Application of clustered resampling methods in assessing accuracy of cross-validated estimators of cross-classification probabilities of nearest-neighbor classifiers

Abstract

In several applications objects of interest can be investigated only if they are observed from a long distance. Remotely sensed raster data are collected by sophisticated sensors. These remotely sensed data are transformed into digital images showing special properties of the observed objects by differently colored pixels. Transformation of elements of remotely sensed data into colored pixels can be realized by classifiers. The case with a fi??nite number of classes (colors) is considered. In geomatics such classes can be different species of trees, e.g. spruce, pine and broad-leaved. In this paper the nearest neighbor (NN-) classifi??ers are considered. The NN-classifiers are widely used in the analysis of statistical data in many areas of research. The most important their characteristics are the cross-classifi??cation (CC-) probabilities. The crossvalidated (CV -) estimators of the CC-probabilities can be obtained using training sets. The accuracy of NN-classifiers is characterized by the distributions of the deviations of the CV -estimators from the true values of CC-probabilities, which depend on the training sets. Resampling from NN-clusters of connected NN-points can be used to obtain consistent estimators for the distributions of the deviations in the case of 1NN-classifi??ers. Two numerical experiments illustrate the suggested methods of resampling

Keywords

remotely sensed data; nearest-neighbor classifiers; estimation; cross-validation; point processes; clusters; resampling

Published in

Research report (Centre of Biostochastics)
2005, volume: 2005, number: 1, pages: 1-35
Publisher: SLU

SLU Authors

• Belyaev, Yuri

  • Department of Forest Economics, Swedish University of Agricultural Sciences
    

UKÄ Subject classification

Earth Observation

Permanent link to this page (URI)

https://res.slu.se/id/publ/6846