University of Twente Student Theses

Login

Contextual image classification with Support Vector Machine

Goumehei, Elham (2010) Contextual image classification with Support Vector Machine.

[img] PDF
1MB
Abstract:Wide range of remotely sensed data obtained from different sensors is currently available. This data requires to be analyzed to get information. One way to analyze remote sensing data is classification. Choosing a suitable classification algorithm is important to efficiently use this large data set. Several approaches have been introduced where the contextual information is one of the applicable introduced models in classification of remote sensing data. For characterizing contextual information Markov Random Field (MRF) has been found an efficient tool. Application of MRF is based on maximum a posterior (MAP) estimation. It is employed as prior probability density function (p.d.f.). For the conditional p.d.f. often Maximum Likelihood Classification (MLC) is used where assumes classes are normally distributed. This assumption is not always a correct assumption. This research proposed a new MRF-SVM model that explores Support Vector Machine (SVM) instead of MLC. Since Implementation of SVM presented an improved classification results compare to other classifiers like MLC (Foody & Mathur, 2004; Foody, et al., 2006; Huang, et al., 2002; Pal, 2006; Pal & Mathur, 2005). The introduced model uses Simulated Annealing (SA) for energy minimization. Contribution of prior and conditional models was controlled by a smoothness parameter. SVM offers some flexibility choice of penalty parameter value and a kernel function. Influence of these choices was considered. SVM assigns label to classes while in application of SVM as conditional p.d.f. class probabilities are required. To compute class probabilities for SVM Plott’s theory was used. Using class probabilities from Plott’s theory model was implemented on image synthesized from an agricultural scene. The accuracy of produced results was assessed by means of kappa coefficient (k ). In addition, reproducibility of results was evaluated by standard deviation of ten runs of model for ten different input images. Results indicate sufficient classification accuracy where the maximum k is 0.95. During the procedure effect of class separability was investigated too. Also, performance of the model was compared to MRF based on normal distribution assumption. An illustration of MRF-SVM implementation on Synthetic Aperture Radar (SAR) image was presented to demonstrate applicability of the developed model for classification of real data. In conclusion, the experimental results prove the effectiveness of the developed model. Performance of the model on synthetic data in terms of accuracy and reproducibility is acceptable. The model gives high k value while use real images may reduce it. Also employed image is smooth that may positively affect classification accuracy. The strength of the model is observed through classification results of exponentially distributed classes. The results of MRF-SVM for both normally and exponentially distributed classes are nearly identical whereas MRF based on MLC does not behave similarly for data with different probability distributions. In terms of computational time, the new model has similar number of iteration as MRF based on MLC. The study shows that the MRF-SVM model is applicable for classification of remotely sensed data. Key words: Markov Random Field, Support Vector Machine, maximum a posterior, Maximum Likelihood, class probability, Simulated Annealing, class separability, class distribution, exponential distribution classes.
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:https://purl.utwente.nl/essays/92480
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page