Accuracy Assessment for Boolean and Fuzzy Classificaiton in Tripoli, Libya

Abdulhakim Khmag, Alexis Comber, and Peter Fisher
Department of Geography, University of Leicester, Leicester, LEI 7RH, United Kingdom
{aek9, ajs36, pffl}@le.ac.uk

Abstract: Satellite imagery is a longstanding and effective resource for environmental analysis and monitoring at local, regional and global scales. Thematic map accuracy continues to be problematic; especially when Boolean representations are used as each image pixel is assumed to be pure and is classified to one and only one class. In reality the pixel may be mixed, containing many classes. Fuzzy classifications may be useful as multiple class memberships are assigned. A membership function is defined for each class against the feature value (digital numbers) and membership values of a class to belong to a particular pixel are determined based on function definition. Quantifying classification accuracy is an important aspect of map production as it allows confidences to be attached to the classifications for their effective end use. Accuracy measures serve as the analysis of errors, arising from the classification process due to complex interactions between the spatial structure of landscape, classification algorithms, land cover change and sensor resolutions. The accuracy of Boolean classifications may be assessed in a number of different ways and traditionally the error matrix is generates overall accuracy, producer and user accuracy and kappa coefficients. A number of additional measures, for example, classification success index and Tau coefficient. Therefore, other accuracy measures may appropriately including the fuzziness in the classification outputs and/or reference (ground) data. These include Euclidean, entropy, cross-entropy, and LI distances, fuzzy set operators, and fuzzy error matrix based measure. Generally, the confusion matrix, compares ground observations for a given set of validation samples with the classification result. Accuracy assessment usually includes three essential mechanisms: sampling design, response design, and estimation and analysis procedures. Selection of a suitable sampling strategy is a critical step, the major components of a sampling plan include sampling unit (pixels or polygons), sampling design, and sample size. Possible sampling designs include random, stratified random, systematic, double, and cluster sampling. This paper compares different accuracy assessment measures for Boolean and Fuzzy classified images at different dates in Tripoli city, Libya and the results show that classification accuracy related to a range of different factors. The result also shows the use of kappa coefficients for accuracy assessment gives good result for both Fuzzy and Boolean.

Keywords: Accuracy assessment, Boolean classification, fuzzy classifications, kappa coefficient, error matrix