Sensitivity of Pixel-Based Classifiers to Training Sample Size in Case of High Resolution Satellite Imagery

Document Type : Original Article

Authors

1 Civil Engineering Department, Faculty of Engineering, Menoufia University, Egypt

2 Surveying Engineering Department, Faculty of Engineering, Benha University, Egypt

3 GIS & Surveying Engineer, Menofya Governorate, Egypt

Abstract

Thematic maps representing the characteristics of the Earth’s surface have been widely used as a primary input in many land related studies. Classification of remotely sensed images is an effective way to produce these maps. The value of the map is clearly a function of the accuracy of the classification. Selecting proper size of samples and classification method are essential issues to produce accurate thematic maps. In the present study, training data sets at various sizes used to investigate the effect of the training set size on the classification accuracy. Six supervised classification methods with different characteristics were applied to produce land use/land cover thematic map of the study area. The used classifier include: Parallelepiped, Minimum Distance, Mahalanobis Distance, Maximum Likelihood, Neural Network and Support Vector Machine (SVM). The results showed that optimum sample size differs from classifier to another. In the case of limited number of training pixels, SVM and maximum likelihood classifiers produced higher classification accuracies than the rest of classifiers.