University of Twente Student Theses


Use of prior information on objects in contextual classification with Markov random fields

Caleb, Obiko Victor Onyimbo (2012) Use of prior information on objects in contextual classification with Markov random fields.

[img] PDF
Abstract:Classification of VHR remote sensing satellite imagery is an area that requires extensive research in the improvement of classification on objects. Classification results on objects are less accurate with class separability that is poor or excellent. This study, therefore, is focused in the introduction of prior information on objects in contextual classification with Markov Random Fields (MRF) aiming at improving classification. The prior information incorporated in the classification (in this research topic) is shape and size, tree crowns being the subject. Tree crowns have a round geometric shape with definite area which is modeled in the prior term of the MRF objective function. A tuning subset of the multispectral image that has a tree crown with grass, bare soil and shadow as the background is used in the estimation of MRF parameters. This study explores optimisation of the MRF parameters: Lambda, lambda segments, temperature, temperature update and lambda shape. Simulated annealing energy minimization algorithm is used to establish the optimal values of temperature and temperature update. Kappa values, producer, user and overall accuracy determine the optimal parameters for lambda and lambda segments. The optimal MRF parameters obtained are applied on the tuning subset; the results produced are more accurate and reproducible. The methodology of this study is a Hierarchical Markov Random Field (HMRF) approach. In level one, the MRF pixel based level, the energy of each class label is minimized, increasing the probability of that pixel being assigned a particular class by penalizing adjacent pixels. This MRF method is iterative and converges when the energy is minimized to zero. In level two, the MRF object based level, shape and area as spatial context are modeled in the prior term of the MRF energy function by the concept of smoothness prior. Accuracy assessment is done by use of a reference panchromatic image subset. The HMRF results of the tuning subset are analysed in the confusion matrix to determine the tree crown pixels that have been correctly classified and those misclassified (Errors of omission and commission). Implementation is done on two different subsets: 1) A well separated three tree crown area, 2) Tree crowns close to each other. The results obtained show that the HMRF method improves classification on tree crowns that are separated and the classification accuracy is low for interlocked ones. In conclusion, the HMRF method developed by integrating shape and area outperforms the MLH in the classification of separated tree crowns. Key words: - Hierarchical Markov Random Fields (HMRF), Simulated Annealing (SA), Iterative Conditional Modes (ICM), Maximum Likelihood Classification (MLC), Tree crown.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page