.</p><p>The improved estimates of prior probabilities and probability densities increase the reliability of posterior probabilities and enhance subsequent decision making, such as maximum posterior probability class selection. Moreover, class areas are estimated more accurately, compared to standard Maximum Likelihood classification.</p><p>Two sources of image regionalization are distinguished. Ancillary data in geographic information systems often divide the image area into regions with different class mixing proportions, in which probabilities are estimated. Otherwise, a regionalization can be obtained by image segmentation. A region based method is presented, being a generalization of connected component labeling in the quadtree domain. It recursively merges leaves in a quadtree representation of a multi-spectral image into segments with arbitrary shapes and sizes. Order dependency is avoided by applying the procedure iteratively with slowly relaxing homogeneity criteria.</p><p>Region fragmentation and region merging, caused by spectral variation within objects and spectral similarity between adjacent objects, are avoided by regarding class homogeneity in addition to spectral homogeneity. As expected, most terrain objects correspond to image segments. These, however, reside at different levels in a segmentation pyramid. Therefore, class mixing proportions are estimated in all segments of such a pyramid to distinguish between pure and mixed ones. Pure segments are selected at the highest possible level, which may vary over the image. They form a non-overlapping set of labeled objects without fragmentation or merging. In image areas where classes cannot be separated, because of spatial or spectral resolution limitations, mixed segments are selected from the pyramid. They form uncertain objects, to which a mixture of classes with known proportion is assigned.</p><p>Subsequently, remotely sensed data are used for taking decisions in geographical information systems. These decisions are usually based on crisp classifications and, therefore, influenced by classification errors and uncertainties. Moreover, when processing spatial data for decision making, the objectives and preferences of the decision maker are crucial to deal with. This thesis proposes to exploit mathematical decision analysis for integrating uncertainties and preferences, on the basis of carefully estimated probabilistic class information. It aims to solve complex decision problems on the basis of remotely sensed data.</p>
Original languageEnglish
QualificationDoctor of Philosophy
Awarding Institution
  • Molenaar, M., Promotor
  • Stein, A., Promotor
Award date12 Oct 1998
Place of PublicationEnschede
Print ISBNs9789061641575
Publication statusPublished - 1998


  • remote sensing
  • image processing
  • theses


Dive into the research topics of 'Probabilistic segmentation of remotely sensed images'. Together they form a unique fingerprint.

Cite this