University of Twente Student Theses


Accuracy assessment of fuzzy classification

Ezeilo, Chekwube Bartholomew (2011) Accuracy assessment of fuzzy classification.

[img] PDF
Abstract:Remote sensing technology captures information about an object. The wealth of information extractable from an object is dependent on the nature of the object and the technique applied in the extraction. Objects in remote sensing can be divided into two forms: (1) Objects that have definite boundaries - they can be easily identified and described they are referred to as crisp objects; (2) Objects whose boundaries are shrouded in mysteries (indefinite-cannot be easily determined and identified), they are referred to as fuzzy objects. Assessing the accuracy of crisp object can be easily done using the error matrix which is by convention, the traditional way of assessing the accuracy of crisp objects. Defining the boundary of a fuzzy object is difficult and assessing the accuracy as well, has not been standardized; hence, it is the focus of this research. The fuzziness in an object is as a result of uncertainty, before an attempt is made to characterise the object, there is need to identify the uncertainty in the object. Several forms of uncertainty exist; however, the uncertainty in the fuzzy object of interest was defined as the vagueness that results when the boundary of the object lies within the zone of transition. The east fork fire burn scar that occurred in April 2004 was the choice of the fuzzy object. MODIS and ASTER images were acquired for classification and generating the reference. MODIS image was classified using four classification techniques (unsupervised crisp (ISODATA), supervised crisp (Maximum likelihood), unsupervised and supervised fuzzy-c-means). Three cases were investigated for each classification scheme, in which 2, 3 and 4 classes were defined as Case A, Case B and Case C respectively. The reference data was generated by using the same classification scheme as MODIS’; the output was degraded to make it comparable with the pixels of the MODIS. The accuracy of the classifications was judged using the entire image as samples. The conventional error matrix was used to assess the crisp outputs; the fuzzy error matrix was used to assess the fuzzy outputs. The crisp and fuzzy outputs were also assessed by determining the association of each class with the reference this was referred to as correlation coefficient determination. The results obtained from the crisp assessment was higher than those obtained from the fuzzy assessment, the correlation coefficient values were higher in the fuzzy outputs than for the crisp outputs. Also, the fuzzy outputs gave a better description of the burn scar phenomenon than that obtained from the crisp descriptions. Key words: Fuzzy objects, Un-supervised and supervised crisp classifications, Un-supervised and supervised fuzzy classifications, Accuracy assessment.
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