The Effects of Sample Size Selection on Classification Accuracy

Creative Commons License

Sarp G., Erener A.

Lecture Notes in Information Technology, vol.30, pp.374-379, 2012 (Peer-Reviewed Journal)


Remote sensing data is used primarily for classification of various features within a scene,

thereby to create a thematic map. The classified images provide possibility to distinguish between

different types of surface features based on their spectral responses or ‘spectral signatures’. However,

a classification process is not fully completed if its accuracy is not determined. One of the most

common classification accuracy assessment methods is preparation of error matrixes. Error matrixes

compare the ground truth data with the corresponding classification result. How the selected numbers

of samples effect the accuracy? In this study, it is aimed to investigate this consideration. Selection of

200, 150, 100, 50 sample points is analyzed in terms of overall accuracy and kappa statistics. In

addition to considering different number of sample points for accuracy assessment, the classification

performance of MSS and Pan-sharpened data is also evaluated. The study is applied to a test region at

Çankaya district of Ankara. The multispectral QuickBird with 2.4 m and pan image with 0.64 m

resolutions are used for the assessment. It is seen that the accuracy changed depend on the sample size